Pith.

Artificial Intelligence · arXiv:2610.23791 v1 · submitted 16 Oct 2026 · 66 references

DrugHarness: An Evolvable Agent Harness for Computational Drug Discovery

Yuxing Chen, Zhisheng Zhang, Cheng Tang, Wei Li, Minghui Zhang, Ziyan Huang, Shaocong Wei, Jiawei Chen

show 15 more authors

Junru Hu, Hongze Zhu, Jianghan Shen, Min Cen, Xinyuan Zhu, Yaxue Zhao, Jubo Wang, Yuqiang Li, Chun-Mei Feng, Zongyuan Ge, Rui Su, Jie Xu, Yirong Chen, Xi Xu, Ming Hu

The paper says

“Computational drug discovery requires agents to compose scientific operations and adapt to changing evidence and constraints. Here we introduce DrugHarness, an evolvable agent harness coupling composable scientific execution with conditional skill induction.”

more of the abstract

Its environment integrates over 80 scientific operations, biomedical resources and task-specific code, preserving associations among candidates, results and artifacts. From execution trajectories and outcome feedback, DrugHarness induces action hypotheses with observable applicability conditions, recommended actions and supporting evidence. An independent review agent evaluates candidate skills derived from these hypotheses through source verification and, when needed, targeted execution trials. Admitted skills guide subsequent tasks and are revised through feedback. Evaluation on MolBench and SMDD-Bench demonstrates capabilities in molecular screening, optimization and constrained design. Without evolution, DrugHarness exceeds reported MolClaw-CC results by 10.8 percentage points in affinity-comparison accuracy and 55% in mean top-three active hits on MolBench. Retrospective evaluation on skill-development tasks shows that evolution increases Qwen's affinity-comparison accuracy from 62.2% to 81.1% and optimization success from 79.5% to 87.2%. In a prospective HPK1 study, DrugHarness achieves a more favorable median docking score than Biomni among system-specific candidates. A synthesized compound achieves a cellular IC50 of 14.19 nM, a 2.64-fold potency improvement over the best input reference, with activity competitive with two selected compounds designed by experienced medicinal chemists. These results support conditional skill evolution for scientific agents and demonstrate a workflow connecting computational drug design with experimental validation.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 19 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

The passes disagreed. Reader: holds up. Skeptic: agrees with the reader. Referee: major revision.

Solid engineering contribution with transparent limitations; the skill evolution evidence is in-domain but the paper doesn't hide that. This line is wrong

4 major objections · 6 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 66 references say.

The paper says, at §4.3, Table 3; §5 (Limitations); Abstract

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.88):

“Skill evolution is evaluated on a single benchmark with a limited number of tasks under a retrospective, in-domain setting; its effectiveness at larger scales and generalization to unseen targets and task families remain to be es- tablished.”

Pith says: Objection 1 of 4 · major

The central claim about conditional skill evolution is supported only by a retrospective, in-domain evaluation on one benchmark (MolBench, 190 tasks). The authors acknowledge this limitation in §5 (lines [b149]–[b150]), but the abstract and conclusion present the evolution results as supporting a general mechanism for scientific agents.

more of objection 1

The HPK1 prospective study (§4.4) does not test evolved skills; it uses the base system without evolution. Therefore the claim that DrugHarness 'supports conditional skill evolution for scientific agents' (abstract) is overstated relative to the evidence. The authors should either (a) add a held-out task split, cross-benchmark transfer test, or prospective evolution experiment, or (b) substantially rephrase the claims throughout the paper to state that evolution benefits were observed only in-domain on MolBench and generalization remains to be established.

The paper says, at §4.3, Table 3

The review did not quote this passage.

Pith says: Objection 2 of 4 · major

Virtual screening Hit@3 drops from 60.0% (w/o Evol.) to 48.0% (Evol.), and Mean Hits@3 drops from 0.72 to 0.64, yet the text in §4.3 describes the revision as leading to improvements ('Revised runs achieved a top-three hit on 12 tasks'). This apparent contradiction needs explicit discussion. The authors attribute the degradation to a different metric focus? Please clarify the overall effect on virtual screening and whether the skill evolution was beneficial for this task family.

The paper says, at §3.2, Eq. (1)–(3); §3.3, Eq. (4); §3.4, Eq. (5)–(6)

The review did not quote this passage.

Pith says: Objection 3 of 4 · major

The skill learning loop relies entirely on LLM-based induction and LLM-based independent review (the learner and reviewer are both LLMs). The paper does not characterize the sensitivity of the evolved skills to the choice of LLM, the size of the batch B_t, or the budget allowance.

more of objection 3

Given that the MolBench tasks are small and share surface-level patterns, the observed improvements could reflect overfitting to quirks of the historical execution data rather than acquisition of transferable scientific strategies. The authors should provide cross-validation or a stability analysis (e.g., repeat induction with different random seeds or different LLMs).

The paper says, at §4.4, Figure 2

The review did not quote this passage.

Pith says: Objection 4 of 4 · major

The caption states 'Lower is favorable' for the docking score panel but docking scores are typically negative, so lower (more negative) is indeed favorable. However, the inhibition prediction panel (b) has positive scores (0–1) where lower is also described as favorable; the CYP and OATP predictions are adverse properties, so lower inhibition is favorable.

more of objection 4

This is correct but the caption could be clearer. More importantly, the comparison with Biomni is based on only 4 system-specific structures per group, after excluding 6 shared structures from 'ten evaluated entries each'. With such small sample sizes, the statistical significance of the median differences should be assessed (e.g., bootstrap confidence intervals).

6 minor objections

The paper says, at §4.1, Experimental Setup

Pith says: minor objection 1 of 6

The abbreviation 'N/R' appears in Table 2 without definition. Define 'N/R' as 'Not Reported' in the table note.

The paper says, at Table 5 and Table 3

Pith says: minor objection 2 of 6

The metric 'Mean GT rank↓' in Table 5 is used without defining 'GT'. In the text, 'GT' likely means 'ground truth', but it should be explicitly defined in the caption or in the metric definitions in Appendix A.1.

show 4 more

The paper says, at §4.3, paragraph beginning 'Optimization experience'

Pith says: minor objection 3 of 6

The sentence 'The learned skills achieve 30/37 correct affinity comparisons and 34/39 successful optimizations, approaching the fixed researcher-written references of 33/37 and 35/39, respectively.' This is an interesting comparison, but it is the first mention of 'researcher-written references'. Explain what these reference skills are and how they were constructed (perhaps in Appendix).

The paper says, at §4.2, paragraph 'Backbone and Runtime Comparisons'

Pith says: minor objection 4 of 6

The description 'DrugHarness/ Qwen3.7-Max attains 62.2% affinity accuracy' is for the configuration without evolution, but the evolution results in §4.3 use Qwen3.7-Max as well. Clarify the relationship: the evolution experiment uses the same backbone but with learned skills.

The paper says, at Appendix D.1, Table 7

Pith says: minor objection 5 of 6

The table footnotes 'Met3/4' and 'Met4/4' are informal. Replace with '3/4 targets met' and '4/4 targets met' for clarity.

The paper says, at Throughout

Pith says: minor objection 6 of 6

Several acronyms (e.g., SMDD, LO, IP, SH, FA) are used extensively without being defined on first use in the main text. Define them at first occurrence in §4.1.

What would settle it

Select a held-out subset of MolBench tasks (e.g., 20% of the 190 tasks, stratified by category and not used in any learning round) and apply the version of DrugHarness with the evolved skills from Table 3 to these held-out tasks, measuring the same metrics (accuracy, Hits@3, success rate). Compare to the no-evolution baseline on the same held-out tasks. If the evolved skills still show improvements comparable to those in Table 3, generalizability is supported. Alternatively, test the evolved skills on a different benchmark (e.g., the SMDD-Bench LO tasks) to see if transfer occurs. A null or reversed result would indicate task-specific overfitting.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 19 Oct 2026

  1. Reader holds up confidence high
  2. Skeptic partly agrees 22 s · 26,336 tokens read, 2,363 written
  3. Editor wrote the verdict above 13 s · 27,221 tokens read, 1,452 written
  4. Referee major revision: 4 major objections, 6 minor 40 s · 28,486 tokens read, 1,944 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

DrugHarness is a well-engineered system that integrates over 80 scientific tools with a skill learning loop. The base system without evolution already beats strong baselines like MolClaw-CC by 10.8 percentage points in affinity accuracy and 55% in mean Hits@3 on MolBench, and outperforms the original SMDD harness on lead optimization. That alone is a useful contribution: a composable, tool-rich harness that works well out of the box.

The novel part is the conditional skill induction and independent review pipeline. The idea of having an LLM induce action hypotheses with applicability conditions, then have a separate reviewer check sources and run targeted trials, then revise based on feedback, is a reasonable design for skill evolution in scientific contexts. The case studies are detailed and show concrete examples of how skills are revised and used.

The prospective HPK1 validation with a synthesized compound showing 14.19 nM IC50 and 2.64-fold improvement over the best reference gives real-world weight to the system's design outputs. That said, the HPK1 study was done without evolved skills, so it does not support claims about skill generalization.

The main soft spot is exactly what the stress-test note flags and what the authors themselves acknowledge in the limitations section (lines b149–b150): the skill evolution is evaluated only retrospectively and in-domain on a single benchmark with a small task set. The improvements from 62.2% to 81.1% on affinity comparison and 79.5% to 87.2% on optimization success are real but could reflect overfitting to quirks of the historical execution data. The paper does not include a held-out task set or cross-benchmark transfer test. The authors call this out honestly, so I'm not going to ding them for something they disclose. However, the abstract and conclusion present the evolution results as support for the general approach without sufficiently highlighting the narrow scope. A careful reader will catch the caveats, but a less careful reader might overinterpret.

Who this is for: researchers building or benchmarking AI agents for drug discovery, especially those interested in skill reuse. The paper is a solid systems contribution with thorough evaluation and honest limitations. It deserves peer review and likely acceptance with a request to make the scope limitations clearer in the abstract.

Recommendation: send to review. This is a well-executed piece of engineering that advances the state of the art in scientific agent systems, even if the skill evolution claim is more modest than the title might suggest.

The strongest objection

The paper's central claim includes conditional skill evolution as a general mechanism for scientific agents, yet the only evidence for this is the retrospective, in-domain evaluation on MolBench (Table 3 and Section 4.3). The skills were learned from historical runs of the same task set and then evaluated on that same set, with no held-out tasks or cross-benchmark transfer test. The paper explicitly acknowledges this limitation (lines [b149]–[b150]), but the abstract and conclusion present the evolution result as support for the general approach. Because the task set is fixed and small (190 tasks, many of which share surface-level patterns), the observed improvements—screening accuracy from 96% to 100%, affinity accuracy from 62.2% to 81.1%, optimization success from 79.5% to 87.2%—could reflect overfitting to the particular tasks or to quirks of the historical execution data rather than learning transferable scientific strategies. The HPK1 prospective study (Section 4.4) does not test evolved skills; it uses the base system without evolution, so it cannot confirm cross-task generalization. Without a held-out or out-of-domain evaluation, the claim that DrugHarness supports conditional skill evolution as a reusable mechanism remains only weakly evidenced.

The reader's weakest assumption focused on LLM reliability in skill induction and review. While that is a valid concern, the paper's own limitation statements and the net positive benchmark results partially mitigate it: the system works despite imperfect LLM components. I identify a more targeted concern: the retrospective, in-domain evaluation design means the evolution results may not generalize. This is a load-bearing issue for the central claim that 'conditional skill evolution' is a broadly applicable mechanism. However, the paper is transparent about this limitation (Section 5), and the overall contribution of the system architecture, benchmark results without evolution, and the prospective HPK1 experimental validation remain strong. Therefore the reader's verdict of ACCEPT with high confidence and medium correctness risk is appropriate, and I recommend no change.

The referee's report

If the results hold, DrugHarness moves the state of the art for agent-based drug discovery in several respects. The environment engineering (80+ operations, shared candidate representation, traceable tool-to-code handoffs) is thorough and will be a useful infrastructure contribution. The prospective HPK1 study with experimental validation goes beyond typical benchmark-only evaluations and provides a concrete demonstration of the workflow. The skill evolution mechanism, while incompletely tested, is well motivated and the design choices (conditional applicability, independent review via the same tools, feedback revision) are sensible. The paper's significance is therefore moderate to high for the execution and prospective components, but lower for the skill-learning component because the evidence for generalization is missing.

Pith's reading of the paper (a machine's summary, not the authors')

DrugHarness is a harness that lets computational drug discovery agents improve by learning conditional skills from their own execution traces. It integrates over 80 scientific operations and preserves associations among candidates, results, and artifacts. After each task, a learner module compares trajectories and outcome feedback to induce action hypotheses with explicit applicability conditions; an independent reviewer checks those hypotheses using the same scientific tools and admits reliable ones into a skill library. Without evolution, DrugHarness exceeded reported MolClaw-CC results by 10.8 percentage points in affinity-comparison accuracy and 55% in mean top-three active hits on MolBench. With evolution, retrospective evaluation showed that learned skills raised Qwen's affinity comparison accuracy from 62.2% to 81.1% and optimization success from 79.5% to 87.2%, and a prospective HPK1 design produced a synthesized compound with a cellular IC50 of 14.19 nM, a 2.64-fold improvement over the best input reference and competitive with expert-designed compounds.

What carries the argument. The key mechanism is the conditional skill learning loop: (1) a learner that reviews execution experiences to form hypotheses of the form (goal, conditions, recommended action, supporting sources, unresolved questions); (2) an independent review agent that checks sources and conducts targeted execution trials using the same scientific environment; (3) a versioned skill library that guides future executors and is revised through subsequent feedback. This loop turns heterogeneous scientific decisions into reusable, evidence-grounded guidance with explicit scope conditions.

Core claim. The central claim is that a computational drug discovery agent can reliably improve its performance by inducing conditional action hypotheses from its own task traces, having those hypotheses independently reviewed through source checks and targeted scientific trials, and admitting the validated skills into a persistent library that guides future tasks. The claim is supported by benchmark comparisons, retrospective skill-evolution experiments, and a prospective molecular design study that yielded a synthesized compound with cellular potency exceeding all input references and matching compounds designed by experienced medicinal chemists.

The premise it rests on. The paper assumes that the large language model driving both skill induction and independent review can reliably identify correct action hypotheses and catch errors, an assumption that is not formally verified and could fail on complex or ambiguous scientific tasks.

If the paper is right (Pith's inferences, not the authors')

DrugHarness's conditional skill learning could be applied to other scientific domains where agents compose tools (e.g., materials science, protein engineering), reducing the need for hand-crafted workflows.

The independent review mechanism provides a path toward more reliable self-improving agents by catching flawed hypotheses before they guide decisions.

The demonstrated link between computational design and experimental validation in HPK1 suggests that the system can generate candidate molecules with practical therapeutic potential.

The retrospective improvement numbers indicate that even a modest number of tasks can yield transferable skills, suggesting that further scaling could produce broadly competent scientific agents.

If the skill induction and review processes are robust, the method could accelerate drug discovery cycles by iteratively capturing and reusing effective strategies.

The paper's reliance on a single benchmark (MolBench) for skill evolution evaluation means the reported gains may not generalize to other target families or task types until tested in held-out settings—a limitation the authors acknowledge.

Both skill induction and review depend on the fidelity of large language models; if those models produce unreliable hypotheses or miss critical errors, the learning loop could degrade performance, a risk not formally addressed.

The prospective HPK1 study involves only one synthesized compound; broader experimental validation with multiple compounds and full developability profiling (ADMET, selectivity) would be needed to establish practical utility.

The method assumes that the executor can assess skill applicability conditions from its visible state; in tasks where critical conditions are unobservable, the guidance may be misapplied without detection.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Computational drug discovery requires agents to compose scientific operations and adapt to changing evidence and constraints. Here we introduce DrugHarness, an evolvable agent harness coupling composable scientific execution with conditional skill induction. Its environment integrates over 80 scientific operations, biomedical resources and task-specific code, preserving associations among candidates, results and artifacts. From execution trajectories and outcome feedback, DrugHarness induces action hypotheses with observable applicability conditions, recommended actions and supporting evidence. An independent review agent evaluates candidate skills derived from these hypotheses through source verification and, when needed, targeted execution trials. Admitted skills guide subsequent tasks and are revised through feedback. Evaluation on MolBench and SMDD-Bench demonstrates capabilities in molecular screening, optimization and constrained design. Without evolution, DrugHarness exceeds reported MolClaw-CC results by 10.8 percentage points in affinity-comparison accuracy and 55% in mean top-three active hits on MolBench. Retrospective evaluation on skill-development tasks shows that evolution increases Qwen's affinity-comparison accuracy from 62.2% to 81.1% and optimization success from 79.5% to 87.2%. In a prospective HPK1 study, DrugHarness achieves a more favorable median docking score than Biomni among system-specific candidates. A synthesized compound achieves a cellular IC50 of 14.19 nM, a 2.64-fold potency improvement over the best input reference, with activity competitive with two selected compounds designed by experienced medicinal chemists. These results support conditional skill evolution for scientific agents and demonstrate a workflow connecting computational drug design with experimental validation.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-20

The authors' main claim, as the classifier picked it out 1.00

“These results support conditional skill evolution for scientific agents and demonstrate a workflow connecting computational drug design with experimental validation.”

How the main claim is stated: asserted 0.90

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.14 · human participants 0.09 · controlled comparison 0.86 · randomized assignment 0.01 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.05

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Computational drug discovery requires agents to compose scientific operations and adapt to changing evidence and constraints. background 1.00 0.05 0.03 asserted 1.00
Here we introduce DrugHarness, an evolvable agent harness coupling composable scientific execution with conditional skill induction. contribution 1.00 0.90 0.07 asserted 1.00
Its environment integrates over 80 scientific operations, biomedical resources and task-specific code, preserving associations among candidates, results and artifacts. method 0.86 0.84 0.03 asserted 1.00
From execution trajectories and outcome feedback, DrugHarness induces action hypotheses with observable applicability conditions, recommended actions and supporting evidence. method 0.96 0.88 0.05 asserted 1.00
An independent review agent evaluates candidate skills derived from these hypotheses through source verification and, when needed, targeted execution trials. method 0.99 0.81 0.03 asserted 1.00
Admitted skills guide subsequent tasks and are revised through feedback. method 0.99 0.54 0.06 asserted 1.00
Evaluation on MolBench and SMDD-Bench demonstrates capabilities in molecular screening, optimization and constrained design. result 0.88 0.93 0.10 asserted 1.00
Without evolution, DrugHarness exceeds reported MolClaw-CC results by 10.8 percentage points in affinity-comparison accuracy and 55% in mean top-three active hits on MolBench. result 1.00 0.93 0.09 asserted 1.00
Retrospective evaluation on skill-development tasks shows that evolution increases Qwen's affinity-comparison accuracy from 62.2% to 81.1% and optimization success from 79.5% to 87.2%. result 1.00 0.94 0.11 asserted 1.00
In a prospective HPK1 study, DrugHarness achieves a more favorable median docking score than Biomni among system-specific candidates. result 1.00 0.95 0.04 asserted 1.00
A synthesized compound achieves a cellular IC50 of 14.19 nM, a 2.64-fold potency improvement over the best input reference, with activity competitive with two selected compounds designed by experienced medicinal chemists. result 1.00 0.90 0.06 asserted 1.00
These results support conditional skill evolution for scientific agents and demonstrate a workflow connecting computational drug design with experimental validation. interpretation 1.00 0.75 0.87 asserted 0.90

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from simulation or computation 0.61

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-20: 1 request, 8,892 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (11)
Figure 2 of the paper
Figure 2. Computational evaluation of HPK1 candidates. Four system-specific structures per group remain after excluding six shared structures from ten evaluated entries each. Colors: DrugHarness; Biomni. Symbols: molecules (•), medians ( ). Lower is favorable. The four inhibition endpoints were selected post hoc for lower DrugHarness medians; its median hERG inhibition was higher (0.949 versus 0.663) and QED lower (0.3645 versus 0.4085). In Silico Evaluation of Designed Molecules. Among the four system-specific candidates per evaluated set, DrugHarness achieves a more favorable median MOE…
Figure 3 of the paper
Figure 3. Comparison with selected chemist de￾signs. Lower cellular IC50 and higher 1 nM HPK1 inhibition indicate stronger activity. 5 LIMITATIONS AND CONCLUSION Limitations. Systematic comparisons of runtime, computational cost, and resource requirements against other agent frameworks and human experts remain open. Skill evolution is evaluated on a single benchmark with a limited number of tasks under a retrospective, in-domain setting; its effectiveness at larger scales and generalization to unseen targets and task families remain to be es￾tablished. Experimental validation is also limited…
Figure 4 of the paper
Figure 4. Reference and generated BACE1 candidates. B2-09 and B3-02 were both generated from the reference molecule. Structures are aligned by their shared core
Figure 5 of the paper
Figure 5. Candidate-linked assessment and submission. The three assessments refer to the same registered B3-02 molecule. Following assessment, the candidate was written to solution.smi, read back and submit￾ted in the evaluated molecular state. The shared candidate representation links each assessment to the delivered molecular structure, al￾lowing results from different scientific tools to be considered together under the task requirements. Other candidates Reference B2-09 B3-02 Task limit 0.4 0.6 0.8 hERG prediction 0.2 0.4 0.6 AMES prediction (a) Two-target view of the evaluated pool 0.70…
Figure 6 of the paper
Figure 6. Predicted property trade-offs among BACE1 candidates. (a) All 32 evaluated structures, including the reference; shading marks the region meeting the hERG and AMES limits. (b) Four predictions divided by their task upper limits. Connected pairs compare B2-09 and B3-02 at each endpoint; ratios at or below one meet that target. 23
Figure 7 of the paper
Figure 7. Scaffold replacement and structural inspection. (a) The reference ligand and submitted candidate have different ring scaffolds and share a sulfonamide group. (b) The predicted Boltz-2 complex was retrieved and used as the code’s sole CIF input. For each selected residue r, the script computed the minimum distance to an atom of the predicted ligand L. The full 21-residue report was read before the same candidate was submitted. Arrows follow the recorded scientific handoffs. 24
Figure 8 of the paper
Figure 8. Local static geometry of the predicted complex. Panel (b) shows the nearest-ligand distances recorded during task execution. The pocket overlay in (a) uses structures aligned after execution for visualiza￾tion. Orange points identify His94 and Thr198; the remaining selected residues are blue. DrugHarness carried the predicted complex from a specialized tool into scientific code and re￾turned the geometric report to the executor before submission. Links between the candidate record, predicted complex, code input and derived report keep this sequence traceable to the submitted…
Figure 9 of the paper
Figure 9. QED feedback reused during molecular optimization. (a) The source, first- and second-batch leaders, and submitted candidate, drawn from the recorded molecular structures. Arrows indicate reuse of evaluated structures as baselines in subsequent scripts. (b) All 80 numerical entries from batches containing 26, 26, and 31 records, including repeated structures. Orange, light blue, and dark blue indicate batches 1–3. Diamonds mark batch maxima, the star marks the submitted candidate, and the dashed line denotes the source QED. Three invalid first-batch entries occupy a separate…
Figure 10 of the paper
Figure 10. From scientific feedback to revised affinity guidance. Independent review results and researcher￾selected execution feedback informed the S0-to-S1 revision. The added recovery-and-matching procedure was exercised in a second review on 003 and 021 before its use on task 012. 27
Figure 11 of the paper
Figure 11. CXCR2 candidates and matched affinity records. Molecular structures are task inputs; Ki values were retrieved from BindingDB records associated with Karlstrom et al. (2013). ¨

Showing 10 of the paper's 11 figures. Figures are the authors', reproduced from the paper.

References (66)
  1. [1] Bran, Andres M. and Cox, Sam and Schilter, Oliver and Baldassari, Carlo and White, Andrew D. and Schwaller, Philippe , year =. Nature Machine Intelligence , volume =. doi:10.1038/s42256-024-00832-8 , url = on Pithdoi:10.1038/s42256-024-00832-8
  2. [2] Huang, Kexin and Zhang, Serena and Wang, Hanchen and Qu, Yuanhao and Lu, Yingzhou and Li, Ryan and Roohani, Yusuf and Qiu, Lin and Cao, Shiyi and Li, Gavin and Zhang, Junze and Yin, Di and Wierenga, Rick and Kavi, Deniz and Liu, Sherry and She, Tianwei and Marwaha, Shruti and ... on Pith
  3. [3] CAi Copilot: Reducing Operational Workload in Molecular Design through Intent-Driven Agentic Workflows , author=. arXiv preprint arXiv:2608.06961 , year= on PithPith's reviewarXiv
  4. [4] Han, Kevin and Zhang, Renfei and Wei, Kathy and Mahdavi, Hamed and Mireshghallah, Niloofar and Farimani, Amir Barati , journal = on Pith
  5. [5] Zhang, Lisheng and Wang, Lilong and Sun, Xiangyu and Tang, Wei and Su, Haoyang and Qian, Yuehui and Yang, Qikui and Li, Qingsong and Tang, Zhenyu and Sun, Haoran and others , journal =. 2026 , doi = on Pith
  6. [6] Averly, Reza and Baker, Frazier N and Watson, Ian A and Ning, Xia , booktitle = on Pith
  7. [7] Buniello, Annalisa and Suveges, Daniel and Cruz-Castillo, Carlos and Llinares, Manuel Bernal and Cornu, Helena and Lopez, Irene and Tsukanov, Kirill and Rold. 2025 , journal =. doi:10.1093/nar/gkae1128 , url = on Pithdoi:10.1093/nar/gkae1128
  8. [8] Cerami, Ethan and Gao, Jianjiong and Dogrusoz, Ugur and Gross, Benjamin E. and Sumer, Selcuk Onur and Aksoy, B. 2012 , journal =. doi:10.1158/2159-8290.CD-12-0095 , url = on Pithdoi:10.1158/2159-8290.cd-12-0095
  9. [9] Gao, Jianjiong and Aksoy, B. 2013 , journal =. doi:10.1126/scisignal.2004088 , url = on Pithdoi:10.1126/scisignal.2004088
  10. [10] Zhou, Ying and Zhang, Yintao and Zhao, Donghai and Yu, Xinyuan and Shen, Xinyi and Zhou, Yuan and Wang, Shanshan and Qiu, Yunqing and Chen, Yuzong and Zhu, Feng , year =. Nucleic Acids Research , volume =. doi:10.1093/nar/gkad751 , url = on Pithdoi:10.1093/nar/gkad751
  11. [11] Landrum, Melissa J and Chitipiralla, Shanmuga and Kaur, Kuljeet and Brown, Garth and Chen, Chao and Hart, Jennifer and Hoffman, Douglas and Jang, Wonhee and Liu, Chunlei and Maddipatla, Zenith and others , year =. Nucleic Acids Research , volume =. doi:10.1093/nar/gkae1090 , url = on Pithdoi:10.1093/nar/gkae1090
  12. [12] 2020 , journal =. doi:10.1126/science.aaz1776 , url = on Pithdoi:10.1126/science.aaz1776
  13. [13] 2020 , journal =. doi:10.1038/s41586-020-2493-4 , url = on Pithdoi:10.1038/s41586-020-2493-4
  14. [14] Szklarczyk, Damian and Kirsch, Rebecca and Koutrouli, Mikaela and Nastou, Katerina and Mehryary, Farrokh and Hachilif, Radja and Gable, Annika L and Fang, Tao and Doncheva, Nadezhda T and Pyysalo, Sampo and others , year =. Nucleic Acids Research , volume =. doi:10.1093/nar/gk... on Pithdoi:10.1093/nar/gkac1000
  15. [15] Rosonovski, Summer and Levchenko, Maria and Bhatnagar, Rajat and Chandrasekaran, Umamageswari and Faulk, Lynne and Hassan, Islam and Jeffryes, Matt and Mubashar, Syed Irtaza and Nassar, Maaly and Jayaprabha Palanisamy, Madhumiethaa and others , year =. Nucleic Acids Research ,... on Pithdoi:10.1093/nar/gkad1085
  16. [16] 2025 , journal =. doi:10.1093/nar/gkae1010 , url = on Pithdoi:10.1093/nar/gkae1010
  17. [17] Burley, Stephen K. and Bhikadiya, Charmi and Bi, Chunxiao and Bittrich, Sebastian and Chao, Henry and Chen, Li and Craig, Paul A. and Crichlow, Gregg V. and Dalenberg, Kenneth and Duarte, Jose M. and others , year =. Protein Science , volume =. doi:10.1002/pro.4482 , url = on Pithdoi:10.1002/pro.4482
  18. [18] Velankar, Sameer and Dana, Jos. 2013 , journal =. doi:10.1093/nar/gks1258 , url = on Pithdoi:10.1093/nar/gks1258
  19. [19] Varadi, Mihaly and Bertoni, Damian and Magana, Paulyna and Paramval, Urmila and Pidruchna, Ivanna and Radhakrishnan, Malarvizhi and Tsenkov, Maxim and Nair, Sreenath and Mirdita, Milot and Yeo, Jingi and others , year =. Nucleic Acids Research , volume =. doi:10.1093/nar/gkad1... on Pithdoi:10.1093/nar/gkad1011
  20. [20] Lawson, Catherine L and Berman, Helen M and Chen, Li and Vallat, Brinda and Zirbel, Craig L , year =. Nucleic Acids Research , volume =. doi:10.1093/nar/gkad957 , url = on Pithdoi:10.1093/nar/gkad957
show 46 more
  1. [21] Santos-Martins, Diogo and He, Yiran and Eberhardt, Jerome and Sharma, Parnika and Bruciaferri, Niccol. 2025 , journal =. doi:10.1021/acs.jcim.5c02271 , url = on Pithdoi:10.1021/acs.jcim.5c02271
  2. [22] Westbrook, John D. and Shao, Chenghua and Feng, Zukang and Zhuravleva, Marina and Velankar, Sameer and Young, Jasmine , year =. Bioinformatics , volume =. doi:10.1093/bioinformatics/btu789 , url = on Pithdoi:10.1093/bioinformatics/btu789
  3. [23] Le Guilloux, Vincent and Schmidtke, Peter and Tuffery, Pierre , year =. BMC Bioinformatics , volume =. doi:10.1186/1471-2105-10-168 , url = on Pithdoi:10.1186/1471-2105-10-168
  4. [24] Kriv. 2018 , journal =. doi:10.1186/s13321-018-0285-8 , url = on Pithdoi:10.1186/s13321-018-0285-8
  5. [25] Schake, Philipp and Bolz, Sarah Naomi and Linnemann, Katja and Schroeder, Michael , year =. Nucleic Acids Research , volume =. doi:10.1093/nar/gkaf361 , url = on Pithdoi:10.1093/nar/gkaf361
  6. [26] Bouysset, C. 2021 , journal =. doi:10.1186/s13321-021-00548-6 , url = on Pithdoi:10.1186/s13321-021-00548-6
  7. [27] Zdrazil, Barbara and Felix, Eloy and Hunter, Fiona and Manners, Emma J and Blackshaw, James and Corbett, Sybilla and de Veij, Marleen and Ioannidis, Harris and Lopez, David Mendez and Mosquera, Juan F and others , year =. Nucleic Acids Research , volume =. doi:10.1093/nar/gkad... on Pithdoi:10.1093/nar/gkad1004
  8. [28] Liu, Tiqing and Hwang, Linda and Burley, Stephen K and Nitsche, Carmen I and Southan, Christopher and Walters, W Patrick and Gilson, Michael K , year =. Nucleic Acids Research , volume =. doi:10.1093/nar/gkae1075 , url = on Pithdoi:10.1093/nar/gkae1075
  9. [29] Kim, Sunghwan and Chen, Jie and Cheng, Tiejun and Gindulyte, Asta and He, Jia and He, Siqian and Li, Qingliang and Shoemaker, Benjamin A and Thiessen, Paul A and Yu, Bo and others , year =. Nucleic Acids Research , volume =. doi:10.1093/nar/gkae1059 , url = on Pithdoi:10.1093/nar/gkae1059
  10. [30] Schneuing, Arne and Harris, Charles and Du, Yuanqi and Didi, Kieran and Jamasb, Arian and Igashov, Ilia and Du, Weitao and Gomes, Carla and Blundell, Tom L. and Lio, Pietro and others , year =. Nature Computational Science , volume =. doi:10.1038/s43588-024-00737-x , url = on Pithdoi:10.1038/s43588-024-00737-x
  11. [31] Dunn, Ian and Koes, David R. , year =. Digital Discovery , volume =. doi:10.1039/D5DD00363F , url = on Pithdoi:10.1039/d5dd00363f
  12. [32] Ghorbani, Mahdi and Gendelev, Leo and Beroza, Paul and Keiser, Michael J. , year =. arXiv , doi =. 2401.05370 , archivePrefix = on PithPith's reviewarXiv
  13. [33] Trott, Oleg and Olson, Arthur J. , year =. Journal of Computational Chemistry , volume =. doi:10.1002/jcc.21334 , url = on Pithdoi:10.1002/jcc.21334
  14. [34] Eberhardt, Jerome and Santos-Martins, Diogo and Tillack, Andreas F. and Forli, Stefano , year =. Journal of Chemical Information and Modeling , volume =. doi:10.1021/acs.jcim.1c00203 , url = on Pithdoi:10.1021/acs.jcim.1c00203
  15. [35] Tang, Shidi and Ding, Ji and Zhu, Xiangyu and Wang, Zheng and Zhao, Haitao and Wu, Jiansheng , year =. IEEE/ACM Transactions on Computational Biology and Bioinformatics , volume =. doi:10.1109/TCBB.2024.3467127 , url = on Pithdoi:10.1109/tcbb.2024.3467127
  16. [36] Ding, Ji and Tang, Shidi and Mei, Zheming and Wang, Lingyue and Huang, Qinqin and Hu, Haifeng and Ling, Ming and Wu, Jiansheng , year =. Journal of Chemical Information and Modeling , volume =. doi:10.1021/acs.jcim.2c01504 , url = on Pithdoi:10.1021/acs.jcim.2c01504
  17. [37] Cao, Duanhua and Chen, Geng and Jiang, Jiaxin and Yu, Jie and Zhang, Runze and Chen, Mingan and Zhang, Wei and Chen, Lifan and Zhong, Feisheng and Zhang, Yingying and others , year =. Nature Machine Intelligence , volume =. doi:10.1038/s42256-024-00849-z , url = on Pithdoi:10.1038/s42256-024-00849-z
  18. [38] Passaro, Saro and Corso, Gabriele and Wohlwend, Jeremy and Reveiz, Mateo and Thaler, Stephan and Somnath, Vignesh Ram and Getz, Noah and Portnoi, Tally and Roy, Julien and Stark, Hannes and others , year =. bioRxiv , doi = on Pith
  19. [39] Wohlwend, Jeremy and Corso, Gabriele and Passaro, Saro and Getz, Noah and Reveiz, Mateo and Leidal, Ken and Swiderski, Wojtek and Atkinson, Liam and Portnoi, Tally and Chinn, Itamar and others , year =. bioRxiv , doi = on Pith
  20. [40] Mirdita, Milot and Sch. 2022 , journal =. doi:10.1038/s41592-022-01488-1 , url = on Pithdoi:10.1038/s41592-022-01488-1
  21. [41] Delaney, John S. , year =. Journal of Chemical Information and Computer Sciences , volume =. doi:10.1021/ci034243x , url = on Pithdoi:10.1021/ci034243x
  22. [42] Li, Hao and Cao, He and Feng, Bin and Shao, Yanjun and Tang, Xiangru and Yan, Zhiyuan and Tian, Yonghong and Yuan, Li and Li, Yu , booktitle =. Beyond Chemical. doi:10.52202/085713-5122 , editor = on Pithdoi:10.52202/085713-5122
  23. [43] Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development , author =. Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks , year = on Pith
  24. [44] Ertl, Peter and Schuffenhauer, Ansgar , year =. Journal of Cheminformatics , volume =. doi:10.1186/1758-2946-1-8 , url = on Pithdoi:10.1186/1758-2946-1-8
  25. [45] Swanson, Kyle and Walther, Parker and Leitz, Jeremy and Mukherjee, Souhrid and Wu, Joseph C and Shivnaraine, Rabindra V and Zou, James , year =. Bioinformatics , volume =. doi:10.1093/bioinformatics/btae416 , url = on Pithdoi:10.1093/bioinformatics/btae416
  26. [46] Loeffler, Hannes H. and He, Jiazhen and Tibo, Alessandro and Janet, Jon Paul and Voronov, Alexey and Mervin, Lewis H. and Engkvist, Ola , year =. Journal of Cheminformatics , volume =. doi:10.1186/s13321-024-00812-5 , url = on Pithdoi:10.1186/s13321-024-00812-5
  27. [47] He, Jiazhen and Nittinger, Eva and Tyrchan, Christian and Czechtizky, Werngard and Patronov, Atanas and Bjerrum, Esben Jannik and Engkvist, Ola , year =. Journal of Cheminformatics , volume =. doi:10.1186/s13321-022-00599-3 , url = on Pithdoi:10.1186/s13321-022-00599-3
  28. [48] Fialkov. 2022 , journal =. doi:10.1021/acs.jcim.1c00469 , url = on Pithdoi:10.1021/acs.jcim.1c00469
  29. [49] Saigiridharan, Lakshidaa and Hassen, Alan Kai and Lai, Helen and Torren-Peraire, Paula and Engkvist, Ola and Genheden, Samuel , year =. Journal of Cheminformatics , volume =. doi:10.1186/s13321-024-00860-x , url = on Pithdoi:10.1186/s13321-024-00860-x
  30. [50] Zhao, Andrew and Huang, Daniel and Xu, Quentin and Lin, Matthieu and Liu, Yong-Jin and Huang, Gao , booktitle = on Pith
  31. [51] Jingwei Ni and Yihao Liu and Xinpeng Liu and Yutao Sun and Mengyu Zhou and Pengyu Cheng and Dexin Wang and Erchao Zhao and Xiaoxi Jiang and Guanjun Jiang , year =. 2603.25158 , archivePrefix = on PithPith's reviewarXiv
  32. [52] Yang, Yifan and Gong, Ziyang and Huang, Weiquan and Yang, Qihao and Zhou, Ziwei and Huang, Zisu and Li, Yan and Gao, Xuemei and Dai, Qi and Liu, Bei and Qiu, Kai and Yang, Yuqing and Chen, Dongdong and Yang, Xue and Luo, Chong , year =. 2605.23904 , archivePrefix = on PithPith's reviewarXiv
  33. [53] Tang, Liyan and Rashtchian, Cyrus and Ferng, Chun-Sung and Tomkins, Andrew and Juan, Da-Cheng and Vu, Tu , year =. 2608.27454 , archivePrefix = on PithPith's reviewarXiv
  34. [54] Yikun Zhang and Xiwei Cheng and Tianyu Liu and Yuanqi Du and Wengong Jin , year =. 2605.15461 , archivePrefix = on PithPith's reviewarXiv
  35. [55] Zhang, Qizheng and Hu, Changran and Upasani, Shubhangi and Ma, Boyuan and Hong, Fenglu and Kamanuru, Vamsidhar and Rainton, Jay and Wu, Chen and Ji, Mengmeng and Li, Hanchen and Thakker, Urmish and Zou, James Y and Olukotun, Kunle , booktitle =. Agentic Context Engineering: Ev... on Pith
  36. [56] Gao, Shanghua and Zhu, Richard and Sui, Pengwei and Kong, Zhenglun and Aldogom, Sufian and Huang, Yepeng and Noori, Ayush and Shamji, Reza and Parvataneni, Krishna and Tsiligkaridis, Theodoros and Zitnik, Marinka , year =. 2509.23426v3 , archivePrefix = on PithPith's reviewarXiv
  37. [57] Ding, Keyan and Yu, Jing and Huang, Junjie and Yang, Yuchen and Zhang, Qiang and Chen, Huajun , year =. Nature Computational Science , volume =. doi:10.1038/s43588-025-00849-y , url = on Pithdoi:10.1038/s43588-025-00849-y
  38. [58] Gao, Bowen and Huang, Yanwen and Liu, Yiqiao and Xie, Wenxuan and Ma, Wei-Ying and Zhang, Ya-Qin and Lan, Yanyan , year =. 2503.22164v2 , archivePrefix = on PithPith's reviewarXiv
  39. [59] Li, Lingxiao and Zhang, Haobo and Fan, Ruohao and Chen, Bin and Zhou, Jiayu , year =. Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 , pages =. doi:10.1145/3770855.3819017 , url = on Pithdoi:10.1145/3770855.3819017
  40. [60] Cao, He and Liu, Siyu and Zhang, Fan and Liu, Zijing and Li, Hao and Feng, Bin and Bai, Shengyuan and Chen, Leqing and Xie, Kai and Li, Yu , year =. 2603.03655v1 , archivePrefix = on PitharXiv
  41. [61] Tang, Xiangru and Hu, Tianyu and Ye, Muyang and Shao, Yanjun and Yin, Xunjian and Ouyang, Siru and Zhou, Wangchunshu and Lu, Pan and Zhang, Zhuosheng and Zhao, Yilun and Cohan, Arman and Gerstein, Mark , booktitle = on Pith
  42. [62] Tang, Yuqi and Zhou, Chenyi and Wang, Libin and Ding, Keyan and Zhang, Qiang and Chen, Huajun , year =. 2607.28692v1 , archivePrefix = on PithPith's reviewarXiv
  43. [63] Fu, Yao and Kim, Dong-Ki and Kim, Jaekyeom and Sohn, Sungryull and Logeswaran, Lajanugen and Bae, Kyunghoon and Lee, Honglak , booktitle =. doi:10.52202/079017-3811 , editor = on Pithdoi:10.52202/079017-3811
  44. [64] He, Yu and Yang, Weikai , year =. 2608.15165v1 , archivePrefix = on PithPith's reviewarXiv
  45. [65] Zhang, Lisheng and Wang, Lilong and Sun, Xiangyu and Tang, Wei and Su, Haoyang and Qian, Yuehui and Yang, Qikui and Li, Qingsong and Tang, Zhenyu and Sun, Haoran and Han, Yingnan and Jiang, Yankai and Lou, Wenjie and Zhou, Bowen and Wang, Xiaosong and Bai, Lei and Xie, Zhengwe... on Pith
  46. [66] Karlstr. Substituted 7-amino-5-thio-thiazolo[4,5-. Journal of Medicinal Chemistry , volume =. 2013 , doi = on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment