PaperlayerView exact revision on arXiv

Source-linked AI summary

PMO-Dock: Benchmarking Docking, Specificity, and Generalization in Molecular Optimization

Gor Simonyan, Tatevik Abrahamyan, Narek Abelyan, Tigran Fahradyan, Hrant Khachatrian

arXiv:2610.19011v1cs.LGq-bio.QM

TL;DR

Molecular optimization evaluation is fragmented, and property-based objectives do not capture docking against protein targets. PMO-Dock introduces a held-out, docking-based benchmark spanning hit generation, lead optimization, and binding specificity. Across tasks, no method dominates all others, and methods suit distinct drug-design workflows.

  • Problem

    Molecular optimization evaluation is fragmented, while PMO's simple property-based objectives do not cover docking against protein targets.

    Source: §1 · pp. 1–2

  • Method

    PMO-Dock benchmarks docking-based optimization across 25 tasks, using validation tasks for development and held-out test tasks for final evaluation.

    Source: §1 · p. 2

  • Results

    No generalist methods dominate all tasks, and methods are better suited to distinct drug-design workflows.

    Source: §1 · p. 3

  • Takeaways & Limitations

    PMO-Dock provides a standard for evaluating generalization to unseen docking tasks under a limited test-time oracle budget.

    Source: §6 · p. 10

    The benchmark uses QuickVina and focuses on docking objectives, leaving advanced simulations and requirements such as solubility and non-toxicity for future work.

    Source: A Limitations · p. 13

Abstract

from arXiv · show

The Practical Molecular Optimization (PMO) benchmark standardized evaluation in molecular optimization, but it is built on simple property-based oracles that do not capture structure-based drug design. As the field has moved to docking-based objectives, evaluation practice has become unstandardized, and different methods report results on different docking tasks, often without a shared protocol for fair comparison or sample-efficiency constraints. To address this, we introduce Practical Molecular Optimization for Docking (PMO-Dock), a benchmark and protocol for docking-based optimization consisting of 25 tasks covering hit generation, lead optimization, and a new specificity task requiring strong on-target binding while penalizing off-target interactions. The protocol separates model development from final evaluation: validation tasks are used for hyperparameter tuning, and test tasks are held out for final evaluation and are not used during development. Evaluation on the test tasks is done under a strict oracle budget to reflect realistic drug-discovery settings. We benchmark four diverse high-performing methods spanning different optimization paradigms, Saturn (reinforcement learning), GenMol (discrete diffusion), Genetic-guided GFlowNet, and ChemLlama (LLM-based evolutionary search). Our results show no universally best method across tasks, additional analyses of hyperparameter transfer (cross-task and cross-category) and sensitivity show that transfer is often limited and varies by method, motivating method-dependent selection strategies. This benchmark provides a concrete standard for evaluating generalization to unseen tasks and sample efficiency under a limited test-time oracle budget.

1 Introduction and Related Work

Molecular optimization evaluation has been fragmented, and PMO’s property-based objectives do not cover docking against protein targets. PMO-Dock pairs docking-based tasks with PMO-style evaluation standards and introduces binding specificity tasks.

1 Introduction and Related Work

  • PMO standardized evaluation across 23 optimization tasks, but its simple RDKit-based objectives do not cover docking against protein targets.These objectives include molecular similarity, isomer counts, and drug-likeness metrics.Source: §1 · pp. 1–2
  • Prior docking benchmarks lacked a held-out task protocol for testing generalization to unseen proteins or drug-design settings.The SMINA benchmark measured single-target hit optimization using a learned surrogate, while TARTARUS used a live docking oracle during optimization.Source: §1 · p. 2
  • PMO-Dock introduces 25 docking-based tasks, separated into 11 validation tasks for development and 14 held-out test tasks for evaluation.The benchmark covers hit generation, lead optimization, and binding specificity.Source: §1 · p. 2
  • Binding specificity tasks require strong binding to a target protein or pocket while avoiding off-target interactions.Five tasks address both intra-protein and inter-protein selectivity challenges.Source: §1 · pp. 2–3
  • There are no generalist methods that dominate all tasks; methods are better suited to distinct drug-design workflows.The study also analyzes hyperparameter transferability and sensitivity.Source: §1 · p. 3

2 The PMO-Dock Benchmark

PMO-Dock organizes docking-based molecular optimization into hit generation, lead optimization, and binding-specificity tasks, with separate validation and test sets for development and final evaluation. Its task designs measure distinct objectives, including selective binding against off-target pockets.

2 The PMO-Dock Benchmark

  • PMO-Dock measures optimization power, sample efficiency, practicality, and generalization across hit generation, lead optimization, and docking with specificity.The categories use distinct success measures.Source: §2 · p. 3

2.1 Hit Generation

  • Hit generation maximizes the expected rate of unique molecules satisfying physicochemical and target-specific constraints under a 3,000-evaluation budget.The hit oracle uses QuickVina at exhaustiveness 1.Source: §2.1 · p. 3

2.2 Lead Optimization

  • Lead optimization generates compounds that improve a target-specific docking objective while remaining near a seed molecule and satisfying locality and drug-likeness constraints.The task uses a 1,000-evaluation budget and evaluates the top-1 docking score among molecules meeting the constraints.Source: §2.2 · pp. 3–4

2.3 Specificity Binding

  • Specificity tasks optimize binding to TYK2 JH2 while penalizing interactions with TYK2 JH1 or homologous JAK1 and JAK2 domains.The benchmark includes intra-protein domain, inter-protein pseudokinase, and cross-domain inter-protein specificity comparisons.Source: §2.3 · pp. 4–5

2.4 Validation and Test Split

  • The benchmark separates 11 validation tasks for development and hyperparameter selection from 14 held-out test tasks for final evaluation.Test proteins, lead molecules, and binding pockets should not inform model development or training-data choice.Source: §2.4 · p. 5, §2.5 · p. 5

3 Experiments

The experiments compare four molecular optimization methods using category-specific hyperparameter selection and held-out test tasks. No method leads across all categories: Saturn performs better on hit generation, ChemLlama on specificity, and lead-optimization scores are comparable.

3.1 Optimization Methods

  • The experiments evaluate Saturn, GenMol, Genetic-guided GFlowNets, and ChemLlama, which span reinforcement learning, discrete diffusion, GFlowNets, and LLM-based evolutionary optimization.Saturn uses an augmented memory buffer; GenMol remasks and resamples molecular fragments; Genetic-guided GFlowNet uses genetic operators; ChemLlama proposes analogs from a pool of high-scoring molecules.Source: §3.1 · p. 6

3.2 Hyperparameter Tuning

  • Hyperparameters are selected using mean performance on other validation tasks in the same category, then frozen for test evaluation without using test scores.An all-category selection rule and per-task best settings are reported separately as ablations or an extended-budget ceiling.Source: §3.2 · p. 6

3.3 Validation Tasks

  • On validation tasks, Saturn is strongest for hit generation, ChemLlama for specificity, and lead scores are closer across methods.Figure 2 presents performance under category-specific selection.Source: §3.3 · p. 6

3.4 Test Tasks

  • No method is superior across all test tasks: Saturn performs better on hit generation, ChemLlama on specificity, and methods perform comparably on lead optimization.Figure 3 reports mean scores with 95% bootstrap confidence intervals over evaluation seeds.Source: §3.4 · p. 6, Figure 3 · p. 7

3.5 Evaluation with Extended Budget

  • The extended-budget evaluation selects the best validation hyperparameters independently for each task, isolating per-task ceiling behavior without cross-task transfer.Figure 4 reports validation performance under this unrestricted model-selection scenario.Source: §3.5 · p. 7, Figure 4 · p. 7

3.6 Statistical Significance

  • Figure 3 intervals describe seed-level uncertainty for individual method-task scores, while category-level comparisons require resampling tasks and seeds together.Figure 5 estimates pairwise category-level win percentages using 10,000 bootstrap replicates over task-seed pairs.Source: §3.6 · p. 7

4 Hyperparameter Sensitivity and Transferability

The section tests whether validation-tuned hyperparameters transfer across targets and task categories, and how performance changes away from task-optimal settings. Transfer is method- and task-dependent, limiting when validation configurations can be reused.

4 Hyperparameter Sensitivity and Transferability

  • The analysis measures transfer across validation tasks and categories, then assesses sensitivity through landscape patterns and the gap between transfer-based and per-task-best selection.The protocol restricts oracle use on held-out test tasks, motivating whether validation-tuned settings can be reused without another sweep.Source: §4 · p. 8

4.1 Within-category transfer.

  • Saturn aligns across hit-generation targets, but lead-optimization transfer is fragile for all methods; Genetic GFN is least affected by the source task’s hyperparameters.Saturn, ChemLlama, and GenMol degrade more when moved from target-specific optima.Source: §4.1 · p. 8, Figure 6 · p. 9

4.2 Cross-category transfer.

  • Saturn has the strongest hit-to-lead alignment and stays positive across proteins, while GenMol is essentially uncorrelated; cross-category degradation exceeds within-category degradation for Saturn, ChemLlama, and GenMol.Genetic GFN is also positively aligned, but on a smaller hyperparameter grid, and shows little category-boundary effect.Source: §4.2 · p. 9, Figure 7 · p. 9, Figure 8 · p. 9

4.3 Hyperparameter Sensitivity

  • Transfer-based selection can leave a larger gap to per-task-best validation performance, especially for GenMol; Saturn gains most from category-specific selection, while Genetic GFN is essentially indifferent.The per-task-best setting is an oracle-rich ceiling; Genetic GFN’s smaller sweep grid limits how finely its landscape can be resolved.Source: §4.3 · p. 10

5 Do these Methods Hack Rewards?

The authors examine docking-score reliability and generated-molecule quality. Three independent runs at exhaustiveness 8 were more stable than one run at 16, and manual inspection found chemically valid structures across methods.

5.1 The role of docking exhaustiveness

  • Three independent QuickVina runs at exhaustiveness 8 were more stable than one run at 16 and agreed closely with the costlier 3×ex16 scheme, motivating 3×ex8 for specificity docking.Repeated scores can produce false negatives from a single run, particularly for reward-hacking risk in specificity tasks.Source: §5.1 · p. 10

5.2 Manual Inspection

  • Manual inspection found chemically valid, well-formed molecules across methods, without unusual artifacts, unrealistic substructures, or chemically implausible patterns.A ChemLlama specificity example showed favorable target binding and weak anti-target binding.Source: §5.2 · p. 10

6 Conclusion

PMO-Dock evaluates generalization to unseen molecular-optimization tasks under strict sample-efficiency constraints. Its results reveal method-specific strengths and weaknesses that inform applicability across drug-design workflows.

6 Conclusion

  • PMO-Dock comprises 25 problems spanning hit generation, lead optimization, and specificity binding, with 11 tasks for development and tuning and 14 held out for final evaluation.Source: §6 · p. 10
  • The benchmark evaluates four well-performing optimization algorithms and reports distinct strengths and weaknesses relevant to their applicability in different drug-design workflows.Source: §6 · p. 10
  • The authors present PMO-Dock as a concrete protocol for supporting and measuring algorithmic advancements in docking-based molecular optimization.Source: §6 · p. 10

A Limitations

The benchmark standardizes comparison of docking-based molecular-optimization tasks, but the authors identify limits in its oracle and objective scope.

A Limitations

  • The study uses QuickVina for binding-affinity and docking scores; more advanced simulations may mitigate potential QuickVina hacking and make results more relevant to real-world drug discovery.Source: A Limitations · p. 13
  • The benchmark focuses mainly on docking objectives and does not yet incorporate other drug-development requirements such as solubility and non-toxicity.Source: A Limitations · p. 13

B Detailed Method Descriptions and Hyperparameter Configurations

The appendix describes four optimization methods and their reward or search designs, then examines how well tuned hyperparameters transfer across tasks and how stable test comparisons are under bootstrap resampling.

B.1 Saturn

  • Saturn uses reinforcement learning with a replay buffer of high-reward molecules, optional genetic operators, and task-specific multiplicative rewards that penalize anti-target binding for specificity.The buffer size and reward type were among the tested hyperparameters.Source: §B.1 · p. 13

B.2 GenMol

  • GenMol performs fragment-level discrete diffusion, using iterative unmasking and remasking to generate or modify molecules, with binary rewards for hit, lead, and specificity tasks.The sweep tested 120 combinations of masking, randomness, and softmax temperature.Source: §B.2 · pp. 13–14

B.3 Genetic-guided GFlowNets

  • Genetic-guided GFlowNet combines a pretrained SMILES policy, genetic operators, replay-buffered trajectory-balance updates, and a KL penalty to the pretrained prior.The study fixed β = 50 across categories after representative-task selection, while tuning replay and KL coefficients on validation tasks.Source: §B.3 · p. 14

B.4 ChemLlama

  • ChemLlama uses an LLM to propose analogs from a pool of top molecules, scores candidates, updates the pool, and fine-tunes on molecule–score prompts when improvement stalls.Its reward scores properties against target intervals, including seed similarity for lead optimization and an anti-target penalty for specificity.Source: §B.4 · pp. 14–15

C Additional Hyperparameter Transfer Analyses

D All-but-me hyperparameter selection

  • All-but-me validation selection chooses hyperparameters that maximize average performance across every other validation task, including tasks from other categories.Figure 11 reports the resulting per-task performance and category means; colors are normalized separately by category.Source: D All-but-me hyperparameter selection · p. 15, Figure 11 · p. 17

E Additional Statistical Analyses

E.4 Hit-rate bootstrap on test hit tasks

Bootstrap analyses assess uncertainty in test hit-rate comparisons, while validation-task scatter plots probe transfer across targets and lead-optimization thresholds. Saturn’s hit-task comparisons are consistently positive against competitors, but transfer patterns vary by method and setting.

E.4 Hit-rate bootstrap on test hit tasks

  • For every reported competitor and hit-task summary, Saturn’s paired bootstrap 95% interval for the mean difference lies strictly above zero.The summaries are BRAF, JAK2, and BOTH; the analysis uses paired seed resampling and complements win-percentage results.Source: §E.4 · p. 19, Figure 14 · p. 20

F.1 Cross-Target Hit Generation: Additional Protein Pairs

  • Saturn shows strong positive hit-generation alignment across protein pairs (r = 0.90–0.94, n=16), while GenMol’s correlations remain near zero and other methods vary by pair.ChemLlama is moderate on parp1–fa7 (r=0.61) but weak on 5ht1b–fa7 (r=0.06); Genetic GFN has mixed low-n correlations.Source: §F.1 · p. 20, Figure 15 · p. 20, Figure 16 · p. 21

F.2 Lead Optimization: Threshold Transfer

  • Hyperparameters optimized for lead-optimization similarity threshold δ = 0.4 do not reliably transfer to δ = 0.6, indicating sensitivity to constraint tightness.Pearson correlation is undefined in subsets where one threshold yields constant scores.Source: §F.2 · p. 21, Figure 18 · p. 21

F.3 Specificity Task Transfer

Validation-task transfer analyses show that performance relationships depend on the task pairing and method. Saturn has stronger alignment with held-out 5ht1b hit performance than the other methods, but validation aggregates do not uniformly predict it.

F.3 Specificity Task Transfer

  • Across two validation specificity task pairs, performance correlation varies by method, and strong performance on one pair does not guarantee strong performance on the other.Correlations are more positive for ChemLlama and Saturn, and weaker or more uncertain for GenMol and Genetic GFN.Source: §F.3 · p. 21, Figure 19 · p. 22

F.4 Held-Out Target Generalization

  • For held-out 5ht1b hit performance, Saturn shows substantially stronger positive alignment with validation-side aggregates than the other methods, while aggregate validation performance does not uniformly predict the held-out result.The validation aggregates use fa7 and parp1 hit performance, or combined normalized hit and lead performance.Source: §F.4 · p. 22, Figure 20 · p. 22

G Details of the Docking Pipeline for Specificity Tasks

Specificity docking uses a dedicated, higher-exhaustiveness pipeline and multiple runs. Additional analyses compare compute cost and score behavior at lower exhaustiveness, and assess QuickVina against AutoDock Vina.

G Details of the Docking Pipeline for Specificity Tasks

  • Specificity tasks use QuickVina with exhaustiveness 8 and ligand-centered binding boxes, with three independent runs chosen to balance reliability and computational efficiency.The pipeline was also checked against Deucravacitinib’s experimentally known TYK2 binding preference.Source: G Details of the Docking Pipeline for… · pp. 22–23, §G.1 · p. 23

G.1 Exhaustiveness sensitivity (specificity docking)

  • Three runs at exhaustiveness 8 were more stable than one run at exhaustiveness 16 in small-scale tests, while three runs at exhaustiveness 8 and 16 did not differ substantially.The authors selected three runs at exhaustiveness 8 because higher exhaustiveness greatly increases time and compute.Source: §G.1 · p. 23

H.1 Limited scoring-function precision at exhaustiveness 1

  • At exhaustiveness 1, QuickVina appears to capture the relative ranking of strong candidates, while higher exhaustiveness mostly refines scores.Docking takes about 0.05 s/molecule at exhaustiveness 1 versus about 0.25 s/molecule at exhaustiveness 8 on a single CPU core.Source: §H.1 · p. 24

H.2 Adequacy of QuickVina relative to AutoDock Vina

  • QuickVina and AutoDock Vina scores are well correlated across methods and molecule sets, supporting QuickVina’s use as a faster docking oracle for the benchmark.The reported comparison covers all generated molecules and successful molecules across target proteins.Source: §H.2 · p. 25, Table 4 · p. 25

H.3 Physical plausibility of high-scoring poses

The authors assess whether high-scoring generated molecules have physically plausible docked poses using PoseBusters. Generated molecules pass all checks less often than redocked benchmark ligands under the matched docking and PoseBusters protocol.

H.3 Physical plausibility of high-scoring poses

  • The PoseBusters analysis covered 1,335 highest-scoring molecules across four methods, with regenerated local QuickVina poses evaluated after hydrogen preprocessing.The preprocessing removed and re-added hydrogens to avoid PDBQT conversion artifacts; a molecule passed only if it satisfied every PoseBusters check.Source: §H.3 · p. 25, Table 5 · p. 26
  • 59.2% of hit-evaluation molecules and 56.3% of lead-evaluation molecules passed all checks, compared with 42.9% for specificity-evaluation molecules.Per-method pass rates ranged from 50.8% to 60.6%, with ChemLlama lowest.Source: §H.3 · p. 25, Table 5 · p. 26
  • 55.7% of the highest-scoring generated molecules passed all PoseBusters checks, versus 78.3% of successfully docked PoseBusters benchmark ligands under the matched protocol.The benchmark-ligand rate was 77.3% when all 308 ligands were counted.Source: §H.3 · p. 25, Table 5 · p. 26

I Experimental Methodology and Community Resources

The appendix describes hyperparameter sweeps, task-specific thresholds, compute resources, and community resources; it also points to the main paper for the validation/test split and model-selection protocol.

I Experimental Methodology and Community Resources

  • The appendix provides hyperparameter sweep procedures, task-specific thresholds, and community resources, while the validation/test split and model-selection protocol are described in Section 2.4.The task overview is identified as Figure 1.Source: I Experimental Methodology and… · p. 26

I.1 Hyperparameter Sweep Details

  • Independent validation-task sweeps were run for Saturn, GenMol, Genetic-guided GFlowNet, and ChemLlama; test reporting used category-specific selection and one evaluation per method.All tuning was confined to validation tasks, and transfer matrices and selection rules follow Sections 4 and 3.Source: §I.1 · p. 26

I.2.1 Hit Generation Thresholds

  • Hit-generation tasks use target-specific docking-score thresholds, while the specificity objective uses the TYK2 JH2 QuickVina score of Deucravacitinib as a reference for counting hits.The reference score is defined under the protocol in Appendix G.Source: §I.2.1 · p. 26, §I.2.2 · p. 26

I.3 Compute Resources

  • The reported sweep compute used one DGXH100 box with eight 80 GB NVIDIA H100 GPUs and 128 CPU core-processors for QuickVina simulations.The GPU resources were used for hyperparameter sweeps.Source: §I.3 · p. 26

NeurIPS Paper Checklist

  • The authors state that experimental code is provided in the PMO-Dock GitHub repository and that the paper includes no dataset.The checklist also directs readers to Sections 2 and 4 for experiment settings.Source: §2 · p. 29
Loading 2610.19011v1…