PaperlayerView exact revision on arXiv

Source-linked AI summary

Accelerating high-throughput virtual screening through molecular pool-based active learning

David E. Graff, Eugene I. Shakhnovich, Connor W. Coley

arXiv:2012.07127v1q-bio.QMcs.LG

TL;DR

Exhaustive virtual screening is increasingly costly as libraries grow, making it impractical across many distinct protein targets. This paper applies Bayesian optimization to docking and reports an order-of-magnitude reduction in computational cost.

  • Problem

    As virtual libraries grow, exhaustive screening becomes computationally costly and impractical across many distinct protein targets.

    Source: Introduction · p. 3

  • Method

    The paper applies Bayesian optimization to docking, using a surrogate model trained on acquired data to guide subsequent evaluations.

    Source: Introduction · pp. 3–4

  • Results

    Computational cost decreases by an order of magnitude while the fidelity of structure-based methods is preserved.

    Source: Introduction · p. 4

  • Takeaways & Limitations

    Bayesian optimization offers a way to reduce exhaustive docking costs while preserving the fidelity of structure-based methods.

    Source: Introduction · p. 4

    Gaussian-process training does not scale well to millions of acquired data points.

    Source: Introduction · p. 4

Abstract

from arXiv · show

Structure-based virtual screening is an important tool in early stage drug discovery that scores the interactions between a target protein and candidate ligands. As virtual libraries continue to grow (in excess of $10^8$ molecules), so too do the resources necessary to conduct exhaustive virtual screening campaigns on these libraries. However, Bayesian optimization techniques can aid in their exploration: a surrogate structure-property relationship model trained on the predicted affinities of a subset of the library can be applied to the remaining library members, allowing the least promising compounds to be excluded from evaluation. In this study, we assess various surrogate model architectures, acquisition functions, and acquisition batch sizes as applied to several protein-ligand docking datasets and observe significant reductions in computational costs, even when using a greedy acquisition strategy; for example, 87.9% of the top-50000 ligands can be found after testing only 2.4% of a 100M member library. Such model-guided searches mitigate the increasing computational costs of screening increasingly large virtual libraries and can accelerate high-throughput virtual screening campaigns with applications beyond docking.

Introduction

The study applies Bayesian optimization to structure-based virtual screening, using surrogate models to guide docking evaluations and reduce computational costs while preserving the fidelity of structure-based methods.

Introduction

  • Implicit chemical spaces span 10^10 to 10^20 possible compounds (8–10), and high computational cost, including 475 CPU-years in Gorgulla et al. 12, makes exhaustive docking impractical across many distinct protein targets.Source: Introduction · pp. 2–3
  • Bayesian optimization uses a surrogate trained on acquired data to select subsequent evaluations, avoiding brute-force screening of many low-performing compounds.The study treats negative docking score as the black-box objective.Source: Introduction · p. 3
  • Unlike Deep Docking [27], this work models docking scores as continuous values; Gaussian-process training scales poorly to millions of acquired points because of O(N^3) complexity.The introduction also identifies large, discrete design spaces as a relatively unexplored Bayesian-optimization setting.Source: Introduction · p. 4
  • The work reports reducing virtual-library exploration costs by over an order of magnitude while surrogate models prioritize molecules with better docking scores.The authors state that these scores are presumed to be associated with greater likelihood of experimental validation.Source: Introduction · p. 4
  • The study evaluates RF, NN, and MPN surrogates with greedy, UCB, TS, EI, and PI acquisition strategies and varied batch sizes across docking datasets of roughly 10,000 to 100 million molecules.The experiments use MolPAL, software developed and released as open source.Source: Introduction · pp. 4–5

A. Brute Force (previous)

Figure 1 presents an overview of exhaustive brute-force virtual screening and MolPAL, with grey circles representing molecules not yet evaluated.

  • Figure 1 contrasts brute-force (exhaustive) virtual screening with MolPAL and uses grey circles for unevaluated molecules.Source: Figure 1 · p. 5

Results

The Results section examines how MolPAL’s model-guided searches recover top-scoring ligands across docking libraries, including Enamine HTS, and how outcomes vary with model architecture, acquisition strategy, and acquisition size.

Small virtual libraries

Enamine HTS Collection

Ultra-large library

  • The MPN model recovered 87.9% of the top-50,000 scores after exploring 2.4% of the total pool, with EF = 36.6.At batch sizes of 0.2% and 0.1%, recovery was 67.1% (EF = 55.9) and 52.0% (EF = 86.7); UCB yielded 94.8% (EF = 39.5) at 0.4% and 75.5% at 0.2%.Source: Ultra-large library · p. 10
  • Enrichment factors from AmpC and Enamine HTS are not directly comparable because their docking-score precision and chemical spaces differ.AutoDock Vina's lower-precision scores make the top-k score metric more forgiving.Source: Ultra-large library · p. 11

Single-iteration active learning

  • On Enamine HTS, single-iteration greedy MPN screening found 94.9% (± 0.8) of top-1000 scores after exploring 2.4% of the library (EF = 36.5), versus 97.7% (EF = 37.6) for active learning.Source: Figure 6 · p. 12, Single-iteration active learning · p. 12
  • 70.3% of top-1000 scores were found after the first exploration batch for MPN with 0.4% acquisition, up from 0.4% after initialization.This sharp increase suggests that the randomly selected initial batch was informative.Source: Single-iteration active learning · pp. 11–12, Figure 6 · p. 12
  • 77.4% of top-1000 scores (EF = 29.8) were found by MPN when the initial batch was 42,000 molecules and the single acquisition batch was 8,400.This was worse than any active-learning strategy at the same number of function evaluations; NN and RF performance dropped even more with a large initial and small exploration batch.Source: Single-iteration active learning · pp. 12–13

Chemical space visualization

  • Top-1000 compounds clustered in sparse regions of chemical space; NN and MPN acquired more candidates there earlier than RF, which sampled densely only by iteration three and shifted away by iteration five.The chemical-space analysis was qualitative.Source: Chemical space visualization · pp. 13–14

Dynamic convergence criterion

  • On the Enamine HTS collection with greedy acquisition and 0.1% batch size for both initialization and exploration, MPN converged sooner than NN, then RF, reaching 88.7%, 85.4%, and 75.8% of top-1000 scores, respectively.Source: Dynamic convergence criterion · p. 13, Figure 7 · p. 14

Discussion

The discussion finds that greedy acquisition can explore effectively, and that larger libraries are more amenable to surrogate-guided search. It also highlights consistency across trials alongside tradeoffs and limitations in evaluation, batch sizing, and surrogate-model costs.

Effect of acquisition strategy on performance

  • Despite its exploitative design, greedy acquisition explored adequately and outperformed UCB, TS, EI, and PI; unreliable regression uncertainty estimates may explain the alternatives’ poorer results.The discussion notes that greedy acquisition could in principle focus on one region of chemical space.Source: Effect of acquisition strategy on… · pp. 14, 16

Effect of library size

  • Top-k performance generally increased with library size at a constant fractional k, even as relative exploration decreased, alongside greater training-data exposure.The number of training points rose from roughly 500 for Enamine 50k at 1% batch size to nearly 8,400 for Enamine HTS at 0.4% batch size.Source: Effect of library size · p. 16
  • Prediction MSE decreased with library size from 0.3504 to 0.2617 to 0.1520, while Spearman’s ρ rose from 0.6699 to 0.7826 to 0.9094.The authors suggest larger libraries are not proportionally more chemically diverse, making their scores easier to predict with a proportional increase in training data.Source: Effect of library size · pp. 16–17

Consistency across repeated trials

  • Performance was consistent across datasets and configurations, and repeated trials appeared to converge on the same optimal library subset regardless of initialization.The authors hypothesize that relatively smooth structure-property landscapes and a lack of statistical uncertainty contribute to this convergence.Source: Consistency across repeated trials · p. 17

Limitations of evaluation metrics

Optimal batch size for active learning

  • Batch size balances costly sequential retraining and inference against the potentially suboptimal results of selecting all molecules in one iteration; experiments used batches one sixth of exploration size.The authors identify dynamic batch sizes as a topic for future performance engineering.Source: Optimal batch size for active learning · p. 18

Cost of surrogate model (re)training

  • MPN models performed better on the largest datasets but cost more than fingerprint-based NN models, and online learning consistently performed worse than full retraining.The study cautions that its MPN and NN costs are not directly comparable because of implementation, parallelization, and precalculation differences.Source: Cost of surrogate model (re)training · p. 19

Conclusion

The study demonstrates Bayesian optimization for prioritizing compounds in structure-based virtual screening libraries, reducing docking calculations while retaining many top-scoring ligands. It highlights greedy acquisition and message-passing models as effective and presents MolPAL as a basis for broader screening workflows.

Conclusion

  • Across libraries ranging from 10k to 100M ligands, evaluation found strong performance from greedy acquisition and superior performance from message passing neural networks over fingerprint-based feed-forward networks or random forests.Source: Conclusion · p. 19
  • In the 100M-member library screened against 12LS, greedy acquisition identified 87.9% of the top-50000 scoring ligands with a >40-fold reduction in docking calculations, while UCB identified 94.8%.The study describes this as the largest library tested.Source: Conclusion · pp. 19–20
  • The authors propose model-guided prioritization as a drop-in replacement for exhaustive virtual screening when the exact top-k compounds are not required, and note its relevance to experimental high-throughput screening.They describe experimental high-throughput screening as an expensive tool for challenging drug-discovery problems.Source: Conclusion · p. 20
  • Future work will extend MolPAL, test it prospectively on the Enamine REAL database, add surrogate architectures and improved uncertainty estimation, and apply it to other virtual-discovery objectives.Source: Conclusion · p. 20
  • The authors envision making MolPAL more accessible by pipelining the entire virtual-screening process behind a common interface.Source: Conclusion · p. 20

Methods

The methods describe MolPAL’s batched Bayesian optimization workflow for molecular pools, including surrogate-model updates and batch selection for docking-score evaluation. They also specify the docking datasets, acquisition and surrogate configurations, and top-k recovery metrics used to assess the approach.

Batched Bayesian optimization

  • Batched Bayesian optimization scores an initial random set of pool molecules with AutoDock Vina, trains a surrogate, and iteratively acquires batches until a stopping criterion is met.The acquired molecules’ objective values update the training dataset at each iteration.Source: Batched Bayesian optimization · pp. 20–21

Acquisition metrics

Surrogate models

  • Three surrogate architectures were investigated: random forests, feed-forward neural networks using 2048-bit atom-pair fingerprints, and directed message-passing neural networks operating on molecular graphs.Source: Surrogate models · pp. 22–24

Datasets

  • The docking datasets comprise Enamine Discovery Diversity and HTS compounds docked against thymidylate kinase (PDB ID: 4UNN) with AutoDock Vina.Ligands were parsed from Enamine SDFs into SMILES with RDKit and converted to PDBQT using OpenBabel.Source: Datasets · p. 25
  • The AmpC screening dataset published by Lyu et al. was used as-is [34], and receptor preparation used PDB 4UNN chain A with heterogens deleted, missing residues and heavy atoms added, and hydrogens added at pH 7.0.Source: Datasets · p. 25

Evaluation metrics

  • The study evaluated the average top-k docking score found by MolPAL relative to the true top-k average score in the full dataset.This metric is the average of explored top-k scores divided by the full dataset’s true top-k scores.Source: Evaluation metrics · pp. 25–26
  • SMILES measured the fraction of true top-k molecules recovered, while Scores measured the fraction of true top-k scores recovered.Both metrics divide the size of an intersection with the corresponding true top-k set by k.Source: Evaluation metrics · pp. 25–26

Hyperparameter optimization

  • Experiments covered only a small fraction of MolPAL configurations; exhaustive hyperparameter optimization was outside scope despite support for varied fingerprints, preclustering, and deep-learning confidence estimates.Supported options included RDKit, Morgan, and MACCS fingerprints.Source: Hyperparameter optimization · p. 26
  • Broad trends on Enamine 10k and 50k datasets showed only minor performance variations across the settings examined.This finding supported not rigorously screening all possible configurations.Source: Hyperparameter optimization · p. 26

Software design

  • MolPAL iteratively retrained its model on observed data, predicted means and model-dependent uncertainties for pool molecules, and selected the top-m untested candidates for objective evaluation.The Explorer coordinates these steps through MoleculePool, Model, Acquirer, and Objective helper classes.Source: Software design · pp. 26–27
  • All experiments were retrospective, using a LookupObjective with a fully generated dataset as the lookup table for objective values.Source: Software design · p. 27

Supporting Information Available

Additional methods and results are in the supporting information, while code and data for reproducing the study and its figures are available online.

Supporting Information Available

  • The supporting information contains additional methods and results, and code and data needed to reproduce the study and figures are available at a GitHub address.The supplied address is truncated to “https://github.”Source: Supporting Information Available · p. 27

Supporting Information

The supplied Supporting Information passage contains only the partial title “Accelerating High-Throughput Virtual Screening Through.”

Supporting Information

Additional Methods

Additional methods describe MolPAL’s modular software and data handling, alternative uncertainty-estimation models, and online retraining.

Elaboration on software design

Alternative surrogate models

Retraining strategy

  • Online training updated the previous model using only newly acquired data; this strategy applied to NN and MPN models, not RF.Source: Retraining strategy · p. 37

Additional Results · Graphical TOC Entry

Supplementary results describe docking-score distributions, Bayesian optimization performance across libraries and batch sizes, and additional views of search diversity and chemical-space coverage.

Dataset score distributions

  • Docking-score distributions are shown for Enamine 10k, 50k, HTS, and AmpC, with dashed lines marking the 100th, 500th, 1,000th, and 50,000th best scores, respectively.Each distribution uses a bin size of 0.1.Source: Figure S2 · p. 38, Figure S3 · p. 38, Figure S4 · p. 39

Library exploration across separate experiments

  • Across five greedy optimizations, figures track the number of unique SMILES acquired in the 10k, 50k, and HTS libraries against theoretical maximum and minimum bounds.The bounds represent repeated trials selecting distinct subsets versus identical subsets of molecules.Source: Figure S6 · p. 40, Figure S7 · p. 40

Online training strategy

  • For Enamine 10k, 50k, and HTS, figures compare online training with full model retraining over five iterations, using batch sizes of 1.0%, 1.0%, and 0.4%, 0.2%, or 0.1%, respectively.Performance is measured as the percentage of top-100, top-500, or top-1,000 scores found; error bars show ±1 standard deviation across five runs.Source: Figure S8 · p. 41, Figure S9 · p. 41, Figure S10 · p. 42
  • For AmpC, figures compare online training with full model retraining over five iterations at 0.4%, 0.2%, and 0.1% batch sizes.Performance is measured as the percentage of top-50,000 scores found, with error bars showing ±1 standard deviation across three runs.Source: Figure S13 · p. 43, Figure S14 · p. 44, Figure S15 · p. 44

Bayesian Optimization Performance

  • Final performance tables report percentage-based Bayesian optimization results for Enamine 10k, 50k, and HTS using 1.0%, 1.0%, and 0.4%, 0.2%, or 0.1% batch sizes, respectively.Metrics use the top 100, 500, or 1,000 compounds found; values are averages (standard deviations) over five runs, with higher scores better.Source: Table S1 · p. 45, Table S2 · p. 46, Table S3 · p. 47
  • Final AmpC performance tables report percentage-based results for top-50,000 compounds found at 0.4%, 0.2%, and 0.1% batch sizes.Values are averages (standard deviations) over three runs, with higher scores better.Source: Table S6 · p. 50, Table S7 · p. 51, Table S8 · p. 52

Chemical Space Visualization

  • A chemical-space visualization maps Enamine HTS molecules at a specified iteration for greedy acquisition with a 0.1% batch size, marking high-scoring clusters in sparse regions.The 2-D UMAP embedding uses 2048-bit Atom-pair fingerprints and was trained on a random 10% subset of the library; color represents negative docking score, with higher better.Source: Figure S16 · p. 54

Related papers

Loading 2012.07127v1…