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

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

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:2012.07127v1Dec 13, 2020q-bio.QMcs.LG

TL;DR

The paper addresses the impracticality of applying exhaustive screening as virtual libraries grow and considers Bayesian optimization as a strategy for reducing required evaluations. It leverages Bayesian optimization algorithms for docking simulations while noting scalability limitations for Gaussian-process training and challenges in comparing datasets.

Problem
As virtual libraries grow, exhaustive screening becomes impractical and new strategies are needed to mitigate its computational cost.
Method
The work leverages Bayesian optimization algorithms for docking simulations, using surrogate models to reduce required evaluations.
Results
The study demonstrates that surrogate machine learning models can decrease computational costs by an order of magnitude.
Takeaways & Limitations
Bayesian optimization is presented as a framework for reducing the evaluations required in large-library docking simulations.
Takeaways & Limitations
Gaussian-process training does not scale well to millions of acquired data points because of its O(N^3) complexity.

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

Exhaustive structure-based virtual screening becomes increasingly costly as chemical libraries expand, motivating Bayesian-optimization strategies that reduce evaluations while preserving docking-based prioritization.

Motivation: Exhaustive screening spends substantial computational resources evaluating relatively low-performing compounds, making large-library screening impractical for many protein targets.One cited structure-based campaign required 475 CPU-years.
Approach: Bayesian optimization uses a surrogate model trained on acquired data to guide subsequent evaluations and reduce the number of required experiments.The framework is applied to a large, discrete design space that is relatively unexplored for Bayesian optimization.
Approach: This study applies Bayesian optimization to docking while treating docking scores as continuous regression targets rather than binary labels.The approach is intended to preserve structure-based screening fidelity while reducing computational cost.
Results: Surrogate-guided screening reduces computational cost by over an order of magnitude and prioritizes molecules associated with better docking scores across libraries up to 100 million molecules.MolPAL is the open-source software developed for these experiments.

A. Brute Force (previous)

Figure 1 contrasts exhaustive virtual screening with molecular pool-based active learning; grey circles mark molecules that have not been evaluated.

A. Brute Force (previous): Figure 1A depicts brute-force exhaustive virtual screening, while Figure 1B depicts molecular pool-based active learning (MolPAL).Grey circles represent molecules that have not been evaluated.

Results

Across docking datasets, Bayesian optimization found many high-scoring ligands while evaluating only a small fraction of each library. Performance varied with surrogate architecture, acquisition strategy, and batch size, with MPN models often strongest in larger-library tests.

Larger virtual libraries: 74.8% and 72.9% of Enamine 50k top-500 scores were found after exploring 6% with NN and MPN models, respectively.These results corresponded to enrichment factors of 11.3 and 11.0 over the random baseline.
Ultra-large library: 97.7% of Enamine HTS top-1000 scores were found by MPN after exploring 0.4% batches, compared with 2.6% for the random baseline.The RF, NN, and MPN models achieved enrichment factors of 32.4, 36.8, and 37.6, respectively.
Convergence: Higher-capacity models converged sooner and to a higher percentage of top compounds, following MPN > NN > RF.The convergence analysis used greedy acquisition with 0.1% initialization and exploration batches.

Discussion

Across docking datasets and configurations, Bayesian optimization consistently reduced screening effort, with greedy acquisition performing especially well despite its exploitative design. Larger libraries also showed stronger top-k performance, while model choice and batch size introduced computational and methodological trade-offs.

Acquisition strategy: Greedy acquisition consistently provided adequate exploration and outperformed acquisition metrics combining exploration with exploitation.The comparison included UCB, TS, EI, and PI.
Library size: As library size increased, top-k performance increased at a constant fractional k, even as relative exploration size decreased.The authors attribute this partly to greater surrogate-model exposure to training data during optimization.
Library size: MSE decreased from 0.3504 to 0.2617 to 0.1520 across the 10k, 50k, and HTS libraries, while Spearman’s ρ increased from 0.6699 to 0.7826 to 0.9094.These values followed initialization with random 1%, 1%, and 0.4% batches, respectively.
Overall performance: Results showed consistent performance across all datasets and configurations.The paper reports this consistency across Tables S1-S8.
Repeated trials: Repeated trials converged toward the same optimal subset of the library regardless of initialization.The authors hypothesize that relatively smooth structure-property landscapes and limited statistical uncertainty contribute to this pattern.
Evaluation metrics: Direct comparison of average metrics across datasets is challenging because score scales and distributions differ.The average metric is sensitive to score scale, while top-k SMILES can be overly strict when tied scores are arbitrarily ordered.
Batch size: Choosing batch size requires balancing model freshness and computational cost, and the principled balance remains an open challenge.Sequential learning improves model recency but increases retraining and inference costs, whereas very large batches can produce suboptimal performance.
Surrogate models: The MPN model was superior on the largest datasets, but its costs were markedly higher than those of the fingerprint-based NN model.This reflects a trade-off between sample efficiency and surrogate-model cost.

Conclusion

Bayesian optimization was demonstrated for prioritizing compounds in structure-based virtual screening across libraries ranging from 10k to 100M ligands. The evaluation found strong performance for greedy acquisition and message passing neural network surrogates, while extending the approach toward experimental screening and broader virtual discovery.

Conclusion: MolPAL is intended to increase accessibility by placing the virtual-screening process behind a common interface.
Conclusion: The evaluation covered acquisition metrics and surrogate architectures across virtual libraries ranging from 10k to 100M ligands.
Conclusion: Greedy acquisition showed surprisingly strong performance, while message passing neural networks outperformed fingerprint-based feed-forward neural networks and random forests.
Conclusion: 87.9% of the top-50000 scoring ligands were identified in a 100M-member library with a >40-fold reduction in docking calculations using greedy acquisition.Using UCB acquisition identified 94.8% for the same number of objective-function calculations.
Conclusion: The model-guided approach is proposed as a drop-in replacement for exhaustive screening when the exact top-k set is unnecessary.The authors also identify experimental high-throughput screening as a relevant application.
Conclusion: Future work includes prospective screening with MolPAL, additional surrogate architectures, improved uncertainty estimation, and other virtual-discovery objectives.

Methods

The study applies batched Bayesian optimization to a discrete molecular pool, iteratively training surrogate models and acquiring batches for docking evaluation.

Batched Bayesian optimization: Batched Bayesian optimization selects molecules from a discrete virtual library rather than evaluating candidates one at a time.The pool is treated as a candidate set, and points are acquired in batches.
Batched Bayesian optimization: The workflow initializes a dataset with randomly selected molecules, calculates their docking scores, and trains a surrogate model on the resulting data.Docking scores from AutoDock Vina serve as the objective-function values.
Batched Bayesian optimization: Acquisition functions assign utility to unevaluated candidates, and the batch with the largest summed utility is acquired for objective-function evaluation.Utility can reflect predicted value, information gain, or the likelihood of improving the current maximum.
Batched Bayesian optimization: The acquired objective-function values update the dataset, after which the process repeats until a stopping criterion is reached.Stopping criteria include a fixed iteration count or insufficient improvement.
Surrogate models: The study tests random forests, feed-forward neural networks, and directed message-passing neural networks as surrogate architectures.RF and FFNN models use molecular fingerprints, whereas MPNNs operate directly on molecular graphs.
Model training: The experiments use fixed model settings without additional hyperparameter optimization, retraining each surrogate from scratch with all acquired data at every iteration.This design keeps the compared surrogate models and training procedure consistent across iterations.

Supporting Information Available

The supporting information contains additional methods and results, along with code and data for reproducing the study and its figures.

Additional methods and results are provided in the supporting information.
The code and data needed to reproduce the study and its figures are available online.

Supporting Information

The supporting information is associated with accelerating high-throughput virtual screening.

The material concerns high-throughput virtual screening.
The stated focus is acceleration.
The title identifies virtual screening as the application area.

Molecular Pool-Based Active Learning

The paper is authored by David E. Graff, Eugene Shakhnovich, and Connor W. Coley, affiliated with Harvard University and MIT.

The listed authors are David E. Graff, Eugene Shakhnovich, and Connor W. Coley.
The affiliations include Harvard University’s Department of Chemistry and Chemical Biology and MIT’s Department of Chemical Engineering.

Additional Methods

MolPAL uses modular helper classes and disk-backed molecular data to support testing and extending surrogate-model and acquisition-strategy combinations. Its additional model variants include uncertainty-estimation approaches and online training, with some alternatives excluded from experiments for performance or inference-cost reasons.

Software design: MolPAL separates molecule storage, modeling, acquisition, and objective optimization into helper classes for testing different combinations.The Model and Objective classes use minimal abstract interfaces and an adapter pattern for machine-learning libraries.
Software design: MolPAL can interface with Scikit-Learn, PyTorch, and TensorFlow through the Model and Objective adapter interfaces.This design supports straightforward extension with new surrogate models, acquisition metrics, and objective functions.
MoleculePool: MoleculePool stores SMILES strings and optional fingerprints on disk, supporting clustering, vector-input models, and virtual libraries of different sizes.Molecular graphs for the MPN model are recalculated as needed because they cannot currently be stored in memory or on disk.
Confidence estimation: An MPN dropout model estimates uncertainty by averaging 10 stochastic forward passes and using their prediction variance.The model uses dropout probability 0.2 during inference, with the dropout layer placed before the hidden layer.
Confidence estimation: Five-model ensembles and mean-variance estimation models were implemented but excluded from experiments because of lower performance than the base models.The MPN dropout alternative was also excluded because its inference cost was significantly higher.
Training strategy: Online training updates neural-network and MPN surrogates using only newly acquired data, whereas random forests are reinitialized whenever fitted.The alternative to online training retrains the surrogate from scratch using all acquired data.

Additional Results

The additional results report docking-score distributions, repeated-trial molecule diversity, and Bayesian-optimization performance across four libraries. Performance is evaluated by the percentage of top-ranked compounds found as a function of evaluated ligands and across multiple batch sizes.

Docking-score distributions: Docking-score distributions are shown for Enamine 10k, Enamine 50k, Enamine HTS, and AmpC libraries, with dashed lines marking the kth-best-score thresholds.The thresholds are k = 100, 500, 1000, and 50000, respectively.
Greedy optimization diversity: Repeated greedy optimizations quantify the total number of unique SMILES acquired across five trials for the 10k, 50k, and Enamine HTS libraries.The figures include theoretical maximum and minimum bounds corresponding to distinct versus identical selected subsets across trials.
Optimization performance: Bayesian-optimization traces measure the percentage of top-ranked scores found against the number of ligands evaluated across Enamine 10k, 50k, HTS, and AmpC datasets.The traces distinguish acquisition metrics, surrogate architectures, online versus full retraining, and dataset-specific initial and incremental acquisition sizes.
Batch-size comparisons: Enamine HTS experiments compare 0.4%, 0.2%, and 0.1% acquisition batches while targeting the top-1000 compounds.The corresponding figures report performance as a function of the number of evaluated ligands.
Batch-size comparisons: AmpC experiments compare 0.4%, 0.2%, and 0.1% acquisition batches while targeting the top-50000 compounds.Results are reported over three runs in the corresponding performance tables.
Final performance tables: The supplementary tables report final optimization performance as percentages averaged across five runs for Enamine datasets and three runs for AmpC.Higher values are better, with batch sizes specified separately for each dataset.