El Agente Fármaco: A Scalable AI Scientist for Drug Discovery

Marcel Müller; Mohammad Ghazi Vakili; Yoshihiro Ogura; Ella Ashoori; Shimeng Wang; Alessandro S. Nascimento; Eddie Calleja; Yutong Zhao

cs.AINov 2, 2026

El Agente Fármaco: A Scalable AI Scientist for Drug Discovery

Authors: Marcel Müller, Mohammad Ghazi Vakili, Yoshihiro Ogura, Ella Ashoori, Shimeng Wang, Alessandro S. Nascimento, Eddie Calleja, Yutong Zhao,

Organizations: Department of Chemistry, University of Toronto, Lash Miller Chemical Laboratories, 80 St. George Street, ON M5S 3H6, Toronto, Canada · Vector Institute for Artificial Intelligence, 661 University Ave. Suite 710, ON M5G 1M1, Toronto, Canada · Department of Computer Science, University of Toronto, Sandford Fleming Building, 10 King’s College Road, ON M5S 3G4, Toronto, Canada · São Carlos Institute of Physics, University of São Paulo, São Carlos, Brazil · NVIDIA, 2788 San Tomas Expressway, Santa Clara, CA 95051, USA · Acceleration Consortium, 700 University Ave., M7A 2S4, Toronto, Canada · Department of Materials Science & Engineering, University of Toronto, 184 College St., M5S 3E4, Toronto, Canada · Department of Chemical Engineering & Applied Chemistry, University of Toronto, 200 College St. ON M5S 3E5, Toronto, Canada · Institute of Medical Science, 1 King’s College Circle, Medical Sciences Building, Room 2374, Toronto, ON M5S 1A8, Canada · Senior Fellow, Canadian Institute for Advanced Research (CIFAR), 661 University Ave., M5G 1M1, Toronto, Canada · NVIDIA, 431 King St W #6th, M5V 1K4, Toronto, Canada

Abstract

Computational drug discovery offers a growing range of tools for proposing and evaluating molecules, but combining them and deciding what to do next still requires substantial expert effort. Here, we introduce El Agente Fármaco, an AI scientist that uses modern large language model (LLM) technology to turn discovery goals expressed in everyday language into complex, scalable computational screening workflows. The agentic harness selects scientific tools, reviews their results, and adjusts subsequent steps to the target and the user's objectives. Our approach centers on a shared scientific memory: Fármaco Atlas, which stores molecules, structures, and results together with their provenance -- a record of how they were obtained. The agent uses targeted database queries to retrieve the evidence needed for each decision. A single request can launch calculations for an entire collection of molecules, without requiring a separate LLM decision for every compound. This design limits the information the model must process as screening collections grow, while keeping the underlying results available for inspection and further analysis. The integrated tools support molecule generation, prediction of binding and molecular properties, and searches for compounds available for purchase. In a retrospective comparison with an expert-led campaign, the agent proposed chemically distinct candidates with competitive docking scores and higher calculated drug-likeness under explicitly imposed selection criteria. In a 22-task benchmark against plain tool calls and Codex, the agent achieved the highest observed mean model-judged score, with lower median cumulative token use and peak context. Fármaco provides a foundation for flexible, scalable drug discovery workflows that connect scientific reasoning with an enduring, traceable record of computational work.

Figures & tables

Figure 1: Capability overview of Fármaco. Fármaco ( centre ) coordinates workflows and delegates pocket investigation and dataset ingestion to specialist agents. Subset-based workflows select inputs through sql queries, dispatch tool calls, and return run summaries. Its capabilities include format conversion and retrieval utilities ( above ), six categories of scientific tools ( flanking ), and persistent storage in Fármaco Atlas ( below ), shown with selected fields from its core records. The left and right columns group tools by their roles in proposal and evaluation, respectively. Marker colour indicates the deployment modes shown: hosted nim , containerized or remote services, and local Python or cli tools.

Figure 2: Database-driven tool execution. The agent requests a tool using high-level task and subset descriptions. The workflow translates the subset description into a validated sql query, creates a subset_id , and configures the selected tool. The sql database workflow resolves the subset entity into a list of targets, ligands, or target-ligand pairs ( orange ), sets up job specifications per entity, fans out parallelized workflows ( green ), and summarizes information into a run entity ( pink ). The agent only receives a summary of the run; the resulting metrics, structures, or artifacts are stored in the database ( top-right corner ) and are only accessed upon a targeted sql request.

Figure 3: Benchmark quality versus model cost and peak context. Small markers show task-level results. Large markers summarize each configuration. Lines indicate 95% confidence intervals. Metric definitions and statistical methods are provided in appendix D .

	Fármaco	Expert-led campaign
Generated $\rightarrow$ retained for comparison	$705\rightarrow 34$	$\sim$ 600+ $\rightarrow 39$ of 40 purchased
Best docking score (kcal mol -1 )	$-8.5$	$-8.4$
Median docking score (kcal mol -1 )	$-7.3$	$-7.5$
Median Boltz-2 converted affinity score	5.46	5.43
Mean qed	0.69	0.52
Mean rule-of-five violations	0.06	0.54

Table 1: Retrospective computational comparison on TBXT pocket A ′ . Docking scores were calculated using one common protocol for both sets; lower values indicate more favorable predicted binding. Boltz-2 converted affinity scores are

6-\texttt{affinity\_pred\_value}

. Descriptor values were calculated consistently for the 34 compounds selected by Fármaco and the 39 compounds with retrievable structures among the 40 purchased in the expert-led campaign. Property measures include qed and calculated

\log P

, the base-10 logarithm of the predicted octanol–water partition coefficient. Filtering covers admet and structural alerts. The developability comparison is descriptive because the two campaigns used different selection constraints.

Figure 4: Retrospective computational comparison of Fármaco and an independently completed expert-led TBXT campaign. Both campaigns targeted pocket A ′ ; panels (a–c) compare 34 generated compounds selected by the agent with 39 compounds purchased in the expert-led campaign. (a) Common-protocol docking scores against pdb 7ZKF, including five published reference ligands. (b) Boltz-2 converted affinity scores ( $6-\texttt{affinity\_pred\_value}$ ; see Sec. C.1.1 ). (c) Cohort means of qed (higher is better), rule-of-five violations, calculated $\log P$ and molecular weight divided by 100. Calculated $\log P$ is the base-10 logarithm of the predicted octanol–water partition coefficient. Docking and Boltz-2 scores are predictions; experimental affinities are available only for the five reference ligands ( Sec. C.1.4 ).

Figure 5: Adaptive multistage selection of PGK2 ligand candidates using PGK1 as a counter-target. (a) Candidate counts retained at successive stages of the workflow. Across two adaptive search cycles, 369 compounds were registered in Fármaco Atlas, of which 76 passed the property and structural-alert filters, 58 met the prespecified PGK2 TYR242 interaction criterion, and 5 met the Boltz-2 target–counter-target score criterion. (b) PGK2 and PGK1 Boltz-2 affinity scores for the 58 ligands passing Stage 1. The grey dashed diagonal denotes equal PGK2 and PGK1 scores, whereas the orange dashed line denotes the prespecified selection threshold, PGK2 = PGK1 + 0.477. TYR242 denotes tyrosine residue 242. Green markers indicate the five selected candidates, labelled individually; grey markers indicate the remaining compounds.

Ligand	TYR242 interaction	Boltz-2 affinity score (PGK2)	Boltz-2 affinity score (PGK1)	$\Delta$ score
26	Hydrophobic	7.943	7.417	0.526
78	Hydrogen bond	7.140	6.639	0.501
160	Hydrogen bond	8.138	7.454	0.684
351	Hydrophobic	7.914	7.182	0.731
369	Hydrogen bond	7.779	6.703	1.076

Table 2: Computational results for the five PGK2 ligand candidates selected using PGK1 as a counter-target. Scores are the archived Boltz-2 affinity_pic50 outputs ( Sec. C.2 ). Higher scores indicate stronger predicted binding, and

\Delta

score is the PGK2 score minus the PGK1 score. TYR242 denotes tyrosine residue 242; its interactions were identified by plip analysis of the PGK2 DiffDock rank-1 poses.

Figure 6: Computational prioritization of newly generated AmpC candidates. (a) Screening profiles of the 212 candidates from four generation and analogue-search methods. Cells give counts meeting each criterion, with colour indicating the percentage within each set. Each column applies its criterion separately. (b) Boltz-2 converted affinity scores and predicted binding probabilities for the ten shortlisted candidates; labels identify compounds. Higher values are better on both axes.

Candidate	Transformation from EJM42	Calculated $\Delta\Delta G$	Experimental $\Delta\Delta G$
EJM31	ethyl $\to$ methyl	$-1.06\pm 0.87$	$+0.24\pm 0.25$
EJM43	ethyl $\to$ isopropyl	$+2.20\pm 1.22$	$+1.53\pm 0.25$
EJM50	ethyl $\to$ hydroxymethyl	$+0.50\pm 0.72$	$+0.81\pm 0.25$
EJM55	propanamide $\to$ methyl carbamate	failed	$+0.58\pm 0.25$
EJM54	propanamide $\to$ ethyl urea	$-0.05\pm 0.81$	$-0.75\pm 0.25$
EJM45	ethyl $\to$ cyclopropylmethyl	$+2.86\pm 0.65$	$+0.23\pm 0.25$

Table 3: First six rbfe edges of the agent-designed TYK2 network. Ligand labels abbreviate the benchmark identifiers ( Sec. C.5.1 ).

\Delta\Delta G

values (kcal mol -1 ) are relative to EJM42; negative values favour the candidate. Calculated values are Azulene Studio estimates with 5 ps equilibration and 10 ps production per replica and three repeats; the uncertainty is the one Azulene Studio reports. Experimental values are

RT\ln(K_{\mathrm{i},B}/K_{\mathrm{i},A})

at 298.15 K from published

K_{\mathrm{i}}

[ 48 ] , which were not shown to the agent. A and B denote the reference and candidate ligands, respectively.

Ligand	Benzoyl-ring substitution	Boltz-2 $\mathrm{pIC}_{50}$	Boltz-2 binding probability	rbfe $\Delta\Delta G$
Seed	2,6-dichloro	6.500	0.805	—
2	2-chloro-4-fluoro	6.289	0.730	$+1.710\pm 0.114$
5	2,6-difluoro	6.169	0.742	$+0.807\pm 0.205$
15	2-chloro-6-fluoro	6.785	0.828	$+0.018\pm 0.099$

Table 4: TYK2 seed-to-candidate follow-up. Ligand identifiers refer to the archived database. Azulene Studio reports

\Delta\Delta G=\Delta G(B)-\Delta G(\mathrm{seed})

in kcal mol -1 ; positive values favour the seed. All three edges used the same water-free receptor, 0.5 ns equilibration, 5 ns production per replica, and three repeats. Uncertainties are reported by Azulene Studio. Predicted

\mathrm{pIC}_{50}

values were normalized from the archived Boltz-2 outputs as described in Sec. C.5.2 .

Appendix figures & tables21 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Database workflow. Subject jobs fan out through input resolution; parameter-only tools dispatch one job directly to execution. Outcomes join before materialization and summary generation. Diamonds mark decisions; the double-bordered box links to Fig. 8 . Colours follow Fig. 2 .

Figure 8: Subset-selection graph. A generated read-only sql query produces identifiers for validation and subset creation. Errors trigger query regeneration with feedback; three failed attempts lead to an error with attempt history.

Measure	Value
Agent-selected compounds	34
Purchased compounds	39
Mann–Whitney $U$	561.5
Two-sided $p$	0.26

Appendix

Table 5: Statistical comparison of TBXT docking scores. Agent-selected and purchased compounds were docked against pdb 7ZKF using the same protocol. The docking-score distributions were not significantly different, despite small differences in their best and median scores.

Compound	spr $K_{\mathrm{d}}$ ( $\mu$ M)	Vina 7ZKF (kcal mol -1 )	Vina 7ZK2 (kcal mol -1 )
M3	$14\pm 9$	$-7.10$	$-5.90$
M6	$25\pm 7$	$-6.30$	$-6.00$
M7	$14\pm 9$	$-7.40$	$-6.20$
M10	$8\pm 3$	$-6.50$	$-6.40$
M11	$340\pm 8$	$-6.10$	$-5.80$

Appendix

Table 6: Published reference ligands for TBXT pocket A ′ . Apparent dissociation constants

K_{\mathrm{d}}

are the spr values reported by Newman et al. [ 33 ] , given as value

\pm

standard error. Docking scores were calculated here under the common retrospective protocol; lower values indicate more favorable predicted binding. These are the only compounds in this case study with experimental measurements.

Run	Seg.	Tool	Timestamp	Key parameters	Ligands
Target preparation and pocket detection: evidenced by output artifacts
–	A	Structure input	2026-08-09 03:55:59	7ZKF.fasta , A1I1T_ref.sdf	–
–	A	Target registration	2026-08-09 03:56:07	db_get_or_create_target	–
–	A	P2Rank pocket prediction	2026-08-09 03:56:13	version 2.5.1; rank-1 $p=0.015$ , distal	–
–	A	Pocketeer pocket detection	2026-08-09 03:56:13	3 calls; default: 1 pocket, relaxed: 29 pockets	–
–	A	Pocket definition	2026-08-09 03:57:49	centroid computed: $(13.653,-22.965,-6.120)$	–

Appendix

Table 7: Tool invocations for the TBXT campaign. The lower block is taken from the Fármaco Atlas run table. The upper block covers target preparation and pocket detection, whose timestamps were recovered from the archived chat transcript. The resulting pocket centre,

(13.653,-22.965,-6.120)

, was used for all sampling runs. Segment A precedes the operator restart and segment B follows it. The ligands column reports the number of registered ligands; a dash indicates that no new ligands were added. Timestamps are in utc .

Co-crystal	$K_{\mathrm{i}}$ ( $\mu$ M)	Original rank	Rank-1 rmsd (Å)	Best of 10 (Å)	Type Jaccard	Residue Jaccard
9C84	1.31	6,607	0.87	0.87	0.56	0.56
9C6P	0.86	2,216	3.27	3.06	0.42	0.50
9DHL	0.95	8,574	2.76	2.57	0.36	0.64
9C81	322.96	1,544,474	5.55	5.55	0.21	0.23

Appendix

Table 8: Pose and interaction recovery for the four AmpC co-crystal ligands. Ligands were docked into a receptor prepared from pdb 1L2S; the co-crystal structures served as validation references. rmsd were computed over the maximum common substructure after protein superposition, covering 21–22 of 23–24 ligand heavy atoms, and reproduced to the reported precision. Jaccard indices compare plip contacts of the rank-1 docked pose with the union of contacts from both crystal chains, using residue-and-interaction-type or residue alone and omitting chain identity.

K_{\mathrm{i}}

denotes the inhibition constant; Best of 10 is the lowest ligand rmsd among ten docked poses. Type Jaccard compares residue-and-interaction-type contacts, whereas Residue Jaccard compares contacted residues. See Sec. C.3 for mapping and reference-selection details.

Rank	Compound	Generator	$v$	$\mathrm{pIC}_{50}$	predicted_pic50	Binding prob.	Confidence	iptm
1	1718	PrexSyn	0.543	5.457	7.443	0.213	0.915	0.761
2	1721	PrexSyn	0.777	5.223	7.124	0.410	0.966	0.898
3	1663	PocketXMol	0.781	5.219	7.118	0.363	0.958	0.921
4	1572	MolMIM	0.898	5.102	6.959	0.160	0.885	0.648
5	1527	MolMIM	0.973	5.027	6.857	0.248	0.948	0.831
6	1522	MolMIM	1.438	4.563	6.223	0.277	0.972	0.937

Appendix

Table 9: Boltz-2 affinity predictions for the ten shortlisted generated compounds against pdb 1L2S. Compounds are ranked by predicted affinity.

v

is the returned affinity_pred_value ;

\mathrm{pIC}_{50}=6-v

is the scale used throughout this work, and predicted_pic50 is the backend-native field, equal to

1.364\,(6-v)

. Binding prob. is the predicted binding probability, and Confidence is the returned model confidence score. The table also reports the iptm . The generator column identifies each compound’s origin in the recorded workflow and is not part of the Boltz-2 output. These are model outputs for compounds that were never synthesized or assayed; they are not measured potencies and do not establish AmpC activity.

		With database			No database
Turn	Task	Calls	USD	s	Calls	USD	s
1	Dataset ingestion	11	0.25	84	5	0.14	48
2	Docking rank vs. activity	14	0.37	112	5	0.27	77
3	Selection strategy	5	0.19	42	2	0.13	38
4	Retrospective selection	34	0.78	1153	35	2.03	1128
5	Pocket detection	15	1.07	501	5	1.22	218

Appendix

Table 10: Resource comparison for the ten original AmpC turns and successful Boltz-2 completions. Model calls are the number of language-model invocations recorded in the trace for that turn, including those made by sub-agents. Costs are the recorded totals in USD; latency is the server-side turn time in seconds, defined in Sec. C.3 . The Boltz-2 completion is indexed as turn 12 with the database and turn 11 without it; each total covers eleven completed executions and excludes failed and empty recovery attempts. A dash denotes an unavailable model-call count for the database-backed completion and its expanded total. The task labels include admet triage. Each configuration was executed once.

Rank	Candidate	$\mathrm{pIC}_{50}$	Binder probability	Clash atoms	hERG	dili
1	catalogue_003	4.66	0.226	0	0.047	0.013
2	catalogue_001	3.56	0.152	1	0.074	0.017
3	pocketxmol_002	5.50	0.152	1	0.860	0.976
4	pocketxmol_003	4.80	0.098	2	0.610	0.809
5	prexsyn_004_05	4.42	0.207	2	0.863	0.403

Appendix

Table 11: First five candidates from the final KpThiL ranking, all of which passed the tps -pocket heuristic. Converted Boltz-2

\mathrm{pIC}_{50}

, binder probability, and hERG and dili scores are computational predictions, not experimental measurements or calibrated probabilities for these candidates. Clash atoms are ligand heavy atoms whose nearest experimental protein heavy atom was less than 2 Å away. The ranking combines pocket/clash tiers with a weighted score, including additional penalties for hERG, Ames, or dili scores of at least 0.8. Rank 1 was designated a tps -like reference by the agent; ranks 3–5 require synthesis or availability checks.

	Single edge	Network
User turns	4	3
Tool calls	29	25
llm calls	38	24
Total tokens	344,656	396,404
llm cost (USD)	0.73	0.95
Azulene Studio jobs submitted / completed	1 / 1	6 / 5

Appendix

Table 12: User turns and Azulene Studio settings of the two TYK2 sessions. Costs are recorded llm costs in USD and exclude Azulene Studio and other service costs. Cached input tokens are included in the token totals.

Quantity	Recorded value
User turns	28
Tool calls	264
llm calls	467
Total tokens	17,454,238
Cached input tokens	13,686,292
llm cost (USD)	9.82

Appendix

Table 13: Recorded resources and protocol of the TYK2 design session. Costs are recorded llm costs in USD and exclude Azulene Studio, Rowan, and other scientific services. Total tokens include cached input tokens. Simulation lengths are the effective submitted settings, rather than tool defaults; false indicates that protonation was not reassigned.

System	Task	Principal scope
AmpC	Hit ADMET comparison	Predict and compare properties and toxicity risks for 50 compounds.
AmpC	Docking and pose assessment	Assess binding predictions for 20 compounds, with at most three poses each.
AmpC	Hit triage	Apply potency, property, and structural-alert criteria to the supplied screening table.
AmpC	Binding-pocket detection	Identify pockets in PDB 1L2S and compare them with the active site.
AmpC	Screening analysis	Relate docking scores, ranks, and selection strategies to measured hit rates.
AmpC	Reference binding-mode comparison	Reassess prior nominations against the catalytic-site reference ligand.

Appendix

Table 14: Coverage of the 22-task benchmark. All configurations receive the same instructions and original inputs for a task; the fresh-context AmpC follow-up instead inherits its own configuration’s preceding workspace. Counts indicate the requested task scope.

	Fármaco	Plain tool calls	Codex
Agent runs	66	66	66
Mean overall score	8.67	8.30	8.41
95% confidence interval	[8.27, 9.03]	[7.76, 8.79]	[7.89, 8.91]
Pass judge sessions	151/198	123/198	122/198
Median latency (s)	243	154	210
Median model cost (USD)	0.195	0.150	0.296

Appendix

Table 15: Aggregate benchmark results. Scores are means of per-run median overall scores, with 95% confidence intervals from task resampling. Pass counts use judge sessions; all other resource summaries are medians over agent runs. Costs cover the agent’s model calls in US dollars (USD); search-adjusted cost subtracts the domain agents’ search-model charges. Peak context counts input tokens in the largest model call, including cached input; cumulative tokens sum input and output over calls.

Figure 9: Benchmark quality and resource use across agent configurations. Small markers show individual runs; larger markers show the mean overall judge score or median latency, model cost, or peak context. Vertical bars show 95% confidence intervals estimated by resampling tasks. Circles denote Fármaco, squares denote plain tool calls, and triangles denote Codex. The latency, cost, and peak-context axes are logarithmic. Peak context is the largest recorded input-token count for one model call, including cached input.

Figure 10: Judge sub-scores and verdicts. The left panel shows mean task-completion, scientific-correctness, evidence-consistency, and overall scores, using each run’s median over three judge sessions. Bars show 95% confidence intervals from bootstrap resampling of tasks. Circles denote Fármaco, squares denote plain tool calls, and triangles denote Codex. The right panel counts 198 judge sessions per configuration, with displayed percentages rounded to whole numbers. Pass denotes work accepted as delivered, partial denotes useful but incomplete or flawed work, and fail denotes an unsatisfactory result. The judge assigned the overall score directly from its assessment of the complete delivered result.

Figure 11: Per-task comparison of quality, latency, cost, and peak context. Each row is one of the 22 scientific tasks. Markers show the mean overall judge score or median resource metric over the runs of a task and configuration; grey lines span the configurations’ values. Circles denote Fármaco, squares denote plain tool calls, and triangles denote Codex; overlapping markers are drawn in that front-to-back order. Every task–configuration pair has three runs. The latency, cost, and peak-context axes are logarithmic; cost is in US dollars per run and peak context is in input tokens, with k denoting thousands.

Figure 12: Overall judge score versus latency, model cost, and peak context. Small markers show the mean overall score and median resource metric for one task and configuration. Large markers show the configuration-wide mean score and median resource metric; horizontal and vertical lines show their respective 95% task-bootstrap confidence intervals. Circles denote Fármaco, squares denote plain tool calls, and triangles denote Codex. Resource axes are logarithmic; peak context counts the largest model input including cached input.

	Fármaco	Plain tool calls	Codex
Latency (s)	$218\pm 71$	$483\pm 211$	$380\pm 14$
Model calls	$24.0\pm 0.0$	$14.3\pm 4.9$	$49.0\pm 11.1$
Input tokens ( $\times 10^{3}$ )	$113\pm 10$	$221\pm 94$	$1{,}107\pm 331$
Output tokens ( $\times 10^{3}$ )	$5.4\pm 0.4$	$6.4\pm 2.7$	$11.3\pm 1.6$
Total tokens ( $\times 10^{3}$ )	$119\pm 9$	$227\pm 97$	$1{,}118\pm 333$
Cached share of input (%)	$73.9\pm 4.8$	$93.2\pm 0.8$	$95.4\pm 2.2$

Appendix

Table 16: Execution efficiency and result quality for the shape and pharmacophore benchmark task. Values are means

\pm

sample standard deviations over the three benchmark runs per configuration. Token counts include cached input, and model costs cover all recorded model calls in US dollars (USD). Each run’s overall score is the median of its three judge sessions; pass counts use all nine sessions per configuration.

	Fármaco	Plain tool calls	Codex
Compounds ranked per run	49	49	49
Spearman rank correlation	$0.941\pm 0.020$	$0.946\pm 0.002$	$0.906\pm 0.123$
Top-10 overlap	$7.7\pm 0.6$	$7.7\pm 0.6$	$8.0\pm 1.0$
Top-rank agreement	0/3	3/3	3/3

Appendix

Table 17: Agreement of the benchmark shape rankings with the independently sampled reference. Correlation and overlap are means

\pm

sample standard deviations over three runs per configuration. Top-10 overlap counts shared compounds, and top-rank agreement counts runs that select the reference’s first-ranked compound.

	Fármaco database	Plain tool calls
User turns	1	6
llm calls	51	40
Tool calls	29	284
Input tokens ( $\times 10^{3}$ )	447	4,376
Cached input tokens ( $\times 10^{3}$ )	351	3,732
Output tokens ( $\times 10^{3}$ )	20	51

Appendix

Table 18: Agent execution for the 256-ligand STK17B task. Both configurations received the same initial instruction and completed the predictions, ranking, and top-five interaction analysis. User turns include the initial instruction and follow-up prompts. Token counts are cumulative across llm calls; cached input tokens are included in input and total tokens. Costs are recorded llm costs in USD and exclude scientific service costs. Boltz-2 sampling settings differed between configurations, as described in the text.

Figure 13: Predicted binding pose of the top-ranked STK17B ligand. Ligand carbons are shown in yellow, protein carbons and cartoon in grey, and heteroatoms by element. Selected interacting residues are shown in ball-and-stick representation. Visualization prepared with UCSF ChimeraX [ 69 ] .

Drug Discovery

Drug discovery identifies, designs, and tests molecules that can safely affect disease biology.

See all 173 papers

Oct 12, 2026

MADDIE: Multi-Agent Drug Discovery Engine

Drug discovery remains a resource-intensive and time-consuming process, involving billions of dollars and a decade of development to advance a drug to market. While many in silico tools have accelerated this process, operating them still demands substantial time, effort, and expertise from medicinal chemists. Autonomous systems built on Large Language Models (LLMs) have emerged as a promising solution, yet current designs limit their performance. Existing single-agent systems scale poorly to long-horizon, multi-step, and realistic tasks, whereas multi-agent systems typically execute predefined workflows that cannot capture the iterative, non-linear nature of a real drug discovery campaign. Here, we present MADDIE, a multi-agent drug discovery engine. It comprises a central agent, Orchestrator, and four specialized worker agents: HitIdentifier, LeadOptimizer, Synthesizer, and Evaluator, each representing a critical step in a drug discovery campaign. Under Orchestrator's orchestration and coordination, these worker agents are composed into flexible workflows, enabling MADDIE to adapt its strategy and navigate the drug discovery pipeline end-to-end. Across 50 protein targets, MADDIE produces molecules satisfying key property requirements and their synthesis routes, outperforming existing agentic systems, generative models, and general-purpose LLMs by a wide margin. Our analysis reveals how these agents collaborate and that MADDIE can employ different strategies across targets to achieve the campaign's goal, while still being efficient in time and tokens. Together, they demonstrate MADDIE as a low-cost, effective, and efficient autonomous engine for accelerating drug discovery. Code and dataset is available via https://github.com/ninglab/MADDIE.

Ye Liu, Reza Averly, Avery Meyer +4

Biomedical Informatics, The Ohio State University. · Computer Science and Engineering, The Ohio State University. · Medicinal Chemistry and Pharmacognosy, The Ohio State University. +1

Drug Discovery Multi-Agent LLM Systems LLM Agent Orchestration

Ask PDF

Oct 5, 2026

From Benchmark to Bench: Can Agents Survive Real-World Drug Discovery?

Agentic systems increasingly coordinate molecular-design tools, but it is unclear which layer of the stack limits outcomes on real projects. We developed MAGI, an open modular agent that authors objectives, launches and monitors optimization, interprets structure--activity relationships, and revises its strategy accordingly. MAGI generates molecules either directly through the LLM or by delegating to REINVENT 4, with scoring services interchangeable behind a common contract. We tested it across nine retrospective lead-optimization campaigns from three pharmaceutical companies, replayed under fixed temporal cutoffs. Both routes produced valid structures: LLM proposals stayed closer to local chemistry and reached comparable or higher primary activity in fewer operations, whereas REINVENT explored broader chemical space. Whether a campaign met its objective depended on the predictive models, not on the generation route: attainment followed model accuracy on the chemistry proposed, dropping once that chemistry moved outside the model's applicability domain. Separately, a blinded evaluation asked whether the MAGI's output could pass as expert work: chemists were not able to discriminate agentic proposals from held-out compounds, and judged the SAR reasoning broadly plausible yet incomplete. Together, these results position MAGI as a coordination layer pluggable into existing computational chemistry workflows. The ceiling on real projects, however, remains currently set by scorer applicability rather than by tool orchestration.

Pierre Llompart, Levent Guner, Helen Lai +2

AMD Silo AI, Stockholm, Sweden · Molecular AI, Discovery Sciences, R&D, AstraZeneca, Cambridge, UK · Molecular AI, Discovery Sciences, R&D, AstraZeneca, Gothenburg, Sweden +1

AI Agents for Scientific Discovery LLM Agent Orchestration Molecular Generation Drug Discovery

Ask PDF

Oct 3, 2026

Toward a Locally Deployable Agentic Co-Scientist: Small-Model Planning for Early-Stage Drug Discovery

Early-stage computational drug discovery requires coordinating heterogeneous scientific tools across multi-step workflows. We present a lightweight, tool-augmented framework in which a locally deployable compact language model plans calls to 18 modular tools. A Unified Molecular Schema maintains shared molecular records, while a plug-in interface supports tool replacement and extension. We construct 1,263 manually refined query-plan pairs through workflow-graph path coverage and apply LoRA fine-tuning to three compact model families. Under the query-level split, all fine-tuned models generate fully parseable and schema-compliant plans on 47 held-out cross-group queries. Llama 3.2-3B achieves a tool-selection F1 of 0.998, sequence exact match of 0.979, and argument F1 of 0.960. Under the stricter workflow-grouped split, which excludes identical ordered tool sequences across partitions, sequence exact match reaches 0.452 to 0.548, highlighting the remaining difficulty for compact models in generating complete workflow paths unseen during training. These results demonstrate the feasibility of compact, locally deployable planning while identifying compositional generalization as an important direction for further improvement.

Tian Liang, Jiayu Chang, Alejandro F. Frangi +2

Division of Pharmacy and Optometry, University of Manchester, UK · Department of Computer Science, University of Manchester, UK · Centre for Computational Imaging and Modelling in Medicine (CIMIM), University of Manchester, UK +4

LLM Tool Use LLM Agent Planning Large Language Model-Guided Scientific Discovery Small Language Models Drug Discovery

Ask PDF

Abstract

Drug Discovery

Explore similar work

MADDIE: Multi-Agent Drug Discovery Engine

From Benchmark to Bench: Can Agents Survive Real-World Drug Discovery?

Toward a Locally Deployable Agentic Co-Scientist: Small-Model Planning for Early-Stage Drug Discovery