# Source paper archive

- Source title: A trade-off shaping leaf mesophyll architecture emerges from a stochastic geometry model
- Author(s): Rahimi-Majd, M., M. Borsuk, A., Masoomy, H., C. Walker, N., R. Brodersen, C., Nikoloski, Z.
- Source landing URL: https://www.biorxiv.org/content/10.64898/2026.11.18.767598v1
- PDF URL: https://www.biorxiv.org/content/10.64898/2026.11.18.767598v1.full.pdf
- PDF SHA-256: a8acbcd03f24931f97ca3334e10cb973caaa1ee8f019acd3308a98902d0dbe2e
- Converter: pymupdf4llm 0.0.27
- Limits: Automated PDF text extraction; reading order, equations, tables and formatting may be incomplete or incorrect. Images are omitted, and no OCR was performed. Markup and links are neutralized for safe downloading. Consult the original PDF for the authoritative version.

---

&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)

### A trade-off shaping leaf mesophyll architecture emerges from a stochastic geometry model


Milad Rahimi-Majd &#91;1,2&#93;, Aleca M. Borsuk &#91;3&#93;, Hosein Masoomy &#91;1,2&#93;, Nancy C. Walker &#91;4&#93;, Craig R. Brodersen &#91;4&#93;,


and Zoran Nikoloski &#91;*1,2&#93;


1 Bioinformatics Department, Institute of Biochemistry and Biology, University of Potsdam, 14476


Potsdam, Germany

2 Systems Biology and Mathematical Modeling Group, Max Planck Institute of Molecular Plant


Physiology, 14476 Potsdam, Germany

3 New York Botanical Garden, 2900 Southern Boulevard, 10458 Bronx New York, US


4 School of the Environment, Yale University, New Haven, CT, 06511 USA


1 *Contact: nikoloski@mpimp-golm.mpg.de


2 Short title: Stochastic geometric model for leaf mesophyll


3 **Abstract**


4 Leaf mesophyll architecture has a great impact on plant functions, including photosynthe

5 sis. Yet, understanding the principles governing growth, stability, and interspecific variation


6 of leaf mesophyll architecture remains elusive. Here, we develop a computational framework


7 based on stochastic space tessellation to simulate two-dimensional paradermal slices of leaf


8 spongy mesophyll across 40 vascular plant species using three data-derived parameters: leaf


9 porosity, cell arm length, and cell arm diameter. The framework successfully reconstructs


10 leaf slices while preserving six biological, topological, and cell-level traits, with one- to three

11 armed cell geometries generally outperforming circular cells. Importantly, we further develop


12 a theoretical model based on stochastic geometry that closely agrees with the simulations and


13 reveals that leaf porosity strongly influences cell wall construction cost, while cell-arm statis

14 tics shape this relationship. Our parsimoniously parameterized model provides a basis for


15 optimizing leaf architecture to maximize carbon dioxide uptake while minimizing cell wall


16 construction costs.


1


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)

### 17 1 Introduction


18 Plant leaves constitute the primary site of photosynthesis. Within the leaf, the mesophyll is the

19 main photosynthetic tissue that is generally differentiated into two regions: the palisade meso
20 phyll, composed predominantly of densely packed, cylindrical cells, and the spongy mesophyll,

21 characterized by less densely packed cells that range from spherical to lobed &#91;1&#93;. In addition to

22 cells, the mesophyll contains an extensive network of intercellular air space (IAS) between the

23 cells. This airspace facilitates gas exchange between the atmosphere and mesophyll cells &#91;2&#93;. As

24 a result, the IAS contributes to shaping the leaf mesophyll conductance ( _g_ _m_ ) that quantifies the

25 ease with which carbon dioxide ( _CO_ 2 ) moves from stomata to its site of fixation by Rubsico in the

26 chloroplast stroma &#91;3&#93;. Further, IAS contributes to modulating both the gas phase component of

27 conductance ( _g_ _IAS_ ) directly, via diffusional pathways through the IAS, as well as the liquid phase

28 component of mesophyll conductance ( _g_ _LIQ_ ) indirectly, via diffusional pathways through the mes
29 ophyll cell surface area exposed to the IAS &#91;4, 5&#93;. Since mesophyll conductance is regarded as one

30 of the major limitations for photosynthesis &#91;6&#93;, understanding the extent to which IAS architecture

31 determines mesophyll conductance can be used as an entry point to optimizing photosynthesis.

32 Specifically, altering the leaf IAS density and the pattern of its distribution has been empirically

33 shown to improve photosynthetic efficiency in plants, unveiling the direct importance of leaf IAS

34 and cell packing geometry to leaf photosynthesis &#91;7&#93;. Beyond its role in carbon assimilation, IAS

35 architecture substantially affects water-use efficiency and plant resilience to changing climatic con
36 ditions, making it an important target for crop improvement aimed at stability of food production

37 systems &#91;8&#93;.

38 Despite this essential role of IAS in leaf function, growth, and energy production, the mech
39 anisms governing its development, evolution, and structural stability remain poorly understood

40 &#91;8&#93;. Recent studies have proposed hypotheses regarding the formation and organization of the

41 leaf mesophyll IAS from both dynamic and static perspectives: (i) the developmental mechanisms

42 underlying the emergence and evolution of IAS during cell wall formation, cell division, and epi
43 dermal growth &#91;8, 9, 10, 11&#93;; and (ii) the structural and topological principles governing the spatial

44 organization of IAS in mature leaves across plant species &#91;1, 12, 13, 14&#93;.

45 Theories of IAS development in plants posit that it is formed following the establishment of

46 a continuous cellular tissue and arise through three principal developmental mechanisms &#91;8&#93;: (i)

47 lysigeny, involving the generation of tissue cavities via programmed cell death; (ii) schizogeny, in

48 which IAS are created through the physical de-attachment of neighboring cells; and (iii) expan
49 sigeny, whereby existing air spaces are expanded through localized expansion of cell wall regions

50 bordering the intercellular space &#91;8, 9&#93;. While lysigeny is considered to be uncommon in plant

51 leaves, the other two processes are considered as the main mechanisms governing leaf IAS ex

2


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


52 pansion, initiating from multicellular junctions &#91;8, 9, 15&#93;. Live-cell imaging of the microtubule

53 cytoskeleton in developing mesophyll cells has provided experimental support for this view, re
54 vealing a limited contribution of schizogeny and a predominant role for expansigeny in airspace

55 expansion &#91;16&#93;. More specifically, during an initial schizogeny, the cell wall is divided into free

56 walls (exposed to IAS) and shared walls (cell-cell contact walls), and then almost all subsequent

57 expansion occurs based on the selective expansion of the free walls through expansigeny, while the

58 shared walls remain fixed in size &#91;11&#93;.

59 Alongside the development of computational frameworks for simulating plant cell morphogen
60 esis &#91;17, 18, 19&#93;, cell growth based on expansigeny has been modeled using deformable polygons

61 governed by physical laws &#91;10&#93;. The incorporation of key mechanical parameters, _i.e.,_ cell wall

62 stiffness, cell–cell adhesion, and constant boundary pressure, has enabled systematic investigation

63 of the mechanical balance required for spongy mesophyll tissue to remain stable as intercellular

64 void space increases &#91;10&#93;.

65 Following mesophyll development through expansigeny, the final extent of free walls, corre
66 sponding to the mesophyll cell surface area exposed to the IAS ( _S_ _m_ ), plays a critical role in regulat
67 ing the _CO_ 2 diffusion pathway during _g_ _m_ . This exposed surface forms the interface at which _CO_ 2

68 transitions from the gaseous phase within the IAS to the liquid phase inside mesophyll cells, poten
69 tially affecting both conductance phases _g_ _IAS_ and _g_ _LIQ_ &#91;4, 5, 6, 20, 21&#93;. As a result, an increase in

70 _S_ _m_ enlarges the available surface for _CO_ 2 uptake that can enhance the internal availability of _CO_ 2

71 within the mesophyll &#91;2, 7, 20&#93;. In contrast, although shared walls are essential for maintaining

72 the mechanical stability of mesophyll tissue &#91;10&#93;, an increase of their size incurs additional carbon

73 and resource costs for cell wall construction and maintenance while providing no direct benefit to

74 the gas-exchange interface &#91;22&#93;.

75 In parallel with the IAS dynamic growth models, quantitative analyses of mesophyll tissue

76 architecture have provided new insights into the structural properties of the spongy mesophyll,

77 _i.e.,_ the lower mesophyll layer beneath the palisade mesophyll, characterized by spherical to lobed

78 cells. This layer is of particular interest due to its structural variability and its close association with

79 stomatal pores, which are predominantly located on the lower leaf surface &#91;1, 8, 10, 11&#93;. In par
80 ticular, recent work has shown that the leaf spongy mesophyll exhibits species-specific structural

81 order, clarifying traditional descriptions of the tissue as largely irregular and non-ordered (1 and

82 references therein). While leaf spongy mesophyll varied along a spectrum of phenotypes, two cell

83 size–shape parameters, _i.e.,_ cell arm length and cell arm diameter, successfully classified 39 vas
84 cular plant species into two primary structural classes: honeycomblike and non-honeycomblike.

85 The honeycomb-like species exhibit highly ordered spongy mesophyll organization with pris
86 matic, hexagonally-dominant IAS voids, whereas the latter category corresponds to a less ordered,

87 isotropic cellular network &#91;1&#93;. Importantly, the honeycomb-like structures were shown to follow


3


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


88 Euler’s law, revealing a fundamental topological constraint underlying the organization of spongy

89 mesophyll.

90 Beyond broad geometric and topological classifications of the spongy mesophyll, its underly
91 ing geometric principles remain largely unexplored. These principles are key to understanding the

92 constraints governing the expression of leaf morphological traits. Such an understanding could

93 reveal the extent to which individual traits can vary between plant species while maintaining struc
94 tural stability of the mesophyll tissue.

95 Here, building on previous findings, we develop the SCellPack framework based on random

96 space tessellation to reconstruct the 2D structure of the paradermal slices of the spongy mesophyll

97 using only three parameters: mesophyll porosity, cell arm length, and cell arm diameter. SCell
98 Pack successfully reproduces the structural properties of the original slices across several traits.

99 We formulate our model within the framework of a homogeneous Boolean model &#91;23&#93;, providing

100 a mechanistic link between free cell wall organization, leaf porosity, and cell arm size. This frame
101 work identifies the balance between leaf porosity and mean cell arm size, as a key determinant of

102 tissue structure and stability. Our findings demonstrate that a simple stochastic geometry model

103 based on a Poisson point process &#91;23&#93; can capture the geometric organization of the spongy meso
104 phyll across diverse plant species and cell geometries. Finally, using our framework, we derive a

105 data-based relationship between cell wall construction cost and leaf porosity; we then investigate

106 alternative cell morphologies that could shift this relationship, highlighting possible strategies for

107 improving the leaf architecture to maximize _CO_ 2 uptake while minimizing cell wall construction

108 costs.

### 109 2 Results


110 **2.1** **SCellPack provides a simulation framework for leaf 2D structures**


111 To model the geometry of the leaf mesophyll, we developed the stochastic cell-packing, SCellPack,

112 framework based on a stochastic space-tessellation approach. We considered five distinct cell

113 geometries: circular, 1-armed, 2-armed, 3-armed, and 4-armed (Fig. 1), with an additional 2
114 armed geometry having an arm angle of 180 _&#91;◦&#93;_ (2-armed 180 _◦_ ) considered in selected analyses (see

115 section 4.2). This definition was motivated by previous work that identified spherical and triply

116 armed morphologies as representative geometries of plant spongy mesophyll cells across 39 plant

117 species &#91;1&#93;. Specifically, the cells were best characterized either by circular (in 2D) geometry or

118 by lobed geometries based on the tissue skeleton with an average of three arms &#91;1&#93;. Building

119 upon this model, we extended the range of considered morphologies to include cells with both

120 fewer and more arms, thereby enabling the investigation and comparison of additional geometrical

121 configurations (see Ref. &#91;1&#93;).


4


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


122 We used two-dimensional paradermal slices from X-ray micro-computed tomography (mi
123 croCT) images of 40 vascular plant species as the basis for our reconstruction framework (see

124 section 4.1). The simulation box for each species matched the dimensions of the corresponding

125 microCT slice. The fraction of intercellular airspace to total area (porosity; _f_ _IAS_ ) of each parader
126 mal slice, together with the leaf-averaged cell arm length ( _L_ _arm_ ) and arm diameter ( _D_ _arm_ ), served

127 as input parameters for the simulations. For each cell, values of _L_ _arm_ and _D_ _arm_ were randomly

128 sampled from their respective empirical distributions, characterized by the mean and standard devi
129 ation of the measured values. Using these parameters, cells were sequentially positioned at random

130 within the simulation domain, with each cell assigned an independently sampled size. Whenever

131 a newly generated candidate cell’s random placement may result in overlaps with one or more

132 previously placed cells, the cell region that could not be located in free area was clipped such

133 that the shared cell walls followed the boundaries of the existing cells, while the free cell wall

134 remained unconstrained. The simulation stops when the target _f_ _IAS_ (see Fig. 1) or a stopping

135 criterion is reached (see below). This procedure effectively captures the geometrical consequences

136 of expansigeny, but does not provide a dynamic growth model.

137 Following the classification of the previous study &#91;1&#93;, the 40 species were grouped into three

138 categories: non-honeycomb, honeycomb, and neither honeycomb nor spongy (Fig. 1a). The non
139 honeycomb (11 species) and honeycomb (28 species) categories are characterized by circular (in

140 2D) and armed cell geometries, respectively (Fig. 1a). The latter category contains only a single

141 species, characterized by elongated cell arms radiating in different directions &#91;1&#93;.

142 We first assessed the integrity of the SCellPack reconstructions by determining whether the

143 framework accurately reproduced the prescribed target parameter. Specifically, because the slice

144 porosity ( _f_ _IAS_ ) was defined as a target variable based on experimental measurements, its accurate

145 reproduction was a prerequisite for the validity of subsequent analyses (Fig. S1b). The devia
146 tions from the target _f_ _IAS_ may arise from the stopping criterion of the cell-packing procedure,

147 which terminates the simulation after a predefined maximum number of consecutive unsuccess
148 ful placement attempts ( _i.e._, 200), thereby limiting computational cost (section 4.2). We therefore

149 evaluated reconstruction performance across all defined cell geometries, initially considering a

150 sample slice from the representative species within each of the three species categories. No sub
151 stantial deviations were observed between the porosity of the data paradermal slices and that of

152 the corresponding SCellPack-reconstructed slices. This is supported by the low leaf-level average

153 absolute deviation ( _≤_ 0 _._ 03) between the porosity of the data paradermal slices and that of the

154 corresponding SCellPack-simulated images relative to the range of leaf-level average _f_ _IAS_ values

155 across species (0 _._ 24–0 _._ 67) (Fig. S1). The negligible deviation can be explained by the substantial

156 free space remaining even at the minimum _f_ _IAS_ : approximately 25% of the simulation domain

157 remained unoccupied in the densest cases. Because the different cell geometries exhibited compa

5


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


158 rable cell areas, as they are generated from the same cell arm input parameters, this available space

159 was sufficient to reach the prescribed _f_ _IAS_ across geometries.

160 The degree of visual agreement between the reconstructed and original microCT images varied

161 across species categories and cell geometries. Consistent with the classification of Borsuk et al.

162 (2022) &#91;1&#93;, circular cells showed high visual similarity for non-honeycomb species but substantial

163 differences for species characterized by armed cells (Fig. 1). For instance, 4-armed cells showed

164 weak visual similarity across the representative species, whereas 1-, 2-, and 3-armed cell geome
165 tries produced reconstructions that more closely resembled the original microCT images, particu
166 larly for non-honeycomb species (Fig. 1). We note that these representative slices are shown only

167 for qualitative illustration, while quantitative evaluation across all species and cell geometries is

168 provided in the following sections.


6


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


**Figure 1: Comparative illustration of simulation results from the SCellPack framework with**
**real-world** **2D** **spongy** **mesophyll** **patterns.** **a.** Mean cell arm length ( _L_ _arm_ ) and mean arm diameter ( _D_ _arm_ ) of the spongy mesophyll across 40 plant species, with error bars indicating the
corresponding standard deviations (see Ref. &#91;1&#93;). Species are ordered by increasing arm length,
with blue, yellow, and brown shaded regions corresponding to non-honeycomb, honeycomb, and
neither spongy patterns, respectively. **b.** The SCellPack reconstruction for a representative slice
of one species from each species category. The selected non-honeycomb and honeycomb species
correspond to those with the most extreme mean arm lengths, with the smallest and largest _L_ _arm_,
respectively. The first row shows binary heatmaps of 2D paradermal leaf slices from the three
representative species, with cell regions shown in green and intercellular airspace (IAS) in white.
The fraction of intercellular airspace to total area ( _f_ _IAS_ ) and the mesophyll cell surface area exposed to the IAS per unit area ( _S_ _m_ ) are reported in the panel titles. The lower rows show stochastic
simulations generated using SCellPack based on five cell geometries (circular, 1-armed, 2-armed,
3-armed, and 4-armed configurations).


7


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


169 **2.2** **SCellPack** **reconstructs** **the** **2D** **leaf** **mesophyll** **cell** **area** **exposed** **to** **the** **intercellular**

170 **airspace**


171 Mesophyll cell surface area exposed to the IAS ( _S_ _m_ ), which forms the interface between the in
172 tercellular airspace and mesophyll cells, is an essential determinant of leaf structure, linking the

173 three-dimensional organization of mesophyll cells to the efficiency of _CO_ 2 diffusion and uptake.

174 &#91;2, 7, 20&#93;. Given the critical importance of _S_ _m_, this trait was our primary focus for assessing the

175 performance of SCellPack. In the standard leaf physiology measurements, the leaf surface area

176 exposed to the airspace measures the 3D surface area of the cells normalized by the leaf area ( _i.e.,_

177 see Ref. &#91;6&#93;). Since our data consist of 2D paradermal leaf segments, we measured the total ex
178 posed cell perimeter at the IAS interface and normalized it by leaf area in the same manner as for

179 the conventional _S_ _m_ measurements. We retained the conventional notation _S_ _m_ for this slice-level

180 parameter, which represents the mesophyll cell perimeter exposed to the IAS per unit leaf area.

181 We first evaluated the extent to which the SCellPack-generated reconstructions reproduced the

182 observed slice-to-slice variation in _S_ _m_ across individual paradermal slices from three representative

183 species, one from each of the three categories (i.e., honeycomb, non-honeycomb, and neither),

184 using each of the five cell geometries (Fig. 2). Considering the total variation range across the

185 three species, our results showed strong agreement between the measured and reconstructed _S_ _m_

186 for the 1-armed, 2-armed, 3-armed, and circular cell geometries across individual slices of the non
187 honeycomb species with the smallest arm length ( _i.e.,_ _Koelreuteria_ _paniculata_ ), with maximum

188 deviations of 13 _._ 0%, 24 _._ 1%, 29 _._ 3%, and 20 _._ 8%, respectively, relative to the total measured _S_ _m_

189 variation range across the three species (Fig. 2a). The 4-armed geometry exhibited the greatest

190 deviation, with a maximum value of 38 _._ 9%, although the reconstructed values remained consistent

191 with the measured variation (Fig. 2a). In contrast, the two species ( _i.e.,_ _Spinacia_ _oleracea_ and

192 _Platycerium andinum_ ) representing the other categories showed strong agreement across all slices

193 for the armed cell geometries (1-, 2-, 3-, and 4-armed), with maximum deviations of 8 _._ 9% and

194 3 _._ 7%, respectively, relative to the total _S_ _m_ variation range across the three species. However,

195 in both cases, reconstructions based on the circular geometries displayed a small but consistent

196 deviation across all slices, with maximum values of 18 _._ 9% and 8 _._ 7%, respectively (Fig. 2b,c).

197 Next, we assessed these patterns across all 40 species by averaging the measured and recon
198 structed _S_ _m_ values across all slices, yielding a single representative value for each species. We

199 quantified the reconstruction error between observed and average simulated _S_ _m_ values using the

200 absolute logarithmic relative error ( _|LRE_ ( _S_ _m_ ) _|_ ), a symmetric and scale-invariant metric that quan
201 tifies the fold deviation between simulated and observed values (see section 4.5). The analysis

202 showed strong agreement between the measured and reconstructed _S_ _m_ values across species for

203 all five cell geometries (Fig. 2d). Importantly, however, the magnitude of the reconstruction error

204 differed among cell geometries. Using a two-fold deviation, _|LRE_ ( _S_ _m_ ) _|_ _&gt;_ ln(2), as a threshold


8


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


205 for acceptable agreement, no species exceeded this threshold for the 1-armed geometry, whereas

206 two species exceeded it for each of the 2-armed, 3-armed, and 4-armed geometries. These errors

207 were concentrated in the same two species across the three mentioned armed-shape geometries.

208 By contrast, eight species exceeded the two-fold threshold for the circular geometry. Notably,

209 all eight of these species belonged to categories other than non-honeycomb, whereas the circular

210 geometry performed relatively well for non-honeycomb species (Fig. 2d–f). This observation is

211 consistent with the previous classification of non-honeycomb species based on their approximately

212 circular 2D cell shapes. Furthermore, because _D_ _arm_ _&gt;_ _L_ _arm_, the armed cell geometries also

213 closely approximated circular cells in these species, resulting in relatively short arms (see the

214 visual patterns in Fig. 1b), thereby explaining their successful reconstruction performance in non
215 honeycomb species &#91;1&#93;.

216 To obtain a summarized view of the reconstruction performance of each cell geometry across

217 all species, we analyzed the cumulative distributions of _|LRE_ ( _S_ _m_ ) _|_ for each geometry. All 40

218 species were reconstructed within a two-fold deviation for the 1-armed geometry, whereas 95%

219 of species fell within this range for each of the 2-armed, 3-armed, and 4-armed geometries. In

220 contrast, only 80% of species were reconstructed within a two-fold deviation using the circular

221 geometry (Fig. 2e–f). These results indicate that the armed cell geometries generally provided

222 more accurate reconstructions than the circular geometry, with the 1-armed geometry showing the

223 highest proportion of species reconstructed within two-fold deviation.

224 We further calculated the geometric mean fold error ( _GMFE_ ) to summarize the typical mul
225 tiplicative deviation between reconstructed and measured _S_ _m_ values across species (see Section

226 4.5). The 1-armed and 2-armed geometries showed the lowest error, with _GMFE_ = 1 _._ 15 _×_ for

227 both geometries, followed by the 3-armed (1 _._ 18 _×_ ) and 4-armed (1 _._ 23 _×_ ) geometries. By contrast,

228 the circular geometry exhibited a substantially higher error ( _GMFE_ = 1 _._ 67 _×_ ; Fig. 2g).

229 Finally, we assessed the extent to which each cell geometry captured the interspecific varia
230 tion in _S_ _m_ using the Pearson correlation coefficient for the simulated and observed values across

231 species. We found that the correlations were consistently strong and significant across all five

232 geometries, with Pearson’s _r_ ranging from 0 _._ 90 to 0 _._ 92 (Fig. 2h).

233 Taken together, the individual-slice comparisons and the across-species analyses indicate that

234 SCellPack can reproduce both the variation and magnitude of _S_ _m_ across a diverse set of species.

235 While all cell geometries showed strong correlations with the measured values, the reconstruc
236 tion accuracy depended on cell geometry, with the 1-armed and 2-armed geometries consistently

237 showing the lowest errors and the circular geometry showing the poorest overall performance.


9


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


**Figure 2:** **Performance of SCellPack for reconstruction of mesophyll cell perimeter exposed**
**to** **intercellular** **airspace** **(** _S_ _m_ **).** **a–c.** Slice-level variation in _S_ _m_ for three representative species,
comparing experimental data (black) with SCellPack simulations using five cell geometries (i.e.,
circular, 1-armed, 2-armed, 3-armed, and 4-armed), marked with different colors (see legend).
Error bars indicate the standard errors across the ten independent replicates of the simulations.
**d.** Comparison of measured and simulated mean _S_ _m_ across 40 plant species for all five cell geometries. Error bars indicate the standard errors of the slice-level values. Species are ordered by
increasing mean arm length. Blue, yellow, and brown shaded regions indicate non-honeycomb,
honeycomb, and neither spongy patterns, respectively. **e.** Absolute logarithmic relative error of
_S_ _m_ ( _|LRE_ ( _S_ _m_ ) _|_ ), across the 40 species for circular, 2-armed, and 3-armed geometries. Bars indicate the reconstruction errors for the corresponding data points of each geometry in panel **d** . The
horizontal dotted line marks a two-fold deviation, _|LRE|_ = ln(2). **f.** Cumulative distribution
functions of _|LRE_ ( _S_ _m_ ) _|_ for all five cell geometries, showing the probability that the absolute logrelative error is below a given threshold _x_ . The vertical dotted line marks a two-fold deviation,
_|LRE|_ = ln(2). **g.** Geometric mean fold error, _GMFE_ ( _S_ _m_ ), for each cell geometry across the
40 species. **h.** Pearson correlation coefficient, _r_, between simulated and observed mean _S_ _m_ values
across species for each cell geometry. Asterisks indicate statistical significance of the correlation
(*** _p_ _&lt;_ 0 _._ 001). Markers and colors corresponding to the experimental data and simulated cell
geometries are consistent across panels.


238 **2.3** **Reconstructions based on SCellPack preserve the leaf topology**


239 Beyond reproducing the values for _S_ _m_, as a geometric property, we next examined whether the

240 SCellPack-generated image reconstructions also preserved the topology of the spongy mesophyll


10


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


241 in paradermal slices. To this end, we quantified the number of connected components ( _β_ 0 ), corre
242 sponding to the zeroth Betti number, a topological invariant &#91;24, 25&#93;, in both the cell phase ( _β_ 0 _Cell_ )

243 and the IAS phase ( _β_ 0 _IAS_ ) (see section 4.4). These measures capture the connectivity of both

244 phases: the former is relevant to the organization of shared and free cell walls, whereas the latter is

245 an important determinant of _CO_ 2 transport through the IAS in the spongy mesophyll (see Section

246 1). Because the two phases are spatially complementary, connectivity in one phase also reflects, at

247 least qualitatively, the degree of fragmentation and enclosure in the other, corresponding to more

248 complex structural patterns of the spongy mesophyll.

249 The SCellPack framework showed good agreement in reconstructing the connectivity of both

250 the cell and IAS phases across the tested cell geometries (Fig. S2). For _β_ 0 _IAS_, the non-circular
251 geometries showed particularly strong agreement, with 90–92 _._ 5% of species reconstructed within

252 a two-fold deviation and _GMFE_ values ranging from 1 _._ 36 to 1 _._ 38 _×_ (Fig. S3a and Table 1). By

253 contrast, the circular geometry showed substantially lower agreement, with only 27 _._ 5% of species

254 within a two-fold deviation and a _GMFE_ of 2 _._ 86 _×_ . Consistent with these results, the non-circular

255 geometries showed strong and significant correlations between simulated and observed _β_ 0 _IAS_ val256 ues, with Pearson _r_ ranging from 0 _._ 69 to 0 _._ 72 ( _p &lt;_ 10 _&#91;−&#93;_ &#91;6&#93; ). The circular geometry showed a weaker,

257 although still significant, correlation ( _r_ = 0 _._ 55, _p_ _&lt;_ 0 _._ 001), indicating that circular geometries

258 were moderately effective at reproducing interspecific variation in IAS connectivity. The reduced

259 performance of circular geometries for _β_ 0 _IAS_ was particularly pronounced among honeycomb
260 like species. Conversely, circular-cell reconstructions showed relatively good agreement for _β_ 0 _Cell_

261 among non-honeycomb species, consistent with their structural organization &#91;1&#93; (Fig. S2).

262 For _β_ 0 _Cell_, the agreement was more variable across cell geometries, with 60–80% of species

263 reconstructed within a two-fold deviation and _GMFE_ values ranging from 1 _._ 54 to 1 _._ 91 _×_ (Fig.

264 S3b and Table 1). The 2-armed geometry showed the highest proportion of species within two
265 fold (80%) and a low _GMFE_ (1 _._ 54 _×_ ), whereas the 1-armed geometry showed the largest _GMFE_

266 (1 _._ 91 _×_ ). Nevertheless, all five geometries captured significant interspecific variation in _β_ 0 _Cell_, with

267 Pearson _r_ ranging from 0 _._ 57 to 0 _._ 78 ( _p &lt;_ 0 _._ 001). The strongest correlations were observed for the

268 4-armed ( _r_ = 0 _._ 78), 3-armed ( _r_ = 0 _._ 76), and circular ( _r_ = 0 _._ 75) geometries, whereas the 1-armed

269 geometry showed the weakest correlation ( _r_ = 0 _._ 57).

270 Overall, these results indicate that the SCellPack framework robustly reproduces connectivity

271 patterns in the leaf spongy mesophyll and captures their interspecific variation, particularly when

272 using armed cell geometries.


11


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


_β_ 0 _,Cell_ _β_ 0 _,IAS_ _N_ _Cell_ _P_ _Cell_ _A_ _Cell_

_GMFE_ r (p) _GMFE_ r (p) _GMFE_ r (p) _GMFE_ r (p) _GMFE_ r (p)


1-armed 1.91 _×_ 0.57 ( _&lt;_ 0 _._ 001) 1.38 _×_ 0.72 ( _&lt;_ 0 _._ 001) 1.61 _×_ 0.63 ( _&lt;_ 0 _._ 001) 1.18 _×_ 0.97 ( _&lt;_ 0 _._ 001) 1.58 _×_ 0.55 ( _&lt;_ 0 _._ 001)

2-armed 1.54 _×_ 0.71 ( _&lt;_ 0 _._ 001) 1.38 _×_ 0.71 ( _&lt;_ 0 _._ 001) 1.38 _×_ 0.64 ( _&lt;_ 0 _._ 001) 1.30 _×_ 0.97 ( _&lt;_ 0 _._ 001) 1.41 _×_ 0.52 ( _&lt;_ 0 _._ 001)

3-armed 1.74 _×_ 0.76 ( _&lt;_ 0 _._ 001) 1.37 _×_ 0.70 ( _&lt;_ 0 _._ 001) 1.60 _×_ 0.64 ( _&lt;_ 0 _._ 001) 1.56 _×_ 0.97 ( _&lt;_ 0 _._ 001) 1.67 _×_ 0.51 ( _&lt;_ 0 _._ 001)

4-armed 1.79 _×_ 0.78 ( _&lt;_ 0 _._ 001) 1.36 _×_ 0.69 ( _&lt;_ 0 _._ 001) 1.86 _×_ 0.64 ( _&lt;_ 0 _._ 001) 1.82 _×_ 0.96 ( _&lt;_ 0 _._ 001) 1.95 _×_ 0.50 (0 _._ 001)

Circular 1.54 _×_ 0.75 ( _&lt;_ 0 _._ 001) 2.86 _×_ 0.55 ( _&lt;_ 0 _._ 001) 1.89 _×_ 0.58 ( _&lt;_ 0 _._ 001) 1.40 _×_ 0.95 ( _&lt;_ 0 _._ 001) 1.96 _×_ 0.17 (0 _._ 296)


**Table** **1:** **SCellPack** **performance** **in** **reconstructing** **leaf** **traits.** Geometric mean fold error
( _GMFE_ ) and Pearson correlation coefficient ( _r_ ), with _p_ -values shown in parentheses, for five
leaf traits reconstructed using five cell geometries (circular, 1-armed, 2-armed, 3-armed, and 4armed) across 40 species.


273 **2.4** **SCellPack reconstructions preserve cell statistics based on watershed segmentation**


274 We next evaluated the ability of SCellPack to reconstruct single-cell statistics. Specifically, we

275 assessed the extent to which the reconstructed images reproduced the mean cell number per slice

276 ( _N_ _Cell_ ), mean cell area per leaf area ( _A_ _Cell_ ), and mean cell perimeter ( _P_ _Cell_ ) per leaf area, aver
277 aged across all paradermal slices for each species. Cell segmentation was performed using the

278 watershed algorithm applied to the binary images (see section 4.6). The accuracy of the water
279 shed segmentation was evaluated by comparing it with manual segmentation of a representative

280 grayscale paradermal slice for each species (Fig. S6). Considering the variation range across the

281 manually segmented species, watershed segmentation showed high agreement with manual seg
282 mentation at the leaf-averaged level, particularly for _A_ Cell and _P_ Cell (Fig. S7). Therefore, as all

283 subsequent analyses were performed using leaf-averaged statistics, watershed (rather than manual)

284 segmentation was used to quantify the single-cell parameters throughout this study.

285 The reconstruction of single-cell statistics showed greater variation in performance across cell

286 geometries than observed for _S_ _m_ and the topological descriptors (Fig. S4). For _N_ _Cell_, the 2
287 armed geometry showed the highest proportion of species reconstructed within a two-fold devia
288 tion (87 _._ 5%), followed by the 3-armed (85 _._ 0%) and 1-armed (80 _._ 0%) geometries (Fig. S5). The

289 corresponding _GMFE_ values were lowest for the 2-armed geometry (1 _._ 38 _×_ ), followed by the

290 3-armed (1 _._ 60 _×_ ), 1-armed (1 _._ 61 _×_ ), 4-armed (1 _._ 86 _×_ ), and circular (1 _._ 89 _×_ ) geometries (Table 1).

291 Pearson correlations between simulated and observed cell number were significant for all geome
292 tries ( _r_ = 0 _._ 58–0 _._ 64, _p_ _&lt;_ 0 _._ 001), with the strongest correlations for the 2-, 3-, and 4-armed

293 geometries ( _r_ = 0 _._ 64) and the weakest for the circular geometry ( _r_ = 0 _._ 58) (Table 1).

294 For _P_ _Cell_, the 1-armed geometry showed the strongest agreement, with all species reconstructed

295 within a two-fold deviation (100%), followed by the 2-armed (95 _._ 0%) and 3-armed (90 _._ 0%) ge
296 ometries (Fig. S5). These geometries also showed low _GMFE_ values, ranging from 1 _._ 18 _×_ for

297 the 1-armed to 1 _._ 56 _×_ for the 3-armed geometry (Table 1). Agreement remained relatively high for

298 the circular geometry (92 _._ 5% within two-fold; _GMFE_ = 1 _._ 40 _×_ ), but was substantially lower for


12


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


299 the 4-armed geometry (62 _._ 5% within two-fold; _GMFE_ = 1 _._ 82 _×_ ) (Fig. S5 and Table 1). Nev
300 ertheless, Pearson correlations were consistently strong across all geometries ( _r_ = 0 _._ 95–0 _._ 97, all

301 _p_ _&lt;_ 0 _._ 001), indicating that all geometries captured interspecific variation in mean cell perimeter

302 well (Table 1).

303 Moreover, _A_ _Cell_ showed a similar, although more variable pattern. The 2-armed geometry

304 provided the best agreement (85 _._ 0% within two-fold; _GMFE_ = 1 _._ 41 _×_ ), followed by the 3-armed

305 (82 _._ 5%; _GMFE_ = 1 _._ 67 _×_ ) and 1-armed (80 _._ 0%; _GMFE_ = 1 _._ 58 _×_ ) geometries (Fig. S5 and

306 Table 1). The 4-armed and circular geometries showed substantially worse agreement, with only

307 55 _._ 0% of species within two-fold and _GMFE_ values of 1 _._ 95 _×_ and 1 _._ 96 _×_, respectively (Fig. S5

308 and Table 1). Pearson correlations were significant for all armed geometries ( _r_ = 0 _._ 50–0 _._ 55,

309 _p ≤_ 0 _._ 001), whereas the circular geometry showed no significant correlation ( _r_ = 0 _._ 17, _p_ = 0 _._ 296)

310 (Table 1). Thus, unlike the armed geometries, the circular geometry was notably less effective

311 at reconstructing cell area and capturing its interspecific variation. However, its reconstruction

312 performance was relatively better among non-honeycomb species (Fig. S4), consistent with the

313 structural organization discussed above.

314 Overall, the armed geometries generally provided the most consistent reconstruction of single
315 cell statistics, with the 2-armed geometry showing particularly strong performance across _N_ _Cell_

316 and _A_ _Cell_, while the 1-armed geometry performed best for mean _P_ _Cell_ .


317 **2.5** **Stochastic geometric model explains the performance of SCellPack**


318 Since our developed model, SCellPack, is based on a random tessellation process, it can be related,

319 at least in part, to existing models and theoretical frameworks in stochastic geometry. The Boolean

320 model (also known as the Poisson germ–grain model) &#91;23&#93; is a particularly relevant framework

321 in this regard, given its random placement of grains (here, cells) within a defined spatial domain.

322 Specifically, the Boolean model describes a random closed set in which the expected fraction of

323 space not covered by the random set, _ρ_, is given by


_ρ_ = exp( _−λ⟨A⟩_ ) _,_ (1)


324 where _λ_ is the intensity of the homogeneous Poisson process of germ centers, corresponding to

325 the mean number of germs per unit area, and _⟨A⟩_ is the mean area of the grains (cells) &#91;23&#93;. In

326 the same line, the specific perimeter of the closed set, _i.e.,_ the boundary length of the union of the

327 grains per unit area of the underlying plane, is given by


_θ_ = _λ⟨P_ _⟩_ exp( _−λ⟨A⟩_ ) _,_ (2)


13


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


328 where _λ⟨P_ _⟩_ is the mean perimeter of the grains &#91;26&#93;. Substituting Eq. 1 into Eq. 2, and eliminating

329 _λ_, gives

_θ_ = _−_ _&#91;⟨&#93;&#91;P&#93;_ _&#91;⟩&#93;_ (3)

_⟨A⟩_ _&#91;ρ&#93;_ &#91; ln(&#93; _&#91;ρ&#93;_ &#91;)&#93; _&#91;.&#93;_


330 A major difference from our model is that cell placement is sequential in SCellPack. Neverthe
331 less, the porosity and specific perimeter of the resulting cell union are independent of the order in

332 which the cells are placed. As a result, the intercellular air-space fraction ( _f_ _IAS_ ) and the exposed

333 cell perimeter per unit area ( _S_ _m_ ) can be regarded as statistical counterparts of the Boolean model

334 quantities _ρ_ and _θ_, respectively. This correspondence motivated us to derive an analytical solution

335 for _S_ _m_ in the SCellPack model based on equation 3, using the data-based mean cell perimeter,

336 mean cell area, and air-space fraction ( _i.e.,_ porosity) as inputs.

337 However, deriving an analytical expression for _S_ _m_ in SCellPack directly from the Boolean

338 model is not straightforward because of differences in the implementation of cell placement. In

339 SCellPack, candidate cells that are completely overlapped by previously placed cells are rejected

340 (see Section 4.2), which may introduce some deviations from the assumptions of the Boolean

341 model. The major challenge, however, arises from the clipping of cells when they partially overlap

342 with previously placed cells (see Section 4.2). Clipping changes the area and perimeter of individ
343 ual cells relative to their initially sampled values and consequently modifies the resulting cell-area

344 and cell-perimeter distributions. Importantly, however, clipping does not affect the porosity or spe
345 cific perimeter of the resulting structure. This is because clipping occurs within the union of the

346 cells and therefore does not alter either the area or perimeter of the outer boundary of the union.

347 Therefore, the model can still be characterized using the mean cell area and perimeter prior to

348 clipping.

349 Accordingly, here we rely on the arm statistics rather than on the final, truncated cell sizes. The

350 initial cell sizes, prior to clipping, are directly determined by the distributions of the arm statistics.

351 For each cell geometry, the mean perimeter and area of the initial, non-truncated cells can there
352 fore be obtained analytically from the moments of the arm-length and arm-diameter distributions.

353 Specifically, one can derive _⟨A⟩_ and _⟨P_ _⟩_ from the distributions of _L_ _arm_ and _D_ _arm_, while account
354 ing for the error terms associated with the sampling of these parameters. In this line, to assess the

355 performance of the Boolean model, we derived the corresponding expressions for two fundamental

356 geometries, circular and 1-armed cells, as well as for the 2-armed 180 _◦_ geometry. The selection of

357 2-armed 180 _◦_ was motivated by its representation as a simple extension of the 1-armed cell, while

358 consistently showing excellent agreement with the primary 2-armed geometry (characterized by an

359 arm angle of 120 _&#91;◦&#93;_ ) in reconstructing the previously assessed parameters (see Fig. S2 and Fig. S4).

360 These analytical derivations are presented in Section 4.7. Analytical solutions for more complex

361 cell geometries, involving additional arms and different arm angles, require further extensions of


14


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


362 these relations. A detailed derivation for such geometries is left for future work.

363 The formulation of the Boolean model, based on the initial cell sizes rather than the final

364 truncated sizes, also ensures that the Boolean model uses exactly the same inputs as SCellPack,

365 _i.e._, _L_ _arm_ _±_ _σ_ _L_ _arm_, _D_ _arm_, and _f_ _IAS_ . We refer to the values of _ρ_ and _θ_ calculated from these

366 analytical solutions as the theoretical _f_ _IAS_ and _S_ _m_, respectively (see above).

367 Importantly, _S_ _m_ calculated from the Boolean model showed excellent agreement with the cor
368 responding values obtained from the SCellPack simulations across all three cell-shape geometries

369 (Fig. S8 and Fig. 3). For all 40 species (100%), the theoretical values of _S_ _m_ were within two-fold

370 of the SCellPack results for each geometry. The agreement was further supported by highly sig
371 nificant Pearson correlations for the circular ( _r_ = 0 _._ 998), 1-armed ( _r_ = 0 _._ 997), and 2-armed 180 _◦_

372 ( _r_ = 0 _._ 997) geometries. Consistent with these strong correlations, the _GMFE_ error values were

373 very low, at 1 _._ 03 _×_, 1 _._ 04 _×_, and 1 _._ 06 _×_ for the 1-armed, 2-armed, and 2-armed 180 _◦_ geometries, re
374 spectively (Fig. 3). Thus, the Boolean model predictions closely reproduced the SCellPack-derived

375 _S_ _m_ values across all three geometries, and the deviations between theory and data closely followed

376 the same patterns as the deviations between the SCellPack-derived _S_ _m_ values and the empirical

377 data described above (Section 2.2; Fig. S8 and Fig. 3).

378 Together, this formulation analytically links cell arm statistics and leaf porosity to cell wall

379 properties, providing a simple stochastic-geometric framework for describing how these structural

380 features shape leaf architecture as captured by SCellPack.


15


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


**Figure** **3:** **Performance** **of** **the** **Boolean** **model** **in** **estimating** **mesophyll** **cell** **surface** **area** **ex-**
**posed to intercellular airspace (** _S_ _m_ **).** Cumulative distribution functions of the absolute logarithmic relative error _|LRE|_ between: **a.** theoretical Boolean model and SCellPack simulated values
of _S_ _m_, and **b.** theoretical Boolean model and data measured values of _S_ _m_ . Colors indicate different
cell geometries (circular, 1-armed, 2-armed 180 _◦_ ). The vertical dashed lines mark _|LRE|_ = ln(2),
corresponding to 2 _×_ fold-error threshold.


381 **2.6** **The economical geometry for optimal cell-airspace gas-exchange**


382 Beyond the importance of _S_ _m_ for leaf _CO_ 2 conductance, the balance between maximizing free cell

383 wall, as captured by _S_ _m_, and minimizing shared cell walls is another aspect that plays an essential

384 role in determining the structural costs associated with cell wall construction (see section 1). To

385 quantify this cost trade-off, we define the loss fraction as



_f_ _loss_ = 1 _−_ _&#91;S&#93;_ _&#91;m&#93;_

_S_ _w_



_−_ _&#91;S&#93;_ _&#91;b&#93;_

_S_ _w_



_,_ (4)



386 where _S_ _b_ and _S_ _w_ denote the cell perimeter exposed to the image border and the total cell wall

387 perimeter, respectively, both normalized by leaf area, analogous to _S_ _m_ . We note that _S_ _b_ introduces

388 a deviation from the real biological system due to the artificial image boundaries, and its relative

389 contribution diminishes as the image size increases. The defined quantity, _f_ _loss_, represents the

390 fraction of the cell wall that does not contribute to IAS gas exchange because it is located at

391 cell–cell interfaces, and therefore constitutes an additional structural cost of cell wall construction.

392 Based on the Boolean model, using _S_ _b_ = 0, _S_ _w_ = _λ⟨P_ _⟩_, along with equations 1, 2, and 3,


16


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


393 while treating _ρ_, and _θ_ as the counterparts of _f_ _IAS_, and _S_ _m_, respectively, we obtain


_f_ _loss_ = 1 _−_ _f_ _IAS_ _._ (5)


394 This relation indicates a perfect linear trade-off between leaf porosity and the fraction of shared

395 cell walls predicted by the Boolean model (see Section 2.5).

396 Whilst the Boolean model treats overlapping regions between cells as valid cell area, SCell
397 Pack prevents cell overlap, consistent with biological constraints. By definition, this results in

398 shorter shared wall lengths due to cell truncation, _i.e._, a smaller _S_ _w_, while maintaining consistent

399 _S_ _m_ . Interestingly, we observed an excellent correlation between _f_ _loss_ and ( _f_ _IAS_ ) across the SCell
400 Pack reconstructions generated using the different cell geometries (Fig. 4a-e). Specifically, the

401 SCellPack reconstructions exhibited very strong relationships, with coefficients of determination
402 consistently high across all cell geometries ( _R_ &#91;2&#93; _≥_ 0 _._ 96) (Fig. 4a-e). The fitted relationships were

403 also remarkably consistent across geometries, with slopes of _−_ 0 _._ 88, _−_ 0 _._ 88, _−_ 0 _._ 87, _−_ 0 _._ 86, and

404 _−_ 0 _._ 87 for the 1-, 2-, 3-, 4-armed, and circular geometries, respectively (Fig. 4a-e). These findings

405 suggest that, despite differences in cell geometry, the SCellPack reconstructions are in line with

406 the linear trade-off between _f_ _loss_ and _f_ _IAS_, predicted by the Boolean model.

407 Similarly, we observed a moderate to strong relationship between _f_ _loss_ and _f_ _IAS_ in the mea
408 sured data across all as well as only for the honeycomb species. Specifically, the measured data
409 across all 40 species showed a coefficient of determination of _R_ &#91;2&#93; = 0 _._ 46, with a substantially

410 shallower slope of _−_ 0 _._ 51 and an intercept of 0 _._ 50 (Fig. 4f). Restricting the analysis to honeycomb
411 species strengthened the relationship, yielded _R_ &#91;2&#93; = 0 _._ 61, with a similar slope of _−_ 0 _._ 48. These re
412 sults indicate a more pronounced deviation from the ideal linear trade-off observed in the Boolean

413 model for the measured data, while highlighting _f_ _IAS_ as an important determinant of the costs

414 associated with cell wall construction.

415 The higher slopes and intercepts observed for the SCellPack reconstructions indicate that this

416 framework remains closer to the assumptions of the Boolean model than the measured data in

417 terms of the total cell perimeter. This is the case since both predict a substantially larger total cell

418 perimeter relative to _S_ _m_ (Eq. 4). One possible explanation for this observation is that SCellPack

419 considers even very small non-overlapping portions of candidate cells as valid cells, which can

420 generate additional small cells and increase the total cell perimeter (see Section 4.2). To test this

421 hypothesis, we introduced an additional constraint through the clipping fraction parameter ( _T_ ),

422 which defines the upper-bound fraction of a candidate cell that can be truncated by previously

423 placed cells (see Section 4.8).

424 We tested and compared _T_ values in the range &#91;0 _._ 1 _,_ 1 _._ 0&#93; with a step of 0 _._ 1, where _T_ = 1 _._ 0

425 corresponds to the default SCellPack setting, with no truncation constraint except rejection of


17


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


426 candidate cells with complete overlap (sections 4.2 and 4.8). The SCellPack-generated results
427 showed strong linear relationships across all tested _T_ values, with _R_ &#91;2&#93; _≥_ 0 _._ 87 at _T_ = 0 _._ 1 and
428 _R_ &#91;2&#93; _≥_ 0 _._ 95 for _T_ _&gt;_ 0 _._ 1 (Fig. S9). Importantly, both the slope and intercept of the linear _f_ _loss_ - _f_ _IAS_

429 fits decreased systematically with decreasing _T_, following similar trends across cell geometries

430 (Fig. 4g-k and Fig. S9). Moreover, depending on cell geometry, the SCellPack-fitted relationships

431 most closely matched the measured data around _T_ = 0 _._ 3 _−_ 0 _._ 4 (Fig. 4g-k and Fig. S9).

432 To assess the ability of SCellPack to reconstruct leaf structural parameters using the _T_ value

433 that provided the best agreement with the experimental _f_ _loss_ – _f_ _IAS_ relationship, we next evaluated

434 all previously assessed traits, namely _S_ _m_, _β_ 0 _,IAS_, _β_ 0 _,Cell_, _N_ _Cell_, _P_ _Cell_, and _A_ _Cell_, using SCellPack

435 configured with _T_ = 0 _._ 4. Among the cell geometries, the 2-armed cells consistently showed com
436 parable or even higher reconstruction accuracy across species for all traits relative to the primary

437 cases with _T_ = 1, whereas the other geometries showed notable declines in the reconstruction

438 for one or more traits (see Fig. S10). These results support the relevance of incorporating the _T_

439 constraint into SCellPack, particularly for specific cell geometries and structural traits.

440 In summary, we showed that _f_ _loss_ is negatively correlated with _f_ _IAS_ across plant species. This

441 indicates that plant species exhibiting low porosity incur higher cell wall construction costs relative

442 to the free cell wall perimeter available for _CO_ 2 transport. We consistently observed this relation
443 ship across the Boolean model, the SCellPack framework, and the measured data, although the

444 measured data exhibited a notably less steep decline in _f_ _loss_ with increasing _f_ _IAS_ than predicted

445 by either model. We further demonstrated that varying the truncation fraction _T_ changes both the

446 slope and intercept of the resulting linear fits.

447 Specifically, we observe closely matching patterns in the dependence of the linear fits on _T_

448 across all cell-shape geometries (Fig. 4g-k, S9). Varying the value of _T_ modifies the leaf-averaged

449 cell perimeter and area distributions relative to those of the random closed set represented by the

450 Boolean model, while producing similar responses across different cell-shape geometries. There
451 fore, these results suggest that cell size and area distributions are stronger determinants of cell
452 construction costs than cell-shape geometry. Together, these findings provide a basis for identi
453 fying potential economical strategies to modulate crop improvement by optimizing the balance

454 between leaf _f_ _IAS_ and cell-size distributions, thereby maximizing _CO_ 2 uptake while minimizing

455 the costs of cell wall construction.


18


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


**Figure** **4:** **Association** **between** **leaf** **porosity** **and** **cell** **shared** **wall** **fraction.** Linear fits of the
shared-wall fraction ( _f_ _loss_ ) as a function of the intercellular airspace fraction ( _f_ _IAS_ ) for **a.** 1-armed,
**b.** 2-armed, **c.** 3-armed, **d.** 4-armed, and **e.** circular cell geometries, and **f.** measured data. The red
data points and red lines in panels **a-e** represent the SCellPack results at the truncation fraction of
_T_ = 1, while the black data points and black fit in panel **f** represent the measured data; the gray data
points and fit in panel **f** show the subset of honeycomb species. The lower panels **g-k** correspond
to panels **a-e**, respectively, and are aligned in the same columns, showing the corresponding linear
fits across the range of _T_ tested in SCellPack. In panels **g-k**, colored lines represent the linear fit
for each truncation fraction, the black line represents the fit to the measured data, and the dashed
line indicates the ideal trade-off predicted by the Boolean model. The vertical and horizontal error
bars in panels **a-f** represent the standard errors of the mean values across paradermal slices for _f_ _loss_
and _f_ _IAS_, respectively. In the upper panels, _R_ &#91;2&#93;, _α_, and _m_ denote the coefficient of determination,
slope, and intercept of the linear fits, respectively.

### 456 3 Discussion


457 Recent advances in leaf imaging techniques have improved our understanding of the internal struc
458 ture of the leaf mesophyll. However, the general mechanisms governing the development and

459 maintenance of mesophyll structure remain poorly understood. In particular, the spongy meso
460 phyll exhibits remarkable diversity in the spatial organization of cells and intercellular airspaces

461 (IAS) across vascular plant species, yet the mechanisms underlying the formation and structural

462 stability of these distinct architectures remain largely unresolved &#91;1, 8&#93;.

463 At the same time, improving _CO_ 2 uptake through mesophyll conductance while minimizing

464 the resource costs associated with cell wall construction necessitates a mechanistic understand
465 ing of how mesophyll architecture influences these functions. Motivated by this challenge, we


19


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


466 developed SCellPack framework, based on stochastic space tessellation, for simulating the two
467 dimensional architecture of the leaf spongy mesophyll across species. Specifically, SCellPack

468 aims to reconstruct the binary structure, comprising either cell or intercellular airspace phases,

469 of two-dimensional paradermal leaf sections derived from three-dimensional microCT volumes

470 using a small number of data-derived parameters. Building on previous work that categorized

471 the spectrum of spongy mesophyll phenotype into three primary classes based on cell-arm length

472 and diameter &#91;1&#93;, we parameterized SCellPack using a single slice-level parameter, leaf poros
473 ity ( _f_ _IAS_ ), and two leaf-level parameters: cell-arm length ( _L_ _arm_ ) and cell-arm diameter ( _D_ _arm_ ).

474 This parsimonious parameterization provides a key advantage in terms of model simplicity com
475 pared with similar simulation frameworks, which generally involve a larger set of parameters ( _e.g.,_

476 &#91;10, 18, 19&#93;).

477 Motivated by the study of Borsuk et al. (2022) &#91;1&#93;, we configured five cell geometries within

478 the SCellPack framework. In that study, species were categorized according to circular (in 2D) and

479 triply armed cell-shape geometries, which were associated with non-honeycomb and honeycomb
480 like architectures, respectively, as well as a third category comprising cells with radially elongated

481 arms that conform to neither geometry. Building on this classification, we considered the circular

482 geometry and extended the triply armed cell geometry to include both fewer and greater numbers

483 of arms, considering 1-, 2-, 3-, and 4-armed cell geometries.

484 The simulation procedure in SCellPack involves truncating candidate cells such that they oc
485 cupy only the available space on their free-wall sides, while the shared walls formed with previ
486 ously placed cells remain fixed (see section 4.2). This procedure partially captures, at a coarse
487 grained level, expansigeny, a widely accepted mechanism proposed to underlie IAS formation

488 &#91;8, 11&#93;. In particular, the truncation criterion imposes a directional constraint on cell expansion

489 that is effectively consistent with the proposed role of expansigeny in shaping the IAS. Thus, al
490 though the framework represents the static architecture of mature mesophyll, the resulting cell

491 configurations may effectively provide insights into the mechanical constraints associated with

492 IAS formation and organization.

493 To assess the performance of SCellPack, we focused primarily on the cell surface perimeter

494 exposed to the IAS in 2D ( _S_ _m_ ), given its critical role in mesophyll conductance ( _g_ _m_ ) as the inter
495 face for _CO_ 2 transfer from the IAS into the cell, as well as its implications for the costs associated

496 with cell wall construction (see section 1). Our results showed excellent agreement between exper
497 imentally measured and SCellPack-reconstructed _S_ _m_ values at both the slice and leaf levels (Fig. 2

498 and section 2.2). At the leaf level, the performance of SCellPack was assessed using three comple
499 mentary metrics: the percentage of species for which the reconstructed _S_ _m_ was within a two-fold

500 error of the measured value (80% for circular and 95% for armed cell-shape species), the geometric

501 mean fold error ( _GMFE_ = 1 _._ 67 _×_ for circular and _GMFE_ _≤_ 1 _._ 23 _×_ for armed cells), and the


20


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


502 Pearson correlation between measured and simulated values across species ( _r_ _≥_ 0 _._ 9 for all cell

503 geometries) (Fig. 2 and section 2.2). Together, these metrics provided complementary evidence

504 of the high representativeness of SCellPack simulations: the high percentage of species predicted

505 within a two-fold error together with the low _GMFE_ indicates biologically meaningful simu
506 lations at the individual-species level. Moreover, the high Pearson correlation demonstrates the

507 ability of SCellPack to reproduce interspecific variation, independently of the absolute accuracy of

508 individual species-level realizations resulting from SCellPack. Therefore, while simulations from

509 SCellPack result in different realizations, they are highly representative of the analyzed mesophyll

510 architectures.

511 We next found that SCellPack consistently achieved good reconstruction accuracy across the

512 majority of cases, with performance varying among cell geometries and traits, for the reconstruc
513 tion of two topological traits (Section 2.3 and Fig. S2) and three cell-level traits (section 2.4 and

514 Fig. S3). The ability to preserve topological traits ensures that the connectivity of the IAS, which is

515 essential for _CO_ 2 diffusion, and the connectivity of cell objects, which is relevant to the regulation

516 of shared and free cell walls, are maintained in the images generated by SCellPack. In addition,

517 the accurate reconstruction of cell-level statistics, _i.e._, mean perimeter, area, and cell number, im
518 portantly, indicates the biological relevance of cell truncation during cell tessellation in SCellPack

519 framework, in the leaf average level. Overall, the ability of SCellPack to reconstruct leaf parader
520 mal images while preserving six key biological, topological, and anatomical traits demonstrates its

521 capacity to comprehensively reproduce the spatial organization of leaf paradermal slices across a

522 broad range of species representing all previously identified cell-shape categories.

523 Among the five tested cell geometries, the 1-, 2-, and 3-armed geometries consistently ranked

524 among the best-performing configurations across all the species (sections 1, 2.3, and 2.4). In

525 contrast, circular cells performed acceptably primarily for the 11 non-honeycomb species, while

526 in the majority of cases they failed to adequately reconstruct the slices for the remaining species

527 (Figs. 2, S2, and S4). These findings are consistent with the hypothesis proposed by Borsuk et

528 al. (2022) &#91;1&#93;, in which circular cell geometries in 2D were considered the geometric endpoint of

529 triply armed cells as _D_ _arm_ _&gt; L_ _arm_ . This property is also incorporated into our framework (see Fig.

530 1), ensuring that armed cells effectively converge towards a circular geometry under this condition,

531 whereas the converse does not hold.

532 We next showed that SCellPack can in part be captured by the Boolean model, a theoretical

533 framework from stochastic geometry &#91;23&#93;. Specifically, we derived theoretical estimates of _S_ _m_

534 from analytical expressions for the mean cell area and perimeter, parameterized by the empirically

535 measured _L_ _arm_ and _D_ _arm_ for each cell geometry (section 2.5). Our theoretical model for _S_ _m_ is a

536 function of _f_ _IAS_, _L_ _arm_ (and its standard deviation, _σ_ _L_ _arm_ ), and _D_ _arm_ (equation 3), the same set of

537 parameters used as inputs to SCellPack. Importantly, we observed excellent agreement between


21


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


538 the Boolean model estimates and SCellPack simulations, which in turn showed high agreement

539 with the experimental measurements for three representative cell geometries (Fig. 3), demonstrat
540 ing the validity of the theoretical formulation and its consistency with the underlying SCellPack

541 framework.

542 The analytical expressions in our extension of the Boolean model describe the uniform random

543 distributions of candidate cell area and perimeter, parameterized by _L_ _arm_ _±σ_ _L_ _arm_ and _D_ _arm_, equiv
544 alent to the cell distributions generated by SCellPack prior to their truncation during the tessella
545 tion procedure. Thus, these distributions do not correspond directly to the realized distributions

546 of cell area and perimeter after tessellation, nor to those measured from segmented experimen
547 tal images. Nevertheless, because _L_ _arm_ and _D_ _arm_ are directly measured from experimental data,

548 the high accuracy of the _S_ _m_ reconstructions obtained with the Boolean model indicates that these

549 pre-tessellation distributions retain biological relevance. Given that cell overlap is permitted in the

550 Boolean model, these distributions may be interpreted as representing the idealized cell size and

551 geometry under unconstrained growth, in the absence of the shared cell wall constraints that arise

552 through expansigeny.

553 The stability of the spongy mesophyll architecture represents another poorly understood aspect

554 of spongy mesophyll structure &#91;1, 10, 11&#93;. In particular, the structural robustness of the spongy

555 mesophyll despite the presence of a substantial fraction of the mesophyll as void space ( _f_ _IAS_ =

556 0 _._ 24–0 _._ 67 in our dataset; see Fig. S1 for details) highlights the need to understand how porosity

557 is regulated by cell structural parameters. Here, our theoretical model provides a mechanistic

558 framework for understanding the regulation of _f_ _IAS_, indicating that the coordinated regulation of

559 _S_ _m_, _L_ _arm_, and _D_ _arm_ contributes to the maintenance of this architecture, as described by Eq. 3.

560 Finally, we assessed the proportion of shared cell wall relative to the total cell wall ( _f_ _loss_ ) as a

561 measure of the structural investment in cell wall construction that is unavailable for _CO_ 2 exchange

562 between the IAS and the cell interior (section 2.6). For the idealized case described by the Boolean

563 model, we showed a perfect linear trade-off between the two fraction parameters, _f_ _loss_ and _f_ _IAS_

564 ( _f_ _loss_ = 1 _−_ _f_ _IAS_ ), noting that the model considers the perimeter of overlapping cell regions

565 as shared cell walls (section 2.6). While accounting for the biologically relevant shared walls

566 of truncated cells in SCellPack, simulations nevertheless showed an excellent linear relationship

567 between _f_ _loss_ and _f_ _IAS_, with only small deviations in the slope and intercept relative to the ideal

568 trade-off predicted by the Boolean model (Fig. 4a-e). Importantly, the experimental data also

569 showed a moderate (strong) relationship across all (honeycomb-like) species, between _f_ _loss_ and

570 _f_ _IAS_, indicating that _f_ _loss_ is closely associated with leaf porosity partuculrly for the honeycomb
571 like species (Fig. 4f). Specifically, the results suggest that species with lower porosity exhibit

572 greater cost-effectiveness in cell wall construction compared with species characterized by a higher

573 proportion of IAS.


22


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


574 Despite exhibiting the same _f_ _loss_ – _f_ _IAS_ correlation trend, the slopes and intercepts of the ex
575 perimental linear fits were substantially lower than those obtained from the SCellPack framework

576 (Fig. 4). This discrepancy indicates that, in real tissues, the maximum attainable _f_ _loss_ is lower and

577 that _f_ _loss_ is less sensitive to variation in _f_ _IAS_ across species than obtained by the SCellPack simu
578 lations. Toward strategies for the optimization of the costs associated with cell wall construction as

579 a function of leaf internal architecture, we explored alternative distributions of truncated cell sizes

580 and shapes in SCellPack. Specifically, we introduced an upper-bound truncation fraction, _T_, rang
581 ing from 0 to 1, where lower values of _T_ shift the cell population toward its ideal, non-truncated

582 geometry (sections 2.6 and 4.8). This approach modifies the spatial distribution of cells from the

583 uniform Poisson distribution assumed in the Boolean model &#91;23&#93;, thereby allowing us to examine

584 how variations in cell size, shape, and spatial organization influence the costs associated with cell

585 wall construction.

586 We observed that decreasing _T_ systematically altered the _f_ _loss_ – _f_ _IAS_ relationship, with both

587 the slope and intercept gradually decreasing until converging with the experimental data fits at

588 _T_ _∼_ 0 _._ 3–0 _._ 4 across all cell geometries (see Fig. 4g-k and Fig. S9). This suggests that the real

589 data sets retain approximately 60–70% of the idealized geometrical characteristics defined by the

590 cell-arm statistics. In addition, the linear fits obtained from the SCellPack simulations across dif
591 ferent _T_ values exhibited a consistent trend across all five cell geometries (Fig. 4g-k and Fig.

592 S9). This consistency suggests that the distribution of cell-arm statistics plays a dominant role

593 in determining the associated construction costs, whereas variations in mean cell area, perimeter,

594 and overall cell morphology appear to have a comparatively smaller effect across the different

595 cell geometries. Together, these findings suggest that the regulation of leaf porosity with final

596 cell arm statistics determines the degree of cost inefficiency associated with cell wall construc
597 tion. Moreover, these findings provide potential pathways for cost–benefit optimization of leaf

598 internal architecture across different cell geometries, with the aim of enhancing _CO_ 2 uptake while

599 minimizing the costs associated with cell wall construction.

600 While SCellPack and the theoretical framework presented here advance our understanding of

601 the geometrical principles underlying leaf tissue architecture and morphogenesis, extending the

602 approach to three-dimensional imaging represents a complementary avenue for future research.

603 The current 2D model could also benefit from incorporating slice-level arm statistics instead of

604 the leaf-level statistics available for and used in the present analysis, potentially improving recon
605 struction accuracy. Furthermore, both the 2D and 3D models could benefit from inclusion of the

606 spatial distribution of stomatal pores. In particular, sub-stomatal cavities require IAS voids that are

607 connected to the rest of the IAS and have been shown to influence mesophyll and IAS architec
608 ture &#91;27, 28&#93;. Extending the formulation to more generalized non-homogeneous Boolean process

609 &#91;23&#93;, and other similar sequential space tessellation approaches such as the dead leaves model


23


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


610 &#91;29, 30, 31&#93; is another direction that could improve the theoretical explanation of the mesophyll

611 structural patterns. Finally, developing predictive models for porosity and cell-arm statistics from

612 available species data sets could enable their prediction for unseen species and genotypes, which,

613 when integrated into SCellPack, could facilitate the reconstruction of leaf mesophyll architecture

614 across previously uncharacterized species and genotypes.

### 615 4 Methods


616 **4.1** **Data set**


617 The data set used in this study consists of X-ray micro-computed tomography (microCT) images

618 of the spongy mesophyll from 40 plant species, originally published by Borsuk _et_ _al._ (2022)

619 &#91;1&#93;. The data set encompasses plant species from 30 genera and 24 families, representing a broad

620 diversity of extant vascular plants &#91;1&#93;. Leaf tissue samples, approximately 1 _._ 5-2 mm wide and 15

621 mm long, were excised near the leaf midpoint, offset 5-10 mm from the mid vein, and imaged

622 at the Advanced Light Source (ALS, Lawrence Berkeley National Laboratory) and the TOMCAT

623 beamline of the Swiss Light Source (SLS). The reconstructed image stacks have pixel resolutions

624 ranging from 0 _._ 1625 to 1 _._ 277 _µ_ m.

625 We used the two-dimensional paradermal slices extracted from the three-dimensional microCT

626 reconstructions together with the accompanying species-level metadata. These metadata include

627 the pixel resolution and the mean and standard deviation of the cell arm length ( _L_ _arm_ ) and arm

628 diameter ( _D_ _arm_ ), which were used to parameterize our analyses.

629 Each species is represented by a leaf tissue sample with different image dimensions and num
630 bers of available paradermal slices through the spongy mesophyll. To preserve the original struc
631 tural information, all analyses and simulations were performed using the native image size and

632 spatial resolution of each species. When simulations required varying the analysis domain or tis
633 sue depth, these parameters were selected within the intervals defined by the original data set,

634 relative to the dimensions of the imported image stack.


635 **4.2** **Simulation**


636 Synthetic paradermal binary images were simulated for the complete data set of 40 species, men
637 tioned above. Because species differed in both the dimensions of the paradermal 2D rectangular

638 box and the number of available paradermal slices, simulations were performed independently for

639 each slice using its original image dimensions, ensuring that all geometric measurements remained

640 directly comparable to the original microCT data.

641 Five idealized cell morphologies were considered: circular cells and cells with one, two, three,


24


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


642 or four protruding arms (hereafter referred to as 1-armed, 2-armed, 3-armed, and 4-armed cells,

643 respectively). The characteristic cell size was sampled from the experimentally measured distri
644 bution for each species. Specifically, arm length ( _A_ _arm_ ) and arm diameter ( _D_ _arm_ ) were indepen
645 dently sampled from normal distributions defined by the measured mean and standard deviation

646 for the corresponding species at the leaf level, as provided in Ref. &#91;1&#93;. For the simulations, these

647 geometric parameters were expressed relative to the simulation box size, thereby preserving the ex
648 perimentally observed proportions between cell arm dimensions and tissue image dimensions for

649 all species. Each armed cell was constructed as the union of buffered line segments. Specifically,

650 each arm was represented by a line segment of length _L_ _arm_ originating from a common center

651 and expanded by a circular buffer of radius _D_ _arm_ _/_ 2, corresponding to the Minkowski sum of the
652 line segment with a circle of radius _D_ _arm_ _/_ 2. The circular buffer was numerically approximated

653 using 32 segments per quarter circle (Python package _Shapely_ version 2 _._ 1 _._ 2, buffer function).

654 Similarly, circular cells were generated by buffering a point at the cell center with radius _L_ _arm_ .

655 This construction produces rounded arm termini and smoothly connected junctions through the

656 union operation, resulting in geometries with well-defined areas and perimeters determined by

657 _L_ _arm_, _D_ _arm_, and arm orientations. When _D_ _arm_ _&gt;_ _L_ _arm_, the armed morphologies increasingly

658 approached a circular shape. The angles between arms for the 2-, 3-, and 4-armed cell geome
659 tries were fixed at 120 _&#91;◦&#93;_, 120 _&#91;◦&#93;_, and 90 _&#91;◦&#93;_, respectively. To assess the effect of arm angle, 2-armed

660 cells with an angle of 180 _&#91;◦&#93;_ were also simulated and referred to as 2-armed 180 _◦_ throughout the text.

661 Additionally, each cell was assigned a random orientation during placement.

662 The simulation was based on a sequential random space-tessellation process in which cells were

663 generated and packed within a rectangular simulation domain. The dimensions of the simulation

664 box were defined from the corresponding image slice dimensions. To maintain consistent scaling

665 between cell size and box size across image slices, the experimentally measured cell arm lengths

666 were normalized by the diagonal of the image slice array, expressed in physical units. Specifically,

667 for each species, the mean ( _L_ _arm_ ) and standard deviation ( _σ_ _L_ ) of the arm length were converted to

668 dimensionless quantities as



_C_ = _&#91;L&#93;_ _&#91;arm&#93;_

_d_ _box_



_,_ _σ_ _C_ = _d_ _&#91;σ&#93;_ _box_ _&#91;L&#93;_



_,_



669 where _d_ _box_ is the diagonal of the rectangular simulation box. The arm length of each candidate

670 cell was then defined as _c × d_ _box_, where _c_ was sampled uniformly from the range _C_ _± σ_ _C_ . To

671 avoid inconsistencies between the sampled cell arm length and arm diameter for an individual cell,

672 the arm diameter was not sampled independently. Instead, for each species, the ratio between the

673 mean arm diameter and mean arm length was calculated as _f_ _DL_ = _&#91;D&#93;_ _L_ _arm_ _&#91;arm&#93;_ &#91;. For each sampled cell, the&#93;
674 arm diameter was subsequently calculated from its sampled arm length using the scaling parameter

675 _f_ _DL_ . Thus, cells with different sampled arm lengths retained the same ratio of _D_ _arm_ _/L_ _arm_ within a


25


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


676 given species. Each candidate cell, with a randomly sampled size and morphology, was generated

677 at a uniformly random position within the simulation box and tested against previously accepted

678 cells. If the candidate cell overlapped previously accepted cells, the overlapping region was clipped

679 against the boundaries of the existing cells, while the remaining free boundary of the candidate cell

680 was left unconstrained. In some cases, a portion of a candidate cell that did not directly overlap

681 an existing cell could nevertheless become disconnected from the largest remaining portion as a

682 consequence of overlaps with other cells. Such disconnected regions were treated as effectively

683 overlapping areas and removed accordingly. If one or more existing cells completely covered the

684 candidate cell, or if the candidate cell completely covered one or more previously accepted cells,

685 the trial was rejected. Cells intersecting the simulation-box boundary were similarly clipped, using

686 the same area-retention criterion as for cell-cell overlap.

687 The porosity ( _f_ _IAS_ ), defined as the fraction of empty area relative to the total image area, was

688 prescribed as the stopping criterion for the cell placement process. Cell placement continued until

689 the accumulated cell area reached the target porosity, at which point the simulation was terminated.

690 Simulations were additionally stopped when a predefined maximum number ( _i.e.,_ 200) of consec
691 utive unsuccessful placement attempts was reached, indicating that further cell insertion was not

692 possible under the overlap constraint. This deviation from the target porosity was considered the

693 simulation error, particularly affecting images with low porosity. In all simulations, an ensemble

694 of 10 independent realizations was generated for each species–cell morphology combination. En
695 semble statistics were calculated from these realizations to obtain the mean structural properties

696 and their associated variability, represented by the mean standard deviation across realizations.

697 All structural statistics of the simulated slices were calculated after converting the _Shapely_ cell

698 geometries into raster images. Each cell was assigned a unique integer label, with the background

699 assigned zero. The labeled images were analyzed using the same image-based procedure applied

700 to the experimental slices. Specifically, the parameters were initially obtained in raster-pixel units

701 and then converted to the simulation scale using the raster pixel scale. All the parameters with

702 a physical length unit are expressed based on the physical length in milimeters by converting the

703 images based on their pixel size.


704 **4.3** **Cell surface area exposed to the intercellular airspace**


705 In the standard leaf physiology measurements, the leaf surface area exposed to the airspace is

706 normalized by the leaf area ( _i.e.,_ see Ref. &#91;6&#93;). Since here our data are leaf 2D paradermal seg
707 ments, we measure the total exposed cell perimeter exposed to the airspace ( _S_ _m_ ) with the same

708 normalization to the leaf area. In these measurements, the cell perimeters were computed using a

709 4-connected neighborhood with the _scikit-image_ Python package (version 0 _._ 26 _._ 0) &#91;32&#93;, using the

710 measure.perimeter() function.


26


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


711 **4.4** **Topological parameters**


712 The number of connected components is a topological descriptor of a binary image, correspond
713 ing to the zeroth Betti number in topological terms &#91;24, 25&#93;. In two dimensions, it is defined

714 as the number of disconnected objects. Accordingly, two topological descriptors were analyzed:

715 the number of cell connected components ( _β_ 0 Cell ) and the number of intercellular airspace (IAS)

716 connected components ( _β_ 0 IAS ). Both phases were labeled using a 4-connected neighborhood, com
717 puted with the _SciPy_ Python package (version 1 _._ 16 _._ 3) &#91;33&#93;. This analysis was performed separately

718 for the cell and IAS phases of both the measured and reconstructed paradermal slices.


719 **4.5** **Reconstruction error metrics**


720 To quantify the deviation between observed, simulated, and theoretically calculated trait values

721 across species, we used the absolute Logarithmic Relative Error ( _|LRE|_ ), defined as



_|_ = _|_ ln( _Y_ _est_ ) _−_ ln( _Y_ _ref_ ) _|,_

)︃



_|LRE|_ = _|_ ln



_Y_ _est_


_Y_

(︃ _ref_



722 where _Y_ _est_ and _Y_ _ref_ denote the estimated and reference trait values, respectively. This metric quan
723 tifies the magnitude of the deviation between estimated and reference values on a multiplicative

724 scale, with _|LRE|_ = 0 indicating perfect agreement. Throughout the text, we express the devia725 tion as the equivalent fold difference, exp( _|LRE|_ ), denoted by _×_ . Specifically, we use a two-fold
726 difference, corresponding to _|LRE|_ = 0 _._ 693 and exp(0 _._ 693) _≈_ 2 _×_, as a threshold for assessing

727 reconstruction performance. This threshold indicates that the estimated value is either twice or half

728 the reference value. A key advantage of _|LRE|_ for model assessment is its scale invariance. This

729 property is particularly relevant for image-derived parameters, whose values may depend on the

730 choice of spatial units (e.g., pixels versus physical dimensions).

731 To summarize the _|LRE|_ across species, we used the geometric mean fold error ( _GMFE_ ),

732 defined as



(︄



)︄



_n_
∑︂


_i_ =1



GMFE = exp



1

_n_



_|LRE_ _i_ _|_



_,_



733 where _i_ indexes species and _n_ is the number of species. The _GMFE_ ranges from 1 to _∞_, with

734 values closer to 1 indicating more accurate reconstructions and larger values indicating greater

735 typical fold deviation between estimated and reference values.


736 **4.6** **Cell segmentation**


737 Cell segmentation was performed using a watershed-based procedure in ImageJ &#91;34&#93; to identify

738 cell boundaries (Fig. S6). Because watershed segmentation removed the cell boundary separating


27


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


739 adjacent cells, cell areas and perimeters derived from this procedure included a small associated

740 approximation. To validate the slice-level performance of the watershed segmentation, we man
741 ually corrected the segmentation results for a representative grayscale paradermal slice of each

742 species. Specifically, we first manually segmented the intercellular airspace in the grayscale im
743 ages and then applied the watershed algorithm to the corresponding binarised cell-tissue masks.

744 Next, extraneous cell boundaries were removed or added by comparing the initial watershed out
745 put with the grayscale images, and incorrectly assigned void pixels were reassigned to the cell

746 phase, or vice versa, following a similar procedure to that described in Ref. &#91;10&#93;. The accuracy

747 of the watershed segmentation was then evaluated by comparing the mean and standard deviation

748 of the resulting cell statistics with those obtained from the manually corrected segmentations for

749 each species (Fig. S6).


750 **4.7** **The cell area and perimeter equations based on cell arm statistics**


751 Here we derived the relationships for the expected area and perimeter of the cells prior to any

752 truncation resulting from cell-cell overlap or intersection with the boundaries of the simulation

753 domain in the SCellPack framework.

754 For the circular geometry, the cell area ( _A_ _c_ ) and perimeter ( _P_ _c_ ) of the initial, non-truncated

755 cells, are determined solely by their randomly sampled arm length _L_, considered as the radius (see
756 section 4.2), given by _πL_ &#91;2&#93; and 2 _πL_, respectively. Hence, the average values of these parameters

757 over the cells generated by SCellpaCk is given by:


_⟨A_ _c_ _⟩_ = _π⟨L_ &#91;2&#93; _⟩,_ _⟨P_ _c_ _⟩_ = 2 _π⟨L⟩._ (6)


758 In the simulations, the cell radius _L_ is sampled uniformly from the interval &#91; _L_ arm _−_ _σ_ _L_ _, L_ arm + _σ_ _L_ &#93;,

759 where _L_ arm and _σ_ _L_ denote the experimentally measured mean and standard deviation of the arm

760 length, respectively. It is easy to show that the first two moments of this distribution are


_⟨L⟩_ = _L_ arm _,_ (7)



761 and

_⟨L_ &#91;2&#93; _⟩_ = _L_ &#91;2&#93;



_L_ (8)
3 _&#91;.&#93;_




&#91;2&#93; arm &#91;+&#93; _&#91;σ&#93;_ _L_ &#91;2&#93;



762 Substituting the equations 7 and 8 into the equation 6, the mean area and perimeter of the initially

763 generated circular cells by SCellPack are




&#91;2&#93; arm &#91;+&#93; _&#91; πσ&#93;_ _L_ &#91;2&#93;



_⟨A_ _c_ _⟩_ = _πL_ &#91;2&#93;



_L_ (9)
3 _&#91;,&#93;_



28


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


764 and

_⟨P_ _c_ _⟩_ = 2 _πL_ arm _._ (10)


765 For the 1-armed cell geometry, the cell is constructed as the Minkowski sum (buffer) of a line

766 segment of length _L_ by a circular radius _D/_ 2, where _D_ is the corresponding sampled arm diameter.

767 The resulting capsule-like geometry consists of a rectangular section of length _L_ and width _D_, with

768 two semicircular ends of radius _D/_ 2 (see Fig. 1b). Its area and perimeter are therefore given by


_A_ 1 = _LD_ + _&#91;πD&#93;_ &#91;2&#93; (11)

4 _&#91;,&#93;_


769 and

_P_ 1 = 2 _L_ + _πD,_ (12)


770 respectively. Here, the first terms in each expression correspond to the rectangular section, while

771 the second terms account for the two semicircular ends. In our simulations, _D_ is scaled propor
772 tionally to _L_ (see section 4.2) as:



_D_ = _f_ _DL_ _L,_ _f_ _DL_ = _&#91;D&#93;_ &#91;arm&#93;

_L_ arm



_._ (13)



773 Therefore, substituting the equation 13 into equations 11 and 12 the average area and perimeter

774 over randomly sampled cells are given by:



_f_ _DL_ + _&#91;π&#93;_ 4 _&#91;f&#93;_ _DL_ &#91; 2&#93;



)︂



_⟨A_ 1 _⟩_ =



(︂



_⟨L_ &#91;2&#93; _⟩,_ _⟨P_ 1 _⟩_ = (2 + _πf_ _DL_ ) _⟨L⟩._ (14)



775 Finally, substituting equations 13, 7 and 8 into equation 14, the mean area and perimeter of the

776 initially generated 1-armed cells by SCellPack become:



arm

_⟨A_ 1 _⟩_ = _D_ arm _L_ arm + _&#91;πD&#93;_ &#91;2&#93; +

4



_D_ arm
(︃ _L_ arm



2
_σ_ _L_ (15)
)︃ 3 _&#91;,&#93;_



arm

+ _&#91;πD&#93;_ &#91;2&#93;

4 _L_ &#91;2&#93;
arm



777 and

_⟨P_ 1 _⟩_ = 2 _L_ arm + _πD_ arm _._ (16)


778 For cells with multiple arms, the area and perimeter depend not only on the arm length and

779 diameter, but also on the angles between the arms. This is because the buffered arms overlap in the

780 region surrounding their common origin. Specifically, for an n-armed cell, the area and perimeter

781 would be the combined armed statistics minus their overlapped sections as


_A_ _n_ = _nA_ 1 _−_ _A_ _overlap_ _,_ (17)


29


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


782 and

_P_ _n_ = _nP_ 1 _−_ _P_ _overlap_ _,_ (18)


783 where _A_ _overlap_ represents the total area correction associated with pairwise and higher-order over
784 laps, while _P_ _overlap_ represents the corresponding reduction in the summed individual-arm perime
785 ters due to boundaries that become internal to the union. For instance, for the cell morphology
786 2-armed 180 _◦_ one can easily show that _A_ _overlap_ = _&#91;πD&#93;_ 4 &#91;2&#93; and _P_ _overlap_ = _πD_, while by substituting

787 these in the previous relations and performing the same procedure, we will have:



_⟨A_ 2 _⟩_ = 2 _D_ arm _L_ arm + _&#91;πD&#93;_ arm &#91;2&#93; +

4



(︃



2 _&#91;D&#93;_ &#91;arm&#93;
_L_ arm



2
_σ_ _L_ (19)
)︃ 3 _&#91;,&#93;_



arm

+ _&#91;πD&#93;_ &#91;2&#93;

4 _L_ &#91;2&#93;
arm



788 and

_⟨P_ 2 _⟩_ = 4 _L_ arm + _πD_ arm _._ (20)


789 **4.8** **Cell truncation factor**


790 To limit the extent to which the clipping process during the SCellPack simulation can reduce a

791 candidate cell from its initially sampled size, as described above, a cell truncation fraction parame
792 ter _T_ was introduced. The parameter _T_ defines the upper-bound fraction of the originally sampled

793 candidate-cell area that can be truncated due to interactions with previously accepted cells or the

794 simulation-box boundary. Based on this definition, _T_ was varied from 0 to 1. A value of _T_ = 1

795 imposes no upper limit on truncation, provided that a non-zero area remains for the candidate

796 cell, whereas _T_ = 0 does not allow any truncation of the initially sampled cell; otherwise, the

797 trial is rejected. Smaller values of _T_ ensure that the placed cells more closely retain their initially

798 sampled size and prescribed morphology (e.g., circular or 3-armed geometries), but also make it

799 more difficult to reach the target _f_ _IAS_, possibly resulting in a larger deviation from the target and,

800 consequently, a higher simulation error.

### 801 5 Code and data availability


802 &#91;The Python functions supporting our analyses are available at the following URL: github.com/MRahimiMajd/SCel&#93;(https://github.com/MRahimiMajd/SCellPack)

803 All the binary MicroCT slices of the species analyzed in the study are publicly available through

804 the Ref. &#91;1&#93;.

### 805 6 Declaration of interest


806 The authors declare no competing interests.


30


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)

### 807 7 Funding


808 M.R-M., H.M., and Z.N. acknowledge the funding support by the NovoNordisk Foundation, Data

809 Science Initiative, project DIRECTION (Grant NNF 21OC0068884 to Z.N.)

### 810 8 Author contributions


811 M.R.-M. conceptualized the study, performed computations, analyzed data, interpreted results, and

812 wrote the paper. A.M.B. curated data, interpreted results, and edited the paper. H.M. performed

813 manual segmentation of the grayscale images and contributed software and investigation. N.C.W.

814 contributed data curation and interpretation. C.R.B. interpreted results and edited the paper. Z.N.

815 conceptualized the study, interpreted results, acquired funding, supervised M.R.-M. and H.M., and

816 wrote the paper. All authors contributed to finalizing the manuscript.

### 817 References


818 &#91;1&#93; Aleca M Borsuk, Adam B Roddy, Guillaume Th´eroux-Rancourt, and Craig R Brodersen.

819 Structural organization of the spongy mesophyll. _New Phytologist_, 234(3):946–960, 2022.


820 &#91;2&#93; Ichiro Terashima, Yuko T Hanba, Danny Tholen, and Ulo &#91;¨&#93; Niinemets. Leaf functional

821 anatomy in relation to photosynthesis. _Plant physiology_, 155(1):108–116, 2011.


822 &#91;3&#93; John R Evans. Mesophyll conductance: walls, membranes and spatial complexity. _New_

823 _Phytologist_, 229(4):1864–1876, 2021.


824 &#91;4&#93; Jaume Flexas, MIQUEL Ribas-Carbo, Antonio Diaz-Espejo, Jeroni Galm´es, and Hip´olito

825 Medrano. Mesophyll conductance to co2: current knowledge and future prospects. _Plant,_

826 _cell &amp; environment_, 31(5):602–621, 2008.


827 &#91;5&#93; Guillaume Th´eroux-Rancourt, Adam B Roddy, J Mason Earles, Matthew E Gilbert, Ma
828 ciej A Zwieniecki, C Kevin Boyce, Danny Tholen, Andrew J McElrone, Kevin A Simonin,

829 and Craig R Brodersen. Maximum co2 diffusion inside leaves is limited by the scaling

830 of cell size and genome size. _Proceedings_ _of_ _the_ _Royal_ _Society_ _B:_ _Biological_ _Sciences_,

831 288(1945):20203145, 2021.


832 &#91;6&#93; J¨urgen Knauer, Matthias Cuntz, John R Evans, Ulo &#91;¨&#93; Niinemets, Tiina Tosens, Linda-Liisa

833 Veromann-J¨urgenson, Christiane Werner, and S¨onke Zaehle. Contrasting anatomical and bio
834 chemical controls on mesophyll conductance across plant functional types. _New Phytologist_,

835 236(2):357–368, 2022.


31


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


836 &#91;7&#93; Christoph Lehmeier, Radoslaw Pajor, Marjorie R Lundgren, Andrew Mathers, Jen Sloan,

837 Marion Bauch, Alice Mitchell, Chandra Bellasio, Adam Green, Daniel Bouyer, et al. Cell

838 density and airspace patterning in the leaf can be manipulated to increase leaf photosynthetic

839 capacity. _The Plant Journal_, 92(6):981–994, 2017.


840 &#91;8&#93; Christopher D Whitewoods. Riddled with holes: Understanding air space formation in plant

841 leaves. _PLoS biology_, 19(12):e3001475, 2021.


842 &#91;9&#93; Kimitsune Ishizaki. Development of schizogenous intercellular spaces in plants. _Frontiers in_

843 _Plant Science_, 6:497, 2015.


844 &#91;10&#93; John D Treado, Adam B Roddy, Guillaume Th´eroux-Rancourt, Liyong Zhang, Chris Am
845 brose, Craig R Brodersen, Mark D Shattuck, and Corey S O’Hern. Localized growth and

846 remodelling drives spongy mesophyll morphogenesis. _Journal_ _of_ _the_ _Royal_ _Society_ _Inter-_

847 _face_, 19(197), 2022.


848 &#91;11&#93; Liyong Zhang and Chris Ambrose. Beauty is more than epidermis deep: How cell divi
849 sion and expansion sculpt the leaf spongy mesophyll. _Current_ _opinion_ _in_ _plant_ _biology_,

850 79:102542, 2024.


851 &#91;12&#93; J Mason Earles, Guillaume Th´eroux-Rancourt, Adam B Roddy, Matthew E Gilbert, Andrew J

852 McElrone, and Craig R Brodersen. Beyond porosity: 3d leaf intercellular airspace traits that

853 impact mesophyll conductance. _Plant Physiology_, 178(1):148–162, 2018.


854 &#91;13&#93; Guillaume Th´eroux-Rancourt, Jos´e Carlos Herrera, Klara Voggeneder, Federica De Berar
855 dinis, Natascha Luijken, Laura Nocker, Tadeja Savi, Susanne Scheffknecht, Moritz Sch
856 neck, and Danny Tholen. Analyzing anatomy over three dimensions unpacks the differ
857 ences in mesophyll diffusive area between sun and shade vitis vinifera leaves. _AoB_ _Plants_,

858 15(2):plad001, 2023.


859 &#91;14&#93; Jeroen DM Schreel, Guillaume Th´eroux-Rancourt, Pamela K Diggle, Craig Brodersen, and

860 Adam B Roddy. Into the spongy-verse: structural differences between leaf and flower meso
861 phyll. _Integrative and comparative biology_, 65(1):81–93, 2025.


862 &#91;15&#93; Joohyun Kang, Kyungyoon Kim, Jooyeon Woo, Thanh Ha Thi Do, and Yuree Lee. Recent in
863 sights into air space formation in plant shoots. _Current Opinion in Plant Biology_, 85:102725,

864 2025.


865 &#91;16&#93; Liyong Zhang, Delanie McEvoy, Yen Le, and Chris Ambrose. Live imaging of microtubule

866 organization, cell expansion, and intercellular space formation in arabidopsis leaf spongy

867 mesophyll cells. _The Plant Cell_, 33(3):623–641, 2021.


32


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


868 &#91;17&#93; Adrienne HK Roeder, Paul T Tarr, Cory Tobin, Xiaolan Zhang, Vijay Chickarmane, Alexan
869 dre Cunha, and Elliot M Meyerowitz. Computational morphodynamics of plants: integrating

870 development over space and time. _Nature_ _reviews_ _Molecular_ _cell_ _biology_, 12(4):265–273,

871 2011.


872 &#91;18&#93; Fr´ed´eric Boudon, J´erˆome Chopard, Olivier Ali, Benjamin Gilles, Olivier Hamant, Arezki

873 Boudaoud, Jan Traas, and Christophe Godin. A computational framework for 3d mechani
874 cal modeling of plant morphogenesis with cellular resolution. _PLoS computational biology_,

875 11(1):e1003950, 2015.


876 &#91;19&#93; Bandan Chakrabortty, Ikram Blilou, Ben Scheres, and Bela M Mulder. A computational

877 framework for cortical microtubule dynamics in realistically shaped plant cells. _PLoS com-_

878 _putational biology_, 14(2):e1005959, 2018.


879 &#91;20&#93; Varsha S Pathare, Nuria Koteyeva, and Asaph B Cousins. Increased adaxial stomatal den
880 sity is associated with greater mesophyll surface area exposed to intercellular air spaces and

881 mesophyll conductance in diverse c4 grasses. _New Phytologist_, 225(1):169–182, 2020.


882 &#91;21&#93; Milad Rahimi-Majd, Alistair Leverett, Arne Neumann, Johannes Kromdijk, and Zoran

883 Nikoloski. Nonlinear models based on leaf architecture traits explain the variability of mes
884 ophyll conductance across plant species. _Plant,_ _Cell_ _&amp;_ _Environment_, 47(12):5158–5171,

885 2024.


886 &#91;22&#93; Yusuke Mizokami, Riichi Oguchi, Daisuke Sugiura, Wataru Yamori, Ko Noguchi, and Ichiro

887 Terashima. Cost–benefit analysis of mesophyll conductance: diversities of anatomical, bio
888 chemical and environmental determinants. _Annals of Botany_, 130(3):265–283, 2022.


889 &#91;23&#93; Sung Nok Chiu, Dietrich Stoyan, Wilfrid S Kendall, and Joseph Mecke. _Stochastic geometry_

890 _and its applications_ . John Wiley &amp; Sons, 2013.


891 &#91;24&#93; Stephen B Gray. Local properties of binary images in two dimensions. _IEEE Transactions_

892 _on computers_, 100(5):551–561, 1971.


893 &#91;25&#93; Herbert Edelsbrunner and John Harer. _Computational topology:_ _an introduction_ . American

894 Mathematical Soc., 2010.


895 &#91;26&#93; Bruno Galerne. Computation of the perimeter of measurable sets via their covariogram.

896 applications to random sets. _Image Analysis &amp; Stereology_, 30(1):39–51, 2011.


897 &#91;27&#93; Alice L Baillie and Andrew J Fleming. The developmental relationship between stomata and

898 mesophyll airspace. _New Phytologist_, 225(3):1120–1126, 2020.


33


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


899 &#91;28&#93; James M Fischer, Guillaume Th´eroux-Rancourt, Craig R Brodersen, and Andrew DB Leakey.

900 High-resolution microct reveals relationships between stomata and interior leaf anatomy in

901 sorghum. _Plant Physiology_, page kiag414, 2026.


902 &#91;29&#93; D Jeulin. Dead leaves models: from space tesselation to random functions. In _Proc. of the_

903 _Symposium on the Advances in the Theory and Applications of Random Sets_, pages 137–156.

904 World Scientific, 1996.


905 &#91;30&#93; Yann Gousseau and Franc¸ois Roueff. The dead leaves model: general results and limits at

906 small scales. _arXiv preprint math/0312035_, 2003.


907 &#91;31&#93; Mathew D Penrose. Leaves on the line and in the plane. 2020.


908 &#91;32&#93; Stefan Van der Walt, Johannes L Sch¨onberger, Juan Nunez-Iglesias, Franc¸ois Boulogne,

909 Joshua D Warner, Neil Yager, Emmanuelle Gouillart, and Tony Yu. scikit-image: image

910 processing in python. _PeerJ_, 2:e453, 2014.


911 &#91;33&#93; Pauli Virtanen, Ralf Gommers, Travis E Oliphant, Matt Haberland, Tyler Reddy, David Cour
912 napeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, et al. Scipy

913 1.0: fundamental algorithms for scientific computing in python. _Nature methods_, 17(3):261–

914 272, 2020.


915 &#91;34&#93; Johannes Schindelin, Ignacio Arganda-Carreras, Erwin Frise, Verena Kaynig, Mark Longair,

916 Tobias Pietzsch, Stephan Preibisch, Curtis Rueden, Stephan Saalfeld, Benjamin Schmid, et al.

917 Fiji: an open-source platform for biological-image analysis. _Nature methods_, 9(7):676–682,

918 2012.


34


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


919
# **Supporting Information**


35


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)



**Figure** **S1:** **SCellPack** **statistics** **for** **the** **reconstructed** **leaf** **porosity.** **a.** Comparison of the
measured and simulated mean fraction of intercellular airspace to total area ( _f_ _IAS_ ) in paradermal
leaf slices across 40 plant species. Each marker corresponds to a different cell geometry (circular,
1-armed, 2-armed, 2-armed 180 _◦_, 3-armed, and 4-armed). Error bars indicate standard errors of
slice-averaged values. **b.** The bar plot of the absolute error between the measured and simulated
values of _f_ _IAS_ ( _|f_ _IAS_ _&#91;sim&#93;_ _&#91;−&#93;_ _&#91;f&#93;_ _IAS_ _&#91; data&#93;_ _&#91;|&#93;_ &#91;) for three selected cell geometries (circular, 2-armed, and 3-armed)&#93;




_&#91;sim&#93;_

_IAS_ _&#91;−&#93;_ _&#91;f&#93;_ _IAS_ _&#91; data&#93;_



values of _f_ _IAS_ ( _|f_ _IAS_ _&#91;sim&#93;_ _&#91;−&#93;_ _&#91;f&#93;_ _IAS_ _&#91; data&#93;_ _&#91;|&#93;_ &#91;) for three selected cell geometries (circular, 2-armed, and 3-armed)&#93;

across all species. In both panels, species are ordered by increasing arm length, with blue, yellow,
and brown shaded regions indicating non-honeycomb, honeycomb, and neither spongy patterns,
respectively. Markers and colors are consistent across all panels.



36


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


**Figure** **S2:** **SCellPack** **statistics** **for** **the** **reconstructed** **leaf** **topology** **traits.** **a.** Comparison
between measured and simulated mean number of connected components of leaf intracellular
airspace (IAS) phase ( _β_ 0 _IAS_ ) across 40 plant species. Each marker corresponds to a different cell
geometry (circular, 1-armed, 2-armed, 2-armed 180 _◦_, 3-armed, and 4-armed). Error bars indicate
the standard errors of the slice-level values. Species are ordered by increasing mean arm length.
Blue, yellow, and brown shaded regions indicate non-honeycomb, honeycomb, and neither spongy
patterns, respectively. **b.** Absolute logarithmic relative error of _β_ 0 _AIS_ ( _|LRE_ ( _β_ 0 _IAS_ ) _|_ )) across the
40 species for circular, 2-armed, and 3-armed geometries. Bars indicate the reconstruction errors
for the corresponding data points of each geometry in panel **a** . The horizontal dotted line marks
a two-fold deviation, _|LRE|_ = ln(2). **c.** and **d.** Are the same as **a** and **b** for the number of connected components of leaf cell phase ( _β_ 0 _Cell_ ), respectively. Markers and colors corresponding to
the experimental data and simulated cell geometries are consistent across panels.


37


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


**Figure** **S3:** **Distributions** **of** **SCellPack** **reconstruction** **errors** **for** **leaf** **topology.** Cumulative
distribution functions of the absolute logarithmic error ( _|LRE|_ ) between simulated and measured
values for the traits: **a.** number of intercellular airspace (IAS) connected components ( _β_ 0 _IAS_ )
and **b.** number of cell connected components ( _β_ 0 _Cell_ ). Colors indicate different cell geometries
(circular, 1-armed, 2-armed, 3-armed, and 4-armed). The curves show the probability that the
reconstruction error is less than or equal to a given threshold _x_ . The vertical dotted lines mark a
two-fold deviation, _|LRE|_ = ln(2).


38


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


**Figure** **S4:** **SCellPack** **statistics** **for** **the** **reconstructed** **leaf** **cell-level** **traits.** **a.** Comparison
between measured and simulated mean cell number ( _N_ _Cell_ ) across 40 plant species. Each marker
corresponds to a different cell geometry (circular, 1-armed, 2-armed, 2-armed 180 _◦_, 3-armed, and
4-armed). Error bars indicate the standard errors of the slice-level values. Species are ordered by
increasing mean arm length. Blue, yellow, and brown shaded regions indicate non-honeycomb,
honeycomb, and neither spongy patterns, respectively. **b.** Absolute logarithmic relative error of
_N_ _Cell_ ( _|LRE_ ( _N_ _Cell_ ) _|_ )) across the 40 species for circular, 2-armed, and 3-armed geometries. Bars
indicate the reconstruction errors for the corresponding data points of each geometry in panel **a** .
The horizontal dotted line marks a two-fold deviation, _|LRE|_ = ln(2). **c.** and **e.**, are the same as
**a** for the average cell perimeter per leaf area ( _P_ _Cell_ ), and the average cell area per leaf area ( _A_ _Cell_ ),
respectively. **d.** and **f.** are the same as **b** for _P_ _Cell_ ( _|LRE_ ( _P_ _Cell_ ) _|_ ) and _A_ _Cell_ ( _|LRE_ ( _A_ _Cell_ ) _|_ ),
respectively. The measured values in all panels are based on the watershed cell segmentation of
leaf paradermal slices. Markers and colors are consistent across all panels.


39


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


**Figure S5: Distributions of SCellPack reconstruction errors for leaf cell-level traits. Cumula-**
**tive distribution functions of the absolute logarithmic error (** _|LRE|_ **) between simulated and**
**measured values for the traits:** **a.** **the number of cells (** _N_ _Cell_ **),** **b.** **the average cell perimeter**
**per** **leaf** **area** **(** _P_ _Cell_ **),** **and** **c.** **the** **average** **cell** **area** **per** **leaf** **area** **(** _A_ _Cell_ **).** **Colors** **indicate** **dif-**
**ferent cell geometries (circular, 1-armed, 2-armed, 3-armed, and 4-armed).** **The curves show**
**the probability that the reconstruction error is less than or equal to a given threshold** _x_ **.** **The**
**vertical dotted lines mark a two-fold deviation,** _|LRE|_ = ln(2) **.**


40


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)

## **a.** Species 1: Koelreuteria paniculata
### **Grayscale image Watershed segmentation Manual segmentation**

## b. Species 36: Spinacia oleracea

## **c.**
## Species 40: Platycerium andinum


**Figure S6: Comparison of watershed and manual cell segmentation of leaf paradermal slices.**
Segmentation workflow for spongy mesophyll cells from original grayscale microCT images for
three representative species. For each species, the left, middle, and right panels show the original
grayscale image, the watershed segmentation, and the manually corrected watershed segmentation,
respectively.


41


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


**Figure** **S7:** **Comparison** **of** **watershed** **and** **manual** **segmentation** **statistics.** Comparison of
pixel-based statistics derived from watershed and manual cell segmentation of leaf paradermal
slices across 40 plant species, ordered by increasing mean arm length, for: **a.** average cell perimeter
( _P_ _Cell_ ), **b.** average cell area ( _A_ _Cell_ ), and **c.** cell number ( _N_ _Cell_ ). Error bars in the panels **a** and **b**
represent the standard deviations of within-slice mean values. Blue, yellow, and brown shaded
regions indicate non-honeycomb, honeycomb, and neither spongy patterns, respectively.


42


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


**Figure** **S8:** **Performance** **of** **the** **Boolean** **model** **in** **predicting** **mesophyll** **cell** **perimeter** **ex-**
**posed to intercellular airspace (** _S_ _m_ **).** **a.** Comparison of theoretical (Boolean model), measured,
and simulated mean _S_ _m_ across 40 plant species. Each marker corresponds to a different cell geometry (circular, 1-armed, and2-armed 180 _◦_ ). Error bars indicate the standard errors of the slice-level
values. Species are ordered by increasing mean arm length. Blue, yellow, and brown shaded regions indicate non-honeycomb, honeycomb, and neither spongy patterns, respectively. **b.** Absolute
logarithmic relative error of _S_ _m_, _|LRE_ ( _S_ _m_ ) _|_, between theoretical and simulated values across the
40 species and three cell geometries. Bars indicate the reconstruction errors for the corresponding
data points of each geometry in panel **a** . The horizontal dotted line marks a two-fold deviation,
_|LRE|_ = ln(2). **c.** Same as **b**, but showing _|LRE_ ( _S_ _m_ ) _|_ between theoretical and measured values.
Markers and colors are consistent across all panels.


43


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


**Figure S9:** **Statistics of the fitted relationship between leaf porosity and shared wall fraction**
**with different truncation fractions.** **a.** Coefficient of determination ( _R_ &#91;2&#93; ), **b.** absolute slope ( _|α|_ ),
and **c.** intercept ( _m_ ) of the linear fits between shared wall fraction ( _f_ _loss_ ) as a function of the
fraction of intercellular airspace ( _f_ _IAS_ ) for different cell geometries across the range of truncation
fractions ( _T_ ) tested in SCellPack. Colored bars indicate the fitted values for each _T_, while the
dashed black line indicates the corresponding value from the measured data.


44


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.18.767598; this version posted November 19, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.18.767598)

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
&#91;available under a CC-BY-NC-ND 4.0 International license.&#93;(http://creativecommons.org/licenses/by-nc-nd/4.0/)


**Figure S10: Distributions of SCellPack reconstruction errors for leaf traits with altered trun-**
**cation fraction (** _T_ **).** Cumulative distribution functions of the absolute logarithmic error ( _|LRE|_ )
between simulated and measured values for the traits configured with _T_ = 0 _._ 4: **a.** cell perimeter
exposed to intercellular airspace ( _S_ _m_ ), **b.** number of intercellular airspace (IAS) connected components ( _β_ 0 _IAS_ ), **c.** number of cell connected components ( _β_ 0 _Cell_ ), **d.** number of cells ( _N_ _Cell_ ), **e.**
average cell perimeter per leaf area ( _P_ _Cell_ ), and **f.** average cell area per leaf area ( _A_ _Cell_ ). Colors
indicate different cell geometries (circular, 1-armed, 2-armed, 3-armed, and 4-armed). The curves
show the probability that the reconstruction error is less than or equal to a given threshold _x_ . The
vertical dotted lines mark a two-fold deviation, _|LRE|_ = ln(2).


45


