Pith. sign in

REVIEW 3 major objections 3 minor 37 references

Agentic drug-discovery frameworks fail at peptide, in vivo, ML-training, small-team, and multi-objective workflows; the paper shows the bottleneck is architectural, not a lack of model knowledge.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 02:45 UTC pith:U6CU3YP4

load-bearing objection Useful practitioner-grounded gap analysis with a genuinely new diagnostic probe, but the 'architectural rather than epistemic' conclusion outruns the evidence. the 3 major comments →

arxiv 2602.10163 v1 pith:U6CU3YP4 submitted 2026-02-10 q-bio.QM

Beyond SMILES: Evaluating Agentic Systems for Drug Discovery

Edward Wijaya This is my paper
classification q-bio.QM
keywords agentic AIdrug discoverypeptide therapeuticsprotein language modelsin vivo pharmacologycapability gapsmulti-objective optimizationLLM evaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that current agentic AI systems for drug discovery—LLM-based assistants that orchestrate chemistry tools—are systematically built around small-molecule, big-pharma workflows and break down when asked to handle peptide therapeutics, animal-study data, model training or reinforcement learning, small-team budgets, or safety-versus-efficacy trade-offs. Evaluating six agent frameworks against 15 task classes drawn from real peptide-drug projects, the author finds that no framework fully supports any of the 15 classes and that five gaps account for nearly all of the missing coverage. A paired knowledge-probing experiment then isolates where the failure lives: four frontier LLMs answer peptide questions as accurately as small-molecule questions, even though no agent exposes any peptide-aware tools. The paper concludes that the expertise already exists inside the foundation models and that the fix is architectural: build integration pipelines that surface protein language models, ML training loops, in vivo data fusion, and multi-objective reasoning, rather than retraining the models. If right, the roadmap for next-generation frameworks is a capability matrix plus five design requirements, not another foundation model.

Core claim

On its own terms, the paper's central discovery is a diagnosis: the six leading agentic frameworks, assessed against a capability matrix of 15 drug-discovery task classes, show zero full support for any class and coverage scores between 0 and 16.7 percent. Five gaps emerge: small-molecule representation bias (no protein language models or peptide prediction), no bridge between in vivo and in silico data, no support for ML training or reinforcement learning as first-class operations, resource assumptions tied to large pharma, and single-objective optimization that ignores Pareto trade-offs. The paired probe shows the models are not the limiting factor: across 200 matched question pairs, pepti

What carries the argument

The carrying mechanism is a two-part evaluation: (1) a capability matrix that scores six agent frameworks against 15 task classes (from peptide bioactivity prediction to behavioral phenotyping) across five dimensions—molecular representation, computational paradigm support, data modality integration, resource assumptions, and optimization framework—with full/partial/not-supported ratings; and (2) a paired knowledge-probing experiment in which four frontier LLMs answer 50 matched small-molecule/peptide question pairs scored on a 0–3 rubric. The matrix establishes that the gaps are systemic clusters of zero coverage; the probe establishes that the underlying models already possess the relevant

Load-bearing premise

The conclusion that the bottleneck is architectural rather than epistemic assumes that an LLM's correct answers to short peptide questions would translate into successful multi-step tool orchestration if the tools existed; the probe never exercises tool use, and its own authors note the 50-pair design has limited statistical power.

What would settle it

Give one agent framework two matched arms—the same LLM and the same 15 task classes, one arm with only small-molecule tools and one arm with peptide tools, protein language models, and an ML training loop—and compare end-to-end completion rates. If completion rates on peptide task classes remain near zero even after the framework is given peptide-aware tools and fine-tuning primitives, while the LLM's peptide Q&A scores stay high, then missing tools were not the whole story and the diagnosis needs revision; if completion rates rise, the paper's claim is supported.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the diagnosis is right, improving agents means building integration pipelines—protein-language-model fine-tuning, ML/RL training loops, workflow graphs, active-learning loops—rather than training larger or better foundation models.
  • The capability matrix can serve as a practical benchmark: any practitioner can score a framework against the 15 task classes and see immediately where it will fail in peptide or in vivo work.
  • Small-biotech users should expect few-shot, transfer, and active-learning support plus batch-mode operation as prerequisites; agents that assume pharma-scale data and compute will not generalize to their workflows.
  • Multi-objective support is not optional: single-objective or weighted-sum optimization hides the safety-efficacy-stability trade-offs that actually drive candidate selection, so next-generation agents must expose Pareto frontiers and uncertainty.
  • Because peptide knowledge is already in the models, closing the first gap does not require retraining; it requires architectures that surface peptide-aware tools and sequence-native workflows.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the probe tests static question-answering, not multi-step agent execution; a sharper test would give an agent peptide-aware tools and see whether end-to-end task completion rises. Until then, "architectural rather than epistemic" is best read as "the models are not the demonstrated bottleneck in short-answer reasoning."
  • Editorial inference: the same probe design could be extended to other under-supported modalities named in the paper—antibody CDR loops, nucleic-acid therapeutics, imaging phenotypes—to see whether the stranded-knowledge pattern generalizes.
  • Editorial inference: the capability matrix could be repurposed as a living community benchmark, with the 15 task classes scored by multiple raters and updated as frameworks add support; the paper itself notes single-rater assessment as a limitation, and independent scoring would sharpen the gap analysis.
  • Editorial inference: if the active-learning result (roughly one-third fewer assays in the author's peptide projects) holds up, then architecture-level fixes have a directly monetizable consequence—fewer synthesis-and-assay cycles per lead—which gives small biotechs a concrete incentive to demand these features.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper evaluates six existing agentic drug-discovery frameworks (ChemCrow, Coscientist, PharmAgents, ChatInvent, MADD, DiscoVerse) against 15 task classes drawn from peptide therapeutics, in vivo pharmacology, and resource-constrained small-biotech practice. It reports a capability matrix and identifies five gaps: small-molecule representation bias, absence of in vivo–in silico integration, limited ML/RL paradigm support, large-pharma resource assumptions, and single-objective optimization. A paired knowledge-probing experiment on four frontier LLMs is used to argue that the bottleneck is architectural rather than epistemic, because the LLMs answer peptide questions at least as well as small-molecule questions. The paper then derives five design requirements and proposes a capability matrix for next-generation frameworks.

Significance. The paper is a useful practitioner-driven gap analysis. Its main potential contribution is to refocus the field from retraining foundation models toward building integration pipelines. The task taxonomy and the paired knowledge-probing design are concrete, and the supplementary release of questions, raw responses, and scoring data supports reproducibility. The detailed per-framework evidence in Appendix B is also valuable. However, the central 'architectural rather than epistemic' inference is supported only indirectly, and the capability matrix rests on a single rater's subjective assessment. If the claims are carefully scoped to the chosen task set and to the probed LLMs, the design requirements form a plausible roadmap; as written, the paper overgeneralizes.

major comments (3)
  1. [§2.5, §3.1.2] The diagnostic that underpins the 'architectural rather than epistemic' conclusion is mismatched to the systems it diagnoses. The probe tested Kimi K2.5, DeepSeek V3.2, Qwen 3 Next 80B, and Gemini 3 Flash, but the six frameworks are not shown to use these models; ChemCrow and Coscientist are documented with GPT-4 (refs [1,2]) and the others do not disclose model versions. The sentence 'The knowledge exists; no current agent provides a pathway to surface it' (§3.1.2) therefore overgeneralizes: for an agent running an older or less peptide-capable model, the observed zero coverage could be epistemic, not architectural. Moreover, the probe is a static short-answer Q&A task, not an agentic multi-step tool-orchestration task; a null peptide deficit in Q&A does not rule out bottlenecks in tool availability, planning, or execution. Please reframe the conclusion as applying to the probed frontie
  2. [§2.2, §2.4, Table 2] The five gaps are substantially predetermined by the evaluation design. §2.2 states that task classes where agents already perform well were deliberately excluded, and the 15 classes are drawn from the author's peptide/in vivo/resource-constrained workflows. The capability matrix then reports zero coverage for many of these classes, and the five gaps are clusters of that zero coverage. As a scoped gap analysis ('unsupported task classes in this practitioner-derived set') this is coherent, but the paper's language of 'five systematic capability gaps' in agentic drug discovery (Abstract, §3) overstates the generality. The single-rater assessment (§2.4, §5.1) further limits the matrix as evidence. Please either broaden the task-class selection to include known-positive controls and independent raters, or revise the claims to match the deliberately selective scope.
  3. [§5.1, Table 17] The knowledge-probing result depends on LLM-as-judge scores whose expert agreement is low (quadratic-weighted κ=0.22). The paper argues that the paired within-model comparisons are robust to calibration differences because disagreements concentrate at the 2/3 boundary and are balanced across domains, but it does not report the expert-only paired statistics. Please provide the aggregate gap and Wilcoxon results on the expert-scored 80-response subset (or on a fully expert-scored set) to substantiate the robustness claim. The Opus sensitivity check (κ=0.78) shows judge-model stability but is still an LLM judge, not independent ground truth.
minor comments (3)
  1. [§3.1.2, Table 3, Table 16] The sign convention for the aggregate gap is inconsistent. Table 3 defines Gap = peptide − SM (positive indicates peptide advantage); Table 16 defines Gap = SM − peptide (negative indicates peptide advantage). The text reports −0.115 as 'peptide marginally higher', which matches only the Table 16 convention. Please harmonize the definition across the text, tables, and Figure 2 caption.
  2. [§3.4.3] The text 'A 70' at the end of the transfer-learning paragraph appears truncated; it should be completed or removed.
  3. [Figure 2 caption] The caption states 'aggregate gap = −0.115, 95% CI: [−0.255, 0.02]' and 'at or above small-molecule levels'. Under the Table 3 convention, a negative gap means peptide scores are slightly lower; reconcile the caption with the chosen sign convention.

Circularity Check

1 steps flagged

Five capability gaps are entailed by task selection; 'architectural bottleneck' rests on a mismatched knowledge probe.

specific steps
  1. self definitional [§2.2 Task Class Definition; §3 Results: Five Capability Gaps (Table 2 intro)]
    "We deliberately exclude task classes where agents already perform well, such as literature synthesis, retrosynthesis planning, and molecular property lookup, focusing instead on capabilities that remain unsupported. ... The five gaps identified in the following subsections emerge directly from this matrix: clusters of zero-coverage task classes that share underlying architectural limitations."

    The central result, the five capability gaps, is not an independent discovery but a consequence of the evaluation design. §2.4 defines gaps as 'task classes with zero or minimal framework coverage,' and §2.2 constructs the 15 task classes as capabilities that are already 'unsupported' by construction, explicitly excluding areas where agents perform well. Therefore the finding that all six frameworks show zero or near-zero coverage for the selected tasks is forced: the input task set was chosen to contain only tasks expected to be unsupported, and the output is a clustering of those zero-coverage cells. The matrix documents genuine coverage details, but the headline conclusion that current agentic systems have systematic capability gaps—and the corresponding design requirements—are built in

full rationale

The paper's gap taxonomy is circular in a structural sense: the 15 task classes were deliberately selected to exclude tasks where agents already perform well, and gaps were then identified as zero-coverage clusters among those same classes. This makes the five 'critical capability gaps' a re-description of the task-selection criteria rather than an empirically discovered generalization. The knowledge-probing experiment is an independent, welcome check on one premise (frontier LLMs know peptide pharmacology), but it does not break the circularity: it tests Kimi K2.5, DeepSeek V3.2, Qwen 3 Next 80B, and Gemini 3 Flash, whereas the six evaluated frameworks largely used different models (e.g., GPT-4 in ChemCrow/Coscientist), so the inference that 'no current agent provides a pathway to surface it' assumes without evidence that the agent-internal LLMs share the probed models' peptide competence. Additional concerns—single-rater capability assessments, low expert–automated scorer agreement (κ=0.22), and the sign inconsistency in the aggregate gap (−0.115 reported as pro-peptide while Table 3's convention gives positive as peptide advantage)—affect reliability but are not themselves circularity. No load-bearing self-citation chain or imported uniqueness theorem was found; the circularity is concentrated in the task-selection-to-gap inference. Score 6 reflects that the five-gap result is partially forced by construction, while the matrix itself and the probe retain some independent evidentiary content.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The paper introduces no new physical or mathematical entities. Its load-bearing assumptions are representativeness of the task set, accuracy of single-rater judgments, transfer of Q&A probe results to agent architecture, and the validity of LLM-as-judge scoring. The free parameters are hand-chosen evaluation weights and subset sizes that shape the reported coverage scores.

free parameters (3)
  • Partial-support weight = 0.5
    Chosen by hand in §2.4; directly determines the coverage percentages in Table 2 and therefore which task classes count as 'covered'.
  • Expert-validation subset size = 80 (20%)
    Stratified 20% subset for expert validation of LLM judge scores; affects statistical power and the reported kappa = 0.22.
  • Sensitivity-check subset size = 40 responses
    Opus rescoring subset used to claim judge-model stability; small sample limits the strength of that claim.
axioms (6)
  • domain assumption The 15 task classes are representative of real-world drug discovery workflows.
    The task classes are derived from one practitioner's projects at one small biotech (§2.2, Appendix A); no external benchmark validates their representativeness.
  • domain assumption Single-rater capability assessments are accurate.
    All framework-task ratings were performed by the author alone (§2.4); no inter-rater reliability is reported.
  • ad hoc to paper LLM Q&A probing measures the epistemic bottleneck for agentic workflows.
    The probe tests short-answer pharmaceutical knowledge, not multi-step tool use or planning; the conclusion 'architectural rather than epistemic' depends on this transfer (§2.5, §3.1.2).
  • domain assumption The six selected frameworks represent the space of agentic drug-discovery systems.
    Selection criteria exclude foundation models, benchmarks, and single-paradigm ML tools (§2.1); other systems might narrow the gaps.
  • ad hoc to paper Excluding task classes where agents already perform well is appropriate for a gap analysis.
    §2.2 explicitly excludes such tasks, which by construction lowers measured coverage and makes the gaps easier to find.
  • ad hoc to paper LLM-as-judge scores are valid despite low expert agreement.
    The paper asserts paired comparisons are robust to calibration differences (§5.1), but expert-automated kappa = 0.22 and no derivation for robustness is given.

pith-pipeline@v1.3.0-alltime-deepseek · 22980 in / 11439 out tokens · 113464 ms · 2026-08-03T02:45:58.231322+00:00 · methodology

0 comments
read the original abstract

Agentic systems for drug discovery have demonstrated autonomous synthesis planning, literature mining, and molecular design. We ask how well they generalize. Evaluating six frameworks against 15 task classes drawn from peptide therapeutics, in vivo pharmacology, and resource-constrained settings, we find five capability gaps: no support for protein language models or peptide-specific prediction, no bridges between in vivo and in silico data, reliance on LLM inference with no pathway to ML training or reinforcement learning, assumptions tied to large-pharma resources, and single-objective optimization that ignores safety-efficacy-stability trade-offs. A paired knowledge-probing experiment suggests the bottleneck is architectural rather than epistemic: four frontier LLMs reason about peptides at levels comparable to small molecules, yet no framework exposes this capability. We propose design requirements and a capability matrix for next-generation frameworks that function as computational partners under realistic constraints.

Figures

Figures reproduced from arXiv: 2602.10163 by Edward Wijaya.

Figure 1
Figure 1. Figure 1: The Agent Reality Gap in Drug Discovery. Left panel shows computational workflows [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Stranded Knowledge: LLM Peptide Competence vs Agent Capability. Mean scores [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Workflow Complexity: Small Molecules vs Peptides. Top: Small molecule workflow [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: What Agents Can and Cannot Process. Data types grouped by agent accessibility into [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: From LLM-Centric to Multi-Paradigm Orchestration. Top: Current LLM-centric archi [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Big Pharma vs Small Biotech: The Resource Gap. Comparative visualization of com [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The Pareto Frontier Agents Ignore. Two-dimensional scatter plot of candidate compounds [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Knowledge Probing: Per-Category Score Heatmap. Mean scores for each model across five [PITH_FULL_IMAGE:figures/full_fig_p046_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 6 canonical work pages · 1 internal anchor

  1. [1]

    Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes

    Daniil A. Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes. Autonomous chem- ical research with large language models.Nature, 624:570–578, 2023. doi: 10.1038/ s41586-023-06792-0. Coscientist system for autonomous synthesis planning and execution

    2023

  2. [2]

    Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D

    Andres M. Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D. White, and Philippe Schwaller. ChemCrow: Augmenting large-language models with chemistry tools.Nature Machine Intelligence, 6:525–535, 2024. doi: 10.1038/s42256-024-00832-8. arXiv:2304.05376 (preprint April 2023). GPT-4 orchestrating 18 chemistry tools

    Pith/arXiv arXiv 2024

  3. [3]

    Democratising real-world drug discovery through agentic AI.Drug Discovery Today, 31(2):104605, January 2026

    Jiazhen He et al. Democratising real-world drug discovery through agentic AI.Drug Discovery Today, 31(2):104605, January 2026. doi: 10.1016/j.drudis.2026.104605. PMID: 41548711. ChatInvent system deployed at AstraZeneca

    arXiv 2026

  4. [4]

    PharmAgents: Building a virtual pharma with large language model agents

    Bowen Gao, Yanwen Huang, Yiqiao Liu, Wenxuan Xie, Wei-Ying Ma, Ya-Qin Zhang, and Yanyan Lan. PharmAgents: Building a virtual pharma with large language model agents. arXiv preprint arXiv:2503.22164, March 2025. Multi-agent system for target identification and compound selection. 30

    Pith/arXiv arXiv 2025

  5. [5]

    Jaeger, Fan Zhang, Rory Pilgrim, Yossi Matias, Joelle Barral, David Fleet, and Shekoofeh Azizi

    Eric Wang, Samuel Schmidgall, Paul F. Jaeger, Fan Zhang, Rory Pilgrim, Yossi Matias, Joelle Barral, David Fleet, and Shekoofeh Azizi. TxGemma: Efficient and agentic LLMs for thera- peutics.arXiv preprint arXiv:2504.06196, April 2025. Google Research therapeutics-focused language model

    Pith/arXiv arXiv 2025

  6. [6]

    Solovev, Alina B

    Gleb V. Solovev, Alina B. Zhidkovskaya, Anastasia Orlova, et al. MADD: Multi-agent drug discovery orchestra.arXiv preprint arXiv:2511.08217, November 2025. Accepted to EMNLP 2025 Findings. Multi-agent collaboration for molecular design

    arXiv 2025

  7. [7]

    Doktorova

    Xiaochen Zheng, Alvaro Serra, Ilya Schneider Chernov, Maddalena Marchesi, Eunice Mus- vasva, and Tatyana Y. Doktorova. DiscoVerse: Multi-agent pharmaceutical co-scientist for traceable drug discovery and reverse translation.arXiv preprint arXiv:2511.18259, November

    arXiv

  8. [8]

    Shaheen E. Lakhan. The agentic era: Why biopharma must embrace artificial intelligence that acts, not just informs.Cureus, 17(5), May 2025. doi: 10.7759/cureus.83390. PMID: 40322603. Editorial on agentic AI in pharmaceutical industry

    doi:10.7759/cureus.83390 2025

  9. [9]

    AI agents in drug discovery.arXiv preprint arXiv:2510.27130, October 2025

    Srijit Seal et al. AI agents in drug discovery.arXiv preprint arXiv:2510.27130, October 2025. Comprehensive 45-page survey of agentic AI systems

    arXiv 2025

  10. [10]

    Evolutionary-scale prediction of atomic-level protein structure with a language model.Science, 379(6637):1123–1130, 2023

    Zeming Lin, Halil Akin, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model.Science, 379(6637):1123–1130, 2023. doi: 10.1126/science.ade2574. PMID: 36927031

    doi:10.1126/science.ade2574 2023

  11. [11]

    ProtTrans: Toward understanding the language of life through self- supervised learning.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10): 7112–7127, 2022

    Ahmed Elnaggar et al. ProtTrans: Toward understanding the language of life through self- supervised learning.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10): 7112–7127, 2022. doi: 10.1109/TPAMI.2021.3095381. Online July 2021, print October 2022

    arXiv 2022

  12. [12]

    Rodriques

    Odhran O’Donoghue, Aleksandar Shtedritski, John Ginger, Ralph Abboud, Ali Ghareeb, and Samuel G. Rodriques. BioPlanner: Automatic evaluation of LLMs on protocol planning in biology. InProceedings of EMNLP 2023, pages 2676–2694, 2023. doi: 10.18653/v1/2023. emnlp-main.162. arXiv:2310.10632

    Pith/arXiv arXiv 2023

  13. [13]

    Baker, Ziru Chen, Garrett Herb, Boyu Gou, Daniel Adu-Ampratwum, Xia Ning, and Huan Sun

    Botao Yu, Frazier N. Baker, Ziru Chen, Garrett Herb, Boyu Gou, Daniel Adu-Ampratwum, Xia Ning, and Huan Sun. ChemToolAgent: The impact of tools on language agents for chemistry problem solving.arXiv preprint arXiv:2411.07228, November 2024. Accepted to NAACL 2025 Findings

    Pith/arXiv arXiv 2024

  14. [14]

    Agentomics: An agentic system that autonomously develops novel state-of-the-art solutions for biomedical machine learning tasks.bioRxiv, 2026

    Vlastimil Martinek, Andrea Gariboldi, Dimosthenis Tzimotoudis, Mark Galea, Elissavet Zacharopoulou, Aitor Alberdi Escudero, Edward Blake, David ˇCech´ ak, Luke Cassar, Alessan- dro Balestrucci, and Panagiotis Alexiou. Agentomics: An agentic system that autonomously develops novel state-of-the-art solutions for biomedical machine learning tasks.bioRxiv, 20...

    work page doi:10.64898/2026.01.27.702049 2026

  15. [15]

    ML-Agent: Reinforcing LLM agents for autonomous machine learning engineering.arXiv preprint arXiv:2505.23723, 2025

    Zexi Liu, Jingyi Chai, Xinyu Zhu, Shuo Tang, Rui Ye, Bo Zhang, Lei Bai, and Siheng Chen. ML-Agent: Reinforcing LLM agents for autonomous machine learning engineering.arXiv preprint arXiv:2505.23723, 2025. doi: 10.48550/arXiv.2505.23723. Online RL training of LLM agents for autonomous ML engineering

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.23723 2025

  16. [16]

    Rethinking drug design in the artificial intelligence era.Nature Reviews Drug Discovery, 19:353–364, 2020

    Petra Schneider et al. Rethinking drug design in the artificial intelligence era.Nature Reviews Drug Discovery, 19:353–364, 2020. doi: 10.1038/s41573-019-0050-3. PMID: 31801986

    doi:10.1038/s41573-019-0050-3 2020

  17. [17]

    Applica- tions of machine learning in drug discovery and development.Nature Reviews Drug Discovery, 18(6):463–477, 2019

    Jessica Vamathevan, Dominic Clark, Paul Czodrowski, Ian Dunham, Edgardo Ferran, George Lee, Bin Li, Anant Madabhushi, Parantu Shah, Michaela Spitzer, and Shanrong Zhao. Applica- tions of machine learning in drug discovery and development.Nature Reviews Drug Discovery, 18(6):463–477, 2019. doi: 10.1038/s41573-019-0024-5. PMID: 30976107

    doi:10.1038/s41573-019-0024-5 2019

  18. [18]

    Xing, Hao Zhang, Joseph E

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. InAdvances in Neural Information Processing Systems, volume 36, 2023

    2023

  19. [19]

    Large language models generate functional protein sequences across diverse families.Nature Biotechnology, 41:1099–1106, 2023

    Ali Madani et al. Large language models generate functional protein sequences across diverse families.Nature Biotechnology, 41:1099–1106, 2023. doi: 10.1038/s41587-022-01618-2

    doi:10.1038/s41587-022-01618-2 2023

  20. [20]

    PepTune: De novo generation of ther- apeutic peptides with multi-objective-guided discrete diffusion

    Sophia Tang, Yinuo Zhang, and Pranam Chatterjee. PepTune: De novo generation of ther- apeutic peptides with multi-objective-guided discrete diffusion. InProceedings of the 42nd International Conference on Machine Learning (ICML), volume 267 ofProceedings of Ma- chine Learning Research. PMLR, 2025. doi: 10.48550/arXiv.2412.17780. Masked discrete diffusion ...

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.17780 2025

  21. [21]

    DeLisa, Jen-Tsan Ashley Chi, Ray Truant, Hector C

    Leo Tianlai Chen, Zachary Quinn, Madeleine Dumas, Christina Peng, Lauren Hong, Moi- ses Lopez-Gonzalez, Alexander Mestre, Rio Watson, Sophia Vincoff, Lin Zhao, Jianli Wu, Audrey Stavrand, Mayumi Schaepers-Cheu, Tian Zi Wang, Divya Srijay, Connor Monticello, Pranay Vure, Rishab Pulugurta, Sarah Pertsemlidis, Kseniia Kholina, Shrey Goel, Matthew P. DeLisa, ...

    2025

  22. [22]

    ProtGPT2 is a deep unsupervised lan- guage model for protein design.Nature Communications, 13:4348, 2022

    Noelia Ferruz, Steffen Schmidt, and Birte H¨ ocker. ProtGPT2 is a deep unsupervised lan- guage model for protein design.Nature Communications, 13:4348, 2022. doi: 10.1038/ s41467-022-32007-7. 32

    2022

  23. [23]

    Highly accurate protein structure prediction with AlphaFold.Nature, 596: 583–589, 2021

    John Jumper et al. Highly accurate protein structure prediction with AlphaFold.Nature, 596: 583–589, 2021. doi: 10.1038/s41586-021-03819-2. PMID: 34265844

    doi:10.1038/s41586-021-03819-2 2021

  24. [24]

    DeepLabCut: markerless pose estimation of user-defined body parts with deep learning.Nature Neuroscience, 21:1281–1289, 2018

    Alexander Mathis et al. DeepLabCut: markerless pose estimation of user-defined body parts with deep learning.Nature Neuroscience, 21:1281–1289, 2018. doi: 10.1038/s41593-018-0209-y. PMID: 30127430

    doi:10.1038/s41593-018-0209-y 2018

  25. [25]

    KEGG for taxonomy-based analysis of pathways and genomes.Nucleic Acids Research, 51(D1):D587–D592, 2023

    Minoru Kanehisa, Miho Furumichi, Yoko Sato, Masayuki Kawashima, and Mari Ishiguro- Watanabe. KEGG for taxonomy-based analysis of pathways and genomes.Nucleic Acids Research, 51(D1):D587–D592, 2023. doi: 10.1093/nar/gkac963. PMID: 36300620

    work page doi:10.1093/nar/gkac963 2023

  26. [26]

    Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles.Proceedings of the National Academy of Sciences, 102(43):15545–15550, 2005

    Aravind Subramanian et al. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles.Proceedings of the National Academy of Sciences, 102(43):15545–15550, 2005. doi: 10.1073/pnas.0506580102. PMID: 16199517

    doi:10.1073/pnas.0506580102 2005

  27. [27]

    Li, Shanghua Gao, Wanxiang Shen, Valentina Giunchiglia, Andrew Shen, Yepeng Huang, Zhenglun Kong, and Marinka Zitnik

    Pengwei Sui, Michelle M. Li, Shanghua Gao, Wanxiang Shen, Valentina Giunchiglia, Andrew Shen, Yepeng Huang, Zhenglun Kong, and Marinka Zitnik. Medea: An omics AI agent for therapeutic discovery.bioRxiv, 2026. doi: 10.64898/2026.01.16.696667. Omics AI agent with 20 tools for transcriptomics, protein networks, and pathway analysis

    doi:10.64898/2026.01.16.696667 2026

  28. [28]

    DiMasi, Henry G

    Joseph A. DiMasi, Henry G. Grabowski, and Ronald W. Hansen. Innovation in the pharma- ceutical industry: new estimates of R&D costs.Journal of Health Economics, 47:20–33, 2016. doi: 10.1016/j.jhealeco.2016.01.012. PMID: 26928437

    work page doi:10.1016/j.jhealeco.2016.01.012 2016

  29. [29]

    Stokes et al

    Jonathan M. Stokes et al. A deep learning approach to antibiotic discovery.Cell, 180(4): 688–702.e13, 2020. doi: 10.1016/j.cell.2020.01.021. PMID: 32084340

    doi:10.1016/j.cell.2020.01.021 2020

  30. [30]

    Combining generative artificial intelligence and on-chip synthesis for de novo drug design.Science Advances, 7(24), 2021

    Francesca Grisoni et al. Combining generative artificial intelligence and on-chip synthesis for de novo drug design.Science Advances, 7(24), 2021. doi: 10.1126/sciadv.abg3338. PMID: 34117066

    work page doi:10.1126/sciadv.abg3338 2021

  31. [31]

    Tan, Yang Liu, and Lei Xie

    Ryan K. Tan, Yang Liu, and Lei Xie. Reinforcement learning for systems pharmacology- oriented and personalized drug design.Expert Opinion on Drug Discovery, 17(8):849–863,

  32. [32]

    Daniel Reker, Petra Schneider, Gisbert Schneider, and J. B. Brown. Active learning for computational chemogenomics.Future Medicinal Chemistry, 9(4):381–402, 2017. doi: 10.4155/fmc-2016-0197. PMID: 28263088

    work page doi:10.4155/fmc-2016-0197 2017

  33. [33]

    Richard Bickerton, Gaia V

    G. Richard Bickerton, Gaia V. Paolini, J´ er´ emy Besnard, Sorel Muresan, and Andrew L. Hop- kins. Quantifying the chemical beauty of drugs.Nature Chemistry, 4:90–98, 2012. doi: 10.1038/nchem.1243. PMID: 22270643

    doi:10.1038/nchem.1243 2012

  34. [34]

    A multi-objective 33 molecular generation method based on Pareto algorithm and Monte Carlo tree search.Ad- vanced Science, 12(20):2410640, 2025

    Yifei Liu, Yiheng Zhu, Jike Wang, Renling Hu, Chao Shen, Wanglin Qu, Gaoang Wang, Qun Su, Yuchen Zhu, Yu Kang, Peichen Pan, Chang-Yu Hsieh, and Tingjun Hou. A multi-objective 33 molecular generation method based on Pareto algorithm and Monte Carlo tree search.Ad- vanced Science, 12(20):2410640, 2025. doi: 10.1002/advs.202410640. Pareto MCTS achieving 51.6...

    doi:10.1002/advs.202410640 2025

  35. [35]

    Preferential Multi-Objective Bayesian Optimization for Drug Discovery

    Tai Dang, Long-Hung Pham, Sang T. Truong, Ari Glenn, Wendy Nguyen, Edward A. Pham, Jeffrey S. Glenn, Sanmi Koyejo, and Thang Luong. Preferential multi-objective Bayesian optimization for drug discovery.arXiv preprint arXiv:2503.16841, 2025. doi: 10.48550/arXiv.2503.16841. Human-guided preferential multi-objective Bayesian optimization for virtual screenin...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2503.16841 2025

  36. [2022]

    PMID: 35510835

    doi: 10.1080/17460441.2022.2072288. PMID: 35510835

    arXiv 2022

  37. [2025]

    Roche-affiliated researchers