DrugCoding tests whether AI can keep complex drug-discovery workflows on track

A computational drug-discovery project can resemble a relay race: one program generates candidate molecules, another predicts their properties, and another evaluates how they might interact with a biological target. For the hand-offs to work, the software must preserve important details, such as molecular strings and editing instructions, while tracking which steps depend on earlier results. DrugCoding is designed to manage that coordination.

The system takes a research request and supplied information, breaks the request into linked tasks, and routes tasks to specialized scientific models and tools. It uses separate agent components for jobs such as resolving the context, planning the workflow, recommending models and carrying out tool calls. Programs check whether returned results have the expected structure and inputs before accepting them for later steps. The system also records accepted results and their provenance, so downstream tasks can use them.

The researchers describe this as a scientific programming agent: rather than merely answering a question in conversation, it assembles and executes computational work. Its available capabilities include molecular generation, drug–target affinity prediction, docking, virtual screening, and predictions of drug response or ADMET properties—absorption, distribution, metabolism, excretion and toxicity. These are computational predictions, not experimental tests of a medicine.

What the test measured

The main evaluation used the multi-turn portion of the Tool-Calling Dataset for Drug Discovery, or TCDD, a benchmark introduced in earlier work. It contains 1,250 sessions and 2,500 action-level inputs across eight tool interfaces. A session might require an agent to make several tool calls while retaining or revising information from earlier turns. The researchers also tested a smaller, selected subset of 128 sessions that they ranked as especially challenging.

Scoring was strict. A session passed only if the submitted tool operations matched the reference in identity, order and all required argument values. An agent could therefore fail by choosing the wrong tool, omitting a step, or changing a supplied scientific input—even if its written explanation claimed success. The evaluation used seven language-model configurations for the full benchmark and compared DrugCoding with configured versions of Codex, Claude Code and Kimi. On the smaller subset, it was also compared with Claude Science and Rosalind Workbench.

Reported results—and what they mean

DrugCoding’s average session-level Pass@1—the share of sessions completed successfully in a single final submission—was 97.35% on the full benchmark. The report gives Codex, the strongest of the three coding-agent baselines there, an average of 88.11%, a difference of 9.24 percentage points. On the challenging subset, DrugCoding averaged 97.92%, compared with 96.61% for Claude Science and 96.09% for Rosalind Workbench.

An analysis of failures pointed to practical sources of error. Across the full benchmark, DrugCoding had fewer recorded argument mismatches and runtime or parsing failures than the compared coding agents. Examples in the report show other systems lengthening a supplied editing mask or adding characters to a molecular string. Such alterations can make a tool call fail the benchmark, even when the intended operation is understood.

The study also compared multi-agent and single-agent versions of DrugCoding. The multi-agent version averaged 97.35%, against 96.79% for the single-agent version, a 0.56-point difference. It did not perform better in every model configuration. That result cautions against assuming that adding agents automatically improves performance.

A useful but limited measure

The benchmark tests whether an agent can produce the specified computational instructions and preserve their inputs. It does not establish that the selected scientific model is accurate, that a predicted molecule will work in a laboratory, or that any candidate is safe or effective in people. The report itself says further evaluation is needed to determine how reliably correct submissions carry through to executed workflows and validated scientific outcomes.

The comparisons also concern particular system configurations, not every version or default setting of the named products. The authors note that prompting, provider routes and output handling differed, so the results do not isolate which part of DrugCoding explains its advantage. The multi-agent ablation likewise suggests that gains depend on the underlying language model.

Finally, this is an XDeNovo Labs technical report, version 1.0, posted as an arXiv preprint. A preprint is publicly available research that may not have undergone peer review; its results should be treated as provisional. The evidence supports a focused conclusion: on this tool-calling benchmark, DrugCoding was often better at turning multi-turn requests into exactly structured submissions. Whether that reliability translates into better drug-discovery decisions remains an open question.