Fármaco AI agent coordinates drug-discovery tools through a shared scientific database

A new preprint by Marcel Müller and colleagues presents El Agente Fármaco as a coordinator for computational drug discovery. Rather than replacing molecular simulation or property-prediction software, the system uses a large language model to select those tools, interpret their outputs and decide what to do next. Its central architectural choice is to keep detailed scientific records in a database instead of repeatedly passing them through the model’s conversation.

The reported results concern computational workflows, not demonstrated new medicines. The authors evaluate task execution and candidate prioritization, while documenting cases where predictions disagree, tool calls fail or apparently useful scores are misinterpreted.

A database instead of an expanding conversation

Fármaco Atlas stores proteins, molecules, three-dimensional structures, calculation results and workflow records. These records connect outputs to the inputs and computational actions that produced them. Large files remain outside the database but are registered within it.

The agent works with reusable collections called subsets. To assess hundreds of molecules, it can request one workflow referencing a collection rather than make a separate language-model decision for every compound. The execution system resolves the collection, dispatches calculations with bounded concurrency and saves the results. The model receives a compact summary and can retrieve selected details through targeted database queries.

Tool settings are also disclosed selectively: detailed parameters reach a separate configuration step only after a tool has been chosen. This limits information entering the model’s working context, although the additional configuration and database interactions can themselves require more model calls.

Lower token use, with a qualified quality advantage

The researchers compared three configurations across 22 tasks: database-backed Fármaco, the same domain agent using plain scientific tool calls, and Codex. All used the same execution model. Three repetitions per task and configuration produced 198 runs, assessed in 594 separate language-model judge sessions.

Fármaco’s mean overall judge score was 8.67 out of 10, compared with 8.30 for plain tool calls and 8.41 for Codex. However, the confidence intervals overlapped. These are model-judged assessments of delivered computational work, not measurements of drug-discovery success.

Median cumulative token use was 176,523 for Fármaco, 234,014 for plain tool calls and 566,426 for Codex. Fármaco also had the smallest median peak context. But fewer tokens did not make it the cheapest or fastest: plain tool calls had the lowest median recorded model cost and latency. The benchmark covered manageable tasks rather than long-running, large-library discovery campaigns.

A separate 256-ligand experiment illustrated the potential benefit of bulk execution. Database-backed Fármaco completed prediction, ranking and interaction analysis in one user turn, whereas the plain-tool configuration required six. That comparison used different prediction-sampling settings, so it addresses execution overhead rather than prediction quality.

Chemically distinct candidates, not proven stronger binders

For the protein TBXT, Fármaco generated 705 ligand records and ultimately selected 34 compounds. The researchers compared these with 39 compounds from an independently completed expert-led campaign, applying the same external docking protocol to both sets.

The docking-score distributions were not significantly different. Fármaco’s candidates had higher calculated drug-likeness and fewer rule-of-five violations, but the agent had explicitly been instructed to enforce developability constraints that differed from the earlier campaign. The result therefore demonstrates constraint-aware selection, not an unqualified advantage over experts.

The agent-selected compounds were chemically distinct from the purchased set. Neither set’s computational scores established binding or potency. Some expert-selected structures remain withheld during ongoing validation, limiting independent reproduction of structure-based comparisons.

Why prediction checks still matter

Other case studies exposed limits that a coordinating agent cannot remove simply by connecting more tools. In a PGK2 workflow using PGK1 as a counter-target, the agent narrowed 369 generated molecules to five meeting reported property, interaction and model-score criteria. However, an audit found that the archived pose comparisons did not independently establish agreement between docking and co-folding predictions. Experimental binding and selectivity remained future work.

For KpThiL, the strongest predicted affinity score belonged to a pose roughly 27 angstroms from the intended substrate site. In a TYK2 analogue campaign, a candidate favored by Boltz-2 showed approximately unchanged binding relative to its seed in a subsequent relative binding free-energy calculation. Convergence was not established, and experimental improvement was not demonstrated.

The authors envision connecting such workflows to synthesis and biological assays. That experimental feedback loop remains a proposed direction. The preprint’s contribution is a system for coordinating and inspecting computational work—not evidence that automated predictions have resolved the uncertainties of drug discovery.