Skip to main content
NCBI home page

This is a preprint.

It has not yet been peer reviewed by a journal.

The National Library of Medicine is running a pilot to include preprints that result from research funded by NIH in PMC and PubMed.

bioRxiv logoLink to bioRxiv
[Preprint]. 2025 Feb 27:2025.02.25.639879. [Version 1] doi: 10.1101/2025.02.25.639879

A database for large-scale docking and experimental results

Brendan W Hall 1,†, Tia A Tummino 1,†, Khanh Tang 1, John J Irwin 1,*, Brian K Shoichet 1,*
PMCID: PMC11888352  PMID: 40060496

Abstract

The rapid expansion of readily accessible compounds over the past six years has transformed molecular docking, improving hit rates and affinities. While many millions of molecules may score well in a docking campaign, the results are rarely fully shared, hindering the benchmarking of machine learning and chemical space exploration methods that seek to explore the expanding chemical spaces. To address this gap, we develop a website providing access to recent large library campaigns, including poses, scores, and in vitro results for campaigns against 11 targets, with 6.3 billion molecules docked and 3729 compounds experimentally tested. In a simple proof-of-concept study that speaks to the new library’s utility, we use the new database to train machine learning models to predict docking scores and to find the top 0.01% scoring molecules while evaluating only 1% of the library. Even in these proof-of-concept studies, some interesting trends emerge: unsurprisingly, as models train on larger sets, they perform better; less expected, models could achieve high correlations with docking scores and yet still fail to enrich the new docking-discovered ligands, or even the top 0.01% of docking-ranked molecules. It will be interesting to see how these trends develop for methods more sophisticated than the simple proof-of-concept studies undertaken here; the database is openly available at lsd.docking.org.

Graphical Abstract

graphic file with name nihpp-2025.02.25.639879v1-f0001.jpg

Introduction

In the last six years, with the advent of make-on-demand (“tangible”) libraries1,2, readily available chemical space has increased by over four orders of magnitude. Structural modeling of these libraries has supported a concomitant growth in the size and success of prospective molecular docking campaigns3–8, with applications ranging from G protein-coupled receptors (GPCRs)5, 9, 10, to soluble enzymes11, 12, transporters7, and other targets. The first screens of these tangible libraries consisted of only 99 to 138 million molecules1 and have grown to over 1 billion explicitly docked molecules9, 13 more recently. Even larger docking campaigns exceeding 11 billion interpolated synthons14 and adaptively sampling from 69 billion molecules have been undertaken15. While most of these campaigns typically test fewer than 100 docking-prioritized molecules from the larger libraries, for several targets hundreds of molecules have been tested, and overall the number of tested predictions from large-scale docking (LSD) exceeds 10,000 molecules in the open literature16–24.

The results of these campaigns, both experimental and simply docking scores and poses, could benefit the community, acting as benchmarking and training sets for machine learning (ML) among other applications25–36. However, these results have not been made available at scale in a readily accessible way. For instance, a large library docking paper may report docking scores only for the synthesized molecules11, 37, 38 or sometimes only for those that were experimentally active7, 12, representing a tiny portion of the overall docked library or even its high-ranking subset. While some papers report scores for many of the docked molecules, hosted on open-source platforms like GitHub8 or other publicly accessible websites1, 10, 39, information beyond the docking scores, such as docked poses or the energy potential grids used to score the molecules, is rarely provided.

Here we describe an open-source benchmarking set of published docking results for 6.3 billion explicitly evaluated molecules across eleven protein targets (lsd.docking.org). The benchmarking set includes docking scores (using DOCK3.7/3.8), SMILES, poses for top molecules, results for in vitro tested molecules, and the docking energy grids used to produce them. As a simple proof of concept, we consider the ability of ML models trained on this set with the widely used Chemprop framework40 to predict docking scores (subsequently referred to as Chemprop models). Applications of these models to the chemical space exploration technique Retrieval Augmented Docking (RAD)41 are considered.

Results

Organization of large-library docking data.

We began by collecting the source files for the published large-scale docking campaigns run in our lab to date, covering 11 different targets (Table 1, Fig. 1A, Table S1). Within a given target, one or more separate large-scale screens were run, typically differentiated by ligand size (e.g., lead-like versus fragments), docking grids (e.g., input structure differences or different parameter optimizations), or purchasability (e.g., in-stock versus make-on-demand).

Table 1.

Overview of the docking screens included on lsd.docking.org and the number of docking scores and experimental results provided for each target.

Target Compounds with docking scores Compounds experimentally tested
Alpha2AR 30,518,811 82
AmpC 1,568,323,216 1565
CB1R 18,992,691 46
D4 138,312,677 552
EP4R 381,067,069 71
MPro 1,108,167,275 393
MT1R 40,376,489 38
NSP3_Mac1 686,555,212 240
SERT 246,614,514 13
Sigma2 468,639,651 506
5HT2A 1,630,264,067 223
Open in a new tab

Figure 1.

Figure 1.

Open in a new tab

(A) Overview of the directory structure on lsd.docking.org. (B) Types of data available for each screen: docking scores, in vitro experimental data, and poses for the best-scoring molecules.

For each screen, we gathered three levels of data (Fig. 1B). First, we collected what may be the most useful to the ML community, the “Docking Results”. These include the list of scored molecules, their SMILES, and the docking score for each molecule from every screen. To support these, we also collected the docking energy potential grids used to score the molecules in each screen.

Next, we collected 3D poses for the top 500,000 unfiltered molecules from each screen in mol2 format, which can be visualized with software such as Chimera42, PyMoL43, or Maestro44, among others. Finally, we collected the in vitro experimental results and merged them with the docking results for in vitro tested molecules. Together, the set provides docking scores for over 6.3 billion ligand-target pairs and provides over 10 million docking poses for the high-ranking molecules from each docking campaign.

To make these data readily accessible, we developed the website lsd.docking.org. Because some targets have been the focus of multiple large-library screens, and because some papers screened multiple targets, users can search either by target or by paper to access the data. Each level of data, docking results, poses, in vitro results, and docking grids (dockfiles), is a subfolder nested within that of the individual screen. If multiple files are hosted within a folder, a zip file containing the contents is provided for easier dissemination. A README.md is provided within each screen folder explaining the contents of the subfolders.

For example, to access the docking scores for the dopamine D4 receptor screen, a user would click the “Browse by Target” button which navigates to a page showing every available target. Clicking “View” for the D4 target shows a folder called “Lyu_2019” and clicking the folder brings the user to a page showing the folders for each level of data from the screen: “dockfiles,” “Poses,” “InVitroResults,” and “DockingResults.” Finally, clicking on “DockingResults” navigates to a screen with the downloadable file “D4_screen_table.csv.gz” which contains the ZINC IDs, SMILES, and docking scores for every molecule in the screen.

Impact of training set size and sampling on Chemprop performance.

While we do not introduce or apply new ML methods here, we thought it would be interesting to train ML models with these datasets as a proof of concept for the possible utility of this dataset. Accordingly, we applied the well-known framework Chemprop40, 45 on subsets of the docking scores from campaigns against AmpC β-lactamase (AmpC)4, 5HT2A9, and Sigma239, containing 1,468,863,655, 1,630,264,067, and 468,639,651 molecules, respectively. The large datasets allowed us to systematically investigate the effects of training set size and data sampling strategies.

We investigated training set sizes of 1,000, 10,000, 100,000, and 1,000,000 molecules sampled from the datasets using three strategies: random sampling across the entire dataset, random sampling from the top ranking 1% of molecules only, and a stratified approach where 80% of the training set was randomly sampled from the top ranking 1% of molecules and the remaining 20% was sampled from the rest of the dataset. We created test sets by randomly sampling 100,000,000 molecules from each dataset. We evaluated model performance using the Pearson correlation between predicted and true scores across the entire test set and the top 0.01% scoring molecules, ensuring no overlap with the training set. Additionally, we measured the logAUC, which quantifies the fraction of the top 0.01% molecules found as a function of the screened library fraction on the logarithmic scale (Fig. 2C). This metric captures the model’s ability to enrich for the true top 0.01% scoring molecules in its top predictions and is widely used in docking studies46–48.

Figure 2.

Figure 2.

Open in a new tab

(A) Overview of the datasets, training set sizes, sampling strategies, and evaluation metrics used to investigate the ability of Chemprop models to predict DOCK scores. (B) Effects of increasing the training set size on the three evaluation metrics for AmpC. The bars represent the minimum and maximum performance across five replicates. (C) AmpC log recall curves of the top 0.01% scoring molecules for the three sampling strategies using increasing training set sizes. The shaded region represents the minimum and maximum performance across five replicates. The dashed black line represents a random predictor.

Unsurprisingly, increasing the training set size improved model performance across nearly all targets, sampling methods, and evaluation metrics (Fig. 2B, Fig. 2C). For instance, AmpC models trained with random sampling achieved overall Pearson correlations of 0.65, 0.77, 0.83, and 0.86 between predicted and true scores when trained with 1,000, 10,000, 100,000, and 1,000,000 molecules, respectively (Fig. 2B) (we note that even a million molecules are only 0.07% of the overall library docked). Intriguingly, the overall Pearson correlation between a model’s predictions and true scores did not reliably indicate its ability to enrich for the top 0.01% scoring molecules or for the true binders. For example, on the AmpC dataset with 100,000 training molecules (0.007% of the docked library), random sampling achieved a Pearson correlation of 0.83, higher than the 0.76 with stratified sampling (Fig. 2B). However, the same random sampling achieved a logAUC of only 0.49 for the recall of the top 0.01% scoring molecules, whereas stratified sampling achieved a logAUC of 0.77 for this 0.01%--i.e., many more of the highest-ranking molecules were found in the latter’s top predictions than the former’s, despite the former’s better R-value. Similarly, random sampling only achieved a logAUC of 0.47 for the recall of true, novel inhibitors found in the AmpC docking campaign, whereas stratified sampling had a logAUC of 0.87 for these inhibitors (Fig. 2C and Fig. 3B).

Figure 3.

Figure 3.

Open in a new tab

Chemprop model recall curves of the experimental actives after filtering out molecules that share a scaffold with any molecule in the training set for the (A) 5HT2A receptor, (B) AmpC receptor, and (C) Sigma2 receptor. The shaded region represents the minimum and maximum performance across five replicates. The dashed black line represents a random predictor.

Overall, despite relatively high Pearson correlations, few of the top 0.01% scoring molecules and the true ligands were in the top 0.01% of ML predictions – the region realistically inspected during large-scale screening. For example, the random sampling strategy on the AmpC dataset with 1 million training molecules (0.05% of the docked library) achieved the highest Pearson correlation of 0.86, but only 8.4% of the top 0.01% scoring molecules and 13.5% of the true binders were in its top 0.01% predictions. These results also held true for the 5HT2A and Sigma datasets (Fig. 3, Fig. S1, Fig. S2).

Retrieval Augmented Docking using Chemprop models.

We next investigated the utility of the lsd.docking.org benchmarks for evaluating chemical space exploration methods. We assessed the ability of Retrieval Augmented Docking (RAD)41 to find the top 0.01% scoring molecules from billion-scale chemical libraries. Drawing on a technique from large-scale vector databases, as used in social networks, RAD organizes chemical libraries into Hierarchical Navigable Small World (HNSW) graphs49, enabling efficient exploration of chemical space (Fig. 4A). Previous studies found that traversing the HNSW graph with the DOCK3.7 scoring function finds most of the top 0.01% scoring molecules while only needing to explicitly evaluate a fraction of the library. Here, we constructed HNSW graphs for ~1.3 billion molecules from the AmpC and 5HT2A datasets, and ~500 million molecules from the Sigma2 dataset.

Figure 4.

Figure 4.

Open in a new tab

(A) Overview of Retrieval Augmented Docking (RAD). (B) Recall curves of the top 0.01% scoring molecules when evaluating a RAD-prioritized 1% of each library with the DOCK scoring function. (C) 5HT2A recall curves of the top 0.01% scoring molecules when evaluating a RAD-prioritized 1% of the library with the Chemprop models trained with different sampling strategies. The curves are averaged over five replicates. (D) 5HT2A recall curves of the top 0.01% scoring molecules when evaluating a RAD-prioritized 1% of the library with the Chemprop models trained with different training set sizes. The curves are averaged over five replicates.

On these systems, RAD found a large fraction of the top 0.01% scoring molecules while explicitly scoring only 1% of these billion molecule libraries (RAD deprioritizes the rest of the library as it traverses the graph). For the ~1.3 billion molecule libraries, RAD found 84% of AmpC and 75% of 5HT2A top 0.01% scoring molecules (Fig. 4B). For the smaller 500 million molecule Sigma2 library, RAD found 71% of the top 0.01% scoring molecules, again while scoring a RAD-prioritized 1% of the total library (Fig. 4B).

We next asked whether the machine learning models trained to predict DOCK3.7/3.8 scores could effectively replace the DOCK scoring function in the context of RAD. With the best performing ML models, RAD using an ML scoring function found a comparable number of the top 0.01% scoring molecules as the full DOCK3.7/3.8 scoring function, but at much reduced computational cost due to the faster inference speed of machine learning models (~1 sec/core to orient and score a molecule with DOCK3.8, versus ~5 ms/core to do Chemprop ligand preparation and scoring). For example, RAD prioritization of 1% of the ~1.3 billion 5HT2A library found 75% of the top 0.01% scoring molecules using the DOCK3.8 scoring function and 71% using the ML model trained on 1 million molecules with a stratified sampling strategy (Fig. 4C). We do note that the variance in RAD performance across replicates was larger than in the score prediction task (Fig. S3).

Here too, the models trained on larger datasets found more of the top 0.01% scoring molecules from a RAD-prioritized 1% of the library than those trained on smaller datasets. For instance, using the stratified sampling method with 5HT2A, RAD found 14%, 54%, 62%, and 71% of the top 0.01% scoring molecules among its top-prioritized 1% of the library with the models trained on 1,000, 10,000, 100,000, and 1,000,000 molecules, respectively (Fig. 4C). Models trained on larger datasets with the stratified sampling strategy continued to outperform those trained with random sampling or sampling only from the top 1% of scores (Fig. 4D). For example, on the 5HT2A library with models trained on 100,000 molecules, RAD found 62%, 60%, and 53% of the top 0.01% scoring molecules when screening a RAD-prioritized 1% of the library when using stratified sampling, top 1% sampling, and random sampling, respectively (Fig. 4D). These trends were also observed for the AmpC and Sigma2 datasets (Fig. S4 and Fig. S5).

Discussion & Conclusions

There is much interest in training ML methods to reproduce docking results with a view to screening larger libraries faster, and naturally there is longstanding interest in improving scoring for docking and for virtual screening. Large benchmarks of docking results, both computational and experimental, would support these efforts but currently are unavailable to the community. Here we begin to develop an open access benchmark of docking and experimental results (lsd.docking.org). Three results merit emphasis. First, the scale of the benchmark may be appropriate to begin to support these goals. The docking screens are against eleven diverse targets, representing three unrelated enzymes, one transporter, and seven GPCRs (in three families), and over 6.3 billion molecules are scored. Docking scores are provided for all molecules successfully fit to the eleven targets, and for 10 million so too are docking poses. The results are presented in a format that we hope lends itself to machine reading. Second, the experimental activities are provided for 3,729 molecules across the eleven targets. These include both molecules that were experimentally active, as predicted by docking, and molecules that were docking false positives. For three targets, AmpC β-lactamase, Sigma2, and the dopamine D4 receptor, it includes the results not only of testing high-ranking molecules, but molecules across the scoring range—high-scoring, mediocre scoring, and poor scoring, all tested by experiment. As docking false positives can be informative both for scoring function optimization and ML training, these sets may be particularly useful. Third, in proof-of-concept studies, we investigated the sorts of ML model training and methods that the large-scale docking benchmarks might support. In particular, we trained ML models to predict a molecule’s DOCK score from its SMILES string and used these models as the scoring function for the HNSW method, itself drawn from large-scale vector database applications. In both cases, we found that increasing the training set size improved model performance. Despite some models obtaining high Pearson correlations between true scores and predictions, relatively few of the top 0.01% scoring molecules and true binders were in the top ML predictions, highlighting the pitfall of Pearson correlation as the sole evaluation metric. When applied in Retrieval Augmented Docking, the best-performing models were comparable to the full DOCK scoring function while much reducing computational cost.

Several caveats warrant mentioning. The scores reported here are all from DOCK3.7/3.8, a physics-based scoring function. This function has the advantage of allowing campaigns in the multi-billion molecule range, but the scores returned, though apparently correlated with hit-rate1, 4, 39, are offset from free energies of binding. Further, it represents only one of several widely used scoring functions. In the proof-of-concept ML training and optimization, a key limitation is the overlap between test sets and training sets. Although no molecule appeared in both the training and the test sets, we made no other effort to ensure a diverse or dissimilar test set, and model performance would likely decline when evaluated on a more diverse, out-of-distribution test set50–52. Since our goal was to showcase how the benchmarks can be used for optimization, rather than to develop specific optimized models, we hope the reader will see the results in this light. A similar limitation applies to the HNSW graphs, as molecules in the training sets may also appear in the graphs and be evaluated during traversal. However, even in the smaller ~500 million Sigma2 library and with models trained on one million molecules, only 0.2% of the HNSW graph overlaps with the training set. While we do not expect this small overlap to significantly impact the results presented here, it remains a limitation of this study.

These cautions should not obscure the main observations from this study. The scale of docking and experimental results from recent virtual screening campaigns may support benchmarking sets for new ML models and methods. We provide the website lsd.docking.org to easily access these data and demonstrate their potential use in ML applications. As we continue to screen new targets at scale, we will upload new results to the website and welcome contributions from the community. It may particularly benefit from other larger library campaigns using other docking methods, which should be readily added to this resource.

Methods

Construction of the large-scale docking dataset (lsd.docking.org).

We identified directories on our computing cluster containing results from previously published large-scale docking campaigns run in our lab, with help from those who performed them. These directories included docking scores for all molecules in a campaign linked to ZINC IDs, which we used to query the ZINC database for the corresponding SMILES. We excluded a small fraction of ZINC IDs that no longer returned results due to updates of the ZINC database since the original campaigns. We combined the ZINC IDs, docking scores, and retrieved SMILES of all molecules in a campaign to produce the docking datasets available on lsd.docking.org. The directories also contained docking poses for molecules scoring better than a predefined threshold. We collected the docking poses for the top 500,000 scoring molecules without additional filtering and made them available on lsd.docking.org. For some campaigns, fewer than 500,000 poses were provided due to disk failures since the original campaign or due to fewer than 500,000 poses being saved during the original campaign. To complement these datasets, we uploaded the docking grids used to generate the poses and scores. Where available, we collected experimental results from the supplementary information of the corresponding publications. Otherwise, we located the experimental results with help from the lab member who performed the campaign and uploaded the results to lsd.docking.org (Table S1).

Training Chemprop to predict DOCK scores.

We evaluated the effects of three training set acquisition strategies and four training set sizes on the performance of ML models predicting DOCK scores from SMILES strings. The training set acquisition strategies included: random selection from the entire dataset, random selection from only the best 1% of molecules by DOCK score, and selecting 80% of the training data from the best 1% of molecules and the remaining 20% from the remainder of the dataset. We used training set sizes of 1,000, 10,000, 100,000, and 1,000,000.

We assessed model performance on test sets of 100,000,000 molecules using four metrics: the Pearson correlation between the predicted and actual DOCK scores across all test set molecules, the correlation only for the best 0.01% of test set molecules, the enrichment of the top 0.01% scoring molecules, measured by the log area under the curve (logAUC), which reflects recall as a function of ML ranking, and the enrichment of the actual new ligands discovered by experimental testing of top-ranked docking molecules. We applied these metrics to the docking results from three receptor campaigns: AmpC β-lactamase, 5HT2A, and Sigma2.

We randomly subsampled the AmpC, 5HT2A, and Sigma2 docking results from lsd.docking.org to create training sets. For each sampling strategy and training set size, we generated five independent training sets. To account for the validation set, we sampled an additional 20% of molecules. For instance, for a training set size of 10,000, we sampled an additional 2 000 molecules for the validation set, and for 100,000, we sampled an extra 20,000.

To minimize data leakage we split molecules based on their Bemis-Murcko scaffolds to assign them to the training and validation sets, ensuring no scaffold appeared in both sets. This approach produced training sets of approximately 1,000, 10,000, 100,000, and 1,000,000 molecules, with validation sets about 20% as large and containing no overlapping scaffolds.

We trained ML models with the Chemprop40 framework to predict DOCK scores from molecular SMILES strings. Chemprop, a PyTorch based framework, predicts molecular properties using message-passing neural networks on molecular graphs. We used the default parameters, featurizers, and settings. Specifically, we generated featurized molecular graphs from SMILES strings using the “SimpleMoleculeMolGraphFeaturizer,” messages were passed along directed bonds using “BondMessagePassing,” and the graph-level representation was obtained by averaging node-level representations with “MeanAggregation.” The learned representations had a dimensionality of 300, and a single-layer fully connected feed forward network predicted scores from the graph-level representation.

We trained the models to minimize the Root Mean Squared Error (RMSE) between predicted and true DOCK scores over 100 epochs. Training used the Adam optimizer with the Noam learning rate scheduler (initial, maximum, and final learning rates of 10^-4, 10^-3, and 10^-4, respectively). We used batch sizes of 64, batch normalization, and early stopping based on validation RMSE loss, with a patience of 5 epochs.

We gathered a test set for each receptor by randomly sampling 100,000,000 SMILES and DOCK scores for AmpC, 5HT2A, and Sigma2. Using the models trained with different sampling strategies and dataset sizes, we predicted DOCK scores for each molecule in the test set. After removing test set molecules included in a model’s training set, we calculated several metrics.

We calculated the Pearson correlation between predicted and actual DOCK scores for the entire test set and the top 0.01% scoring molecules in the test set. Next, we rank-ordered the ML predictions and plotted the fraction of the top 0.01% scoring molecules recovered against the ML ranking, using a log x-axis. We calculated the area under this curve (logAUC) to evaluate each model’s ability to enrich for the top 0.01% scoring molecules in its top predictions. This metric emphasizes practical utility in prospective screening, where only the top predictions are considered for visual inspection and testing.

Hierarchical Navigable Small World (HNSW) graph construction and traversal.

We demonstrated the utility of these billion-scale DOCK datasets by using them as benchmark sets for the chemical space exploration technique Retrieval Augmented Docking (RAD)41. RAD rapidly organizes chemical libraries into Hierarchical Navigable Small World (HNSW) graphs, which are multi-layered similarity-based structures built on the principles of small-world networks. The DOCK scoring function guides a best-first traversal of the HNSW graph, iteratively docking the neighbors of the best scoring nodes. Previous work found that RAD finds most of the top 0.01% scoring molecules while explicitly docking only a fraction of the library.

We constructed three HNSW graphs, one each for each receptor - AmpC, 5HT2A, and Sigma2 – using libraries from lsd.docking.org. The graphs contained 1,300 572,452, 1,300,808,533, and 468,639,652 molecules, respectively, randomly sampled from the corresponding docking results. We began by generating Morgan fingerprints with a radius of 2 and a length of 512 for all molecules. Using these fingerprints, we built the HNSW graphs with an M of 8, an efConstruction of 400, and the Tanimoto coefficient as the similarity metric. We then greedily traversed the HNSW graphs, as described in previous work41, using both the DOCK score and the predictions from the Chemprop models. We evaluated performance by measuring the fraction of the top 0.01% scoring molecules found while screening a RAD-prioritized 1% of the library.

Supplementary Material

Supplement 1

Table S1: Full breakdown of docking screens available on lsd.docking.org

Figure S1: Effects of increasing the training set size on 5HT2A Chemprop models

Figure S2: Effects of increasing the training set size on Sigma2 Chemprop models

Figure S3: Variability of Retrieval Augmented Docking with Chemprop models

Figure S4: Retrieval Augmented Docking of the AmpC library with Chemprop models

Figure S5: Retrieval Augmented Docking of the Sigma2 library with Chemprop models

media-1.pdf (1.5MB, pdf)

Acknowledgements

We thank E. Fink, S. Gahbauer, J. Lyu, M. Rachman, I. Singh, R. Stein, and F. Liu for assistance with assembling the data that went into lsd.docking.org. We thank M. Castanon and O. Mailhot for help with algorithms for collecting ZINC22 SMILES for hundreds of millions of molecules.

Funding

This work is supported by US NIH grant R35GM122481 (B.K.S.) and US NIH grant R01GM71896 (J.J.I.).

Abbreviations Used

GPCR

G Protein-Coupled Receptor

ML

Machine Learning

RAD

Retrieval Augmented Docking

HNSW

Hierarchical Navigable Small World

logAUC

Logarithmic Area Under the Curve

LSD

Large-Scale Docking

SMILES

Simplified Molecular Input Line Entry System

5HT2A

5-hydroxytryptamine (serotonin) receptor 2A

AmpC

AmpC β -lactamase

RMSE

Root Mean Squared Error

Funding Statement

This work is supported by US NIH grant R35GM122481 (B.K.S.) and US NIH grant R01GM71896 (J.J.I.).

Footnotes

Supporting information

Supporting Information is available free of charge on the ACS Publication website.

The authors declare the following competing financial interests: B.K.S. is co-founder of BlueDolphin LLC, Epiodyne, and Deep Apple Therapeutics, Inc., serves on the SRB of Genentech, the SABs of Schrödinger LLC and of Vilya Therapeutics. J.J.I. is a co-founder of BlueDolphin LLC and of Deep Apple Therapeutics. The authors declare no other competing interests.

Data and Software Availability

All docking results, docking poses, and experimental data are made freely available at lsd.docking.org. Open-source code for performing Retrieval Augmented Docking is available at https://github.com/keiserlab/rad. Open-source code for replicating the Chemprop model training, inference, and integration with RAD is available at https://github.com/bwhall61/lsd.

References

  • (1).Lyu J.; Wang S.; Balius T. E.; Singh I.; Levit A.; Moroz Y. S.; O’Meara M. J.; Che T.; Algaa E.; Tolmachova K.; et al. Ultra-large library docking for discovering new chemotypes. Nature 2019, 566 (7743), 224–229. DOI: 10.1038/s41586-019-0917-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (2).Warr W. A.; Nicklaus M. C.; Nicolaou C. A.; Rarey M. Exploration of Ultralarge Compound Collections for Drug Discovery. J Chem Inf Model 2022, 62 (9), 2021–2034. DOI: 10.1021/acs.jcim.2c00224. [DOI] [PubMed] [Google Scholar]
  • (3).Lyu J.; Irwin J. J.; Shoichet B. K. Modeling the expansion of virtual screening libraries. Nat Chem Biol 2023, 19 (6), 712–718. DOI: 10.1038/s41589-022-01234-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (4).Liu F.; Mailhot O.; Glenn I. S.; Vigneron S. F.; Bassim V.; Xu X.; Fonseca-Valencia K.; Smith M. S.; Radchenko D. S.; Fraser J. S.; et al. The impact of library size and scale of testing on virtual screening. Nat Chem Biol 2025. DOI: 10.1038/s41589-024-01797-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (5).Liu F.; Wu C. G.; Tu C. L.; Glenn I.; Meyerowitz J.; Kaplan A. L.; Lyu J.; Cheng Z.; Tarkhanova O. O.; Moroz Y. S.; et al. Large library docking identifies positive allosteric modulators of the calcium-sensing receptor. Science 2024, 385 (6715), eado1868. DOI: 10.1126/science.ado1868. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (6).Tummino T. A.; Iliopoulos-Tsoutsouvas C.; Braz J. M.; O’Brien E. S.; Stein R. M.; Craik V.; Tran N. K.; Ganapathy S.; Liu F.; Shiimura Y.; et al. Large library docking for cannabinoid-1 receptor agonists with reduced side efects. bioRxiv 2024. DOI: 10.1101/2023.02.27.530254. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (7).Singh I.; Seth A.; Billesbølle C. B.; Braz J.; Rodriguiz R. M.; Roy K.; Bekele B.; Craik V.; Huang X. P.; Boytsov D.; et al. Structure-based discovery of conformationally selective inhibitors of the serotonin transporter. Cell 2023, 186 (10), 2160–2175.e2117. DOI: 10.1016/j.cell.2023.04.010. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (8).Fink E. A.; Xu J.; Hübner H.; Braz J. M.; Seemann P.; Avet C.; Craik V.; Weikert D.; Schmidt M. F.; Webb C. M.; et al. Structure-based discovery of nonopioid analgesics acting through the α2A-adrenergic receptor. Science 2022, 377 (6614), eabn7065. DOI: 10.1126/science.abn7065. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (9).Lyu J.; Kapolka N.; Gumpper R.; Alon A.; Wang L.; Jain M. K.; Barros-Alvarez X.; Sakamoto K.; Kim Y.; DiBerto J.; et al. AlphaFold2 structures guide prospective ligand discovery. Science 2024, 384 (6702), eadn6354. DOI: 10.1126/science.adn6354. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (10).Gahbauer S.; DeLeon C.; Braz J. M.; Craik V.; Kang H. J.; Wan X.; Huang X. P.; Billesbølle C. B.; Liu Y.; Che T.; et al. Docking for EP4R antagonists active against inflammatory pain. Nat Commun 2023, 14 (1), 8067. DOI: 10.1038/s41467-023-43506-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (11).Gahbauer S.; Correy G. J.; Schuller M.; Ferla M. P.; Doruk Y. U.; Rachman M.; Wu T.; Diolaiti M.; Wang S.; Neitz R. J.; et al. Iterative computational design and crystallographic screening identifies potent inhibitors targeting the Nsp3 macrodomain of SARS-CoV-2. Proc Natl Acad Sci U S A 2023, 120 (2), e2212931120. DOI: 10.1073/pnas.2212931120. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (12).Fink E. A.; Bardine C.; Gahbauer S.; Singh I.; Detomasi T. C.; White K.; Gu S.; Wan X.; Chen J.; Ary B.; et al. Large library docking for novel SARS-CoV-2 main protease non-covalent and covalent inhibitors. Protein Sci 2023, 32 (8), e4712. DOI: 10.1002/pro.4712. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (13).Gorgulla C.; Padmanabha Das K. M.; Leigh K. E.; Cespugli M.; Fischer P. D.; Wang Z. F.; Tesseyre G.; Pandita S.; Shnapir A.; Calderaio A.; et al. A multi-pronged approach targeting SARS-CoV-2 proteins using ultra-large virtual screening. iScience 2021, 24 (2), 102021. DOI: 10.1016/j.isci.2020.102021. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (14).Sadybekov A. A.; Sadybekov A. V.; Liu Y.; Iliopoulos-Tsoutsouvas C.; Huang X. P.; Pickett J.; Houser B.; Patel N.; Tran N. K.; Tong F.; et al. Synthon-based ligand discovery in virtual libraries of over 11 billion compounds. Nature 2022, 601 (7893), 452–459. DOI: 10.1038/s41586-021-04220-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (15).Gorgulla C.; Nigam A.; Koop M.; Selim Çınaroğlu S.; Secker C.; Haddadnia M.; Kumar A.; Malets Y.; Hasson A.; Li M.; et al. VirtualFlow 2.0 - The Next Generation Drug Discovery Platform Enabling Adaptive Screens of 69 Billion Molecules. bioRxiv 2023, 2023.2004.2025.537981. DOI: 10.1101/2023.04.25.537981. [DOI] [Google Scholar]
  • (16).Zhu H.; Zhang Y.; Li W.; Huang N. A Comprehensive Survey of Prospective Structure-Based Virtual Screening for Early Drug Discovery in the Past Fifteen Years. Int J Mol Sci 2022, 23 (24). DOI: 10.3390/ijms232415961. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (17).Liu H.; Liang H.; Meng H.; Deng X.; Zhang X.; Lai L. A novel allosteric inhibitor that prevents IKKβ activation. Medchemcomm 2018, 9 (2), 239–243. DOI: 10.1039/c7md00599g. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (18).Ruan H.; Yu C.; Niu X.; Zhang W.; Liu H.; Chen L.; Xiong R.; Sun Q.; Jin C.; Liu Y.; et al. Computational strategy for intrinsically disordered protein ligand design leads to the discovery of p53 transactivation domain I binding compounds that activate the p53 pathway. Chem Sci 2020, 12 (8), 3004–3016. DOI: 10.1039/d0sc04670a. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (19).Wang Q.; Qi Y.; Yin N.; Lai L. Discovery of novel allosteric efectors based on the predicted allosteric sites for Escherichia coli D-3-phosphoglycerate dehydrogenase. PLoS One 2014, 9 (4), e94829. DOI: 10.1371/journal.pone.0094829. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (20).Gorgulla C.; Boeszoermenyi A.; Wang Z. F.; Fischer P. D.; Coote P. W.; Padmanabha Das K. M.; Malets Y. S.; Radchenko D. S.; Moroz Y. S.; Scott D. A.; et al. An open-source drug discovery platform enables ultra-large virtual screens. Nature 2020, 580 (7805), 663–668. DOI: 10.1038/s41586-020-2117-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (21).O’Mahony G.; Petersen J.; Ek M.; Rae R.; Johansson C.; Jianming L.; Prokoph N.; Bergström F.; Bamberg K.; Giordanetto F.; et al. Discovery by Virtual Screening of an Inhibitor of CDK5-Mediated PPARγ Phosphorylation. ACS Med Chem Lett 2022, 13 (4), 681–686. DOI: 10.1021/acsmedchemlett.1c00715. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (22).Schimunek J.; Seidl P.; Elez K.; Hempel T.; Le T.; Noé F.; Olsson S.; Raich L.; Winter R.; Gokcan H.; et al. A community efort in SARS-CoV-2 drug discovery. Mol Inform 2024, 43 (1), e202300262. DOI: 10.1002/minf.202300262. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (23).Pylkkö T.; Tomašič T.; Poso A.; Tammela P. Virtual Screening Assisted Search for Inhibitors of the Translocated Intimin Receptor of Enteropathogenic Escherichia Coli. Chembiochem 2024, 25 (2), e202300638. DOI: 10.1002/cbic.202300638. [DOI] [PubMed] [Google Scholar]
  • (24).Lack N. A.; Axerio-Cilies P.; Tavassoli P.; Han F. Q.; Chan K. H.; Feau C.; LeBlanc E.; Guns E. T.; Guy R. K.; Rennie P. S.; et al. Targeting the binding function 3 (BF3) site of the human androgen receptor through virtual screening. J Med Chem 2011, 54 (24), 8563–8573. DOI: 10.1021/jm201098n. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (25).Yang Y.; Yao K.; Repasky M. P.; Leswing K.; Abel R.; Shoichet B. K.; Jerome S. V. Eficient Exploration of Chemical Space with Docking and Deep Learning. J Chem Theory Comput 2021, 17 (11), 7106–7119. DOI: 10.1021/acs.jctc.1c00810. [DOI] [PubMed] [Google Scholar]
  • (26).Gentile F.; Agrawal V.; Hsing M.; Ton A. T.; Ban F.; Norinder U.; Gleave M. E.; Cherkasov A. Deep Docking: A Deep Learning Platform for Augmentation of Structure Based Drug Discovery. ACS Cent Sci 2020, 6 (6), 939–949. DOI: 10.1021/acscentsci.0c00229. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (27).Graff D. E.; Shakhnovich E. I.; Coley C. W. Accelerating high-throughput virtual screening through molecular pool-based active learning. Chem Sci 2021, 12 (22), 7866–7881. DOI: 10.1039/d0sc06805e. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (28).Graff D. E.; Aldeghi M.; Morrone J. A.; Jordan K. E.; Pyzer-Knapp E. O.; Coley C. W. Self-Focusing Virtual Screening with Active Design Space Pruning. J Chem Inf Model 2022, 62 (16), 3854–3862. DOI: 10.1021/acs.jcim.2c00554. [DOI] [PubMed] [Google Scholar]
  • (29).Sivula T.; Yetukuri L.; Kalliokoski T.; Käsnänen H.; Poso A.; Pöhner I. Machine Learning-Boosted Docking Enables the Eficient Structure-Based Virtual Screening of Giga-Scale Enumerated Chemical Libraries. J Chem Inf Model 2023, 63 (18), 5773–5783. DOI: 10.1021/acs.jcim.3c01239. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (30).Moesgaard L.; Kongsted J. Introducing SpaceGA: A Search Tool to Accelerate Large Virtual Screenings of Combinatorial Libraries. J Chem Inf Model 2024, 64 (21), 8123–8130. DOI: 10.1021/acs.jcim.4c01308. [DOI] [PubMed] [Google Scholar]
  • (31).Ballester P. J.; Mitchell J. B. A machine learning approach to predicting protein-ligand binding afinity with applications to molecular docking. Bioinformatics 2010, 26 (9), 1169–1175. DOI: 10.1093/bioinformatics/btq112. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (32).Wójcikowski M.; Ballester P. J.; Siedlecki P. Performance of machine-learning scoring functions in structure-based virtual screening. Sci Rep 2017, 7, 46710. DOI: 10.1038/srep46710. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (33).Stepniewska-Dziubinska M. M.; Zielenkiewicz P.; Siedlecki P. Development and evaluation of a deep learning model for protein-ligand binding afinity prediction. Bioinformatics 2018, 34 (21), 3666–3674. DOI: 10.1093/bioinformatics/bty374. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (34).Wójcikowski M.; Kukiełka M.; Stepniewska-Dziubinska M. M.; Siedlecki P. Development of a protein-ligand extended connectivity (PLEC) fingerprint and its application for binding afinity predictions. Bioinformatics 2019, 35 (8), 1334–1341. DOI: 10.1093/bioinformatics/bty757. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (35).Abbasi K.; Razzaghi P.; Poso A.; Amanlou M.; Ghasemi J. B.; Masoudi-Nejad A. DeepCDA: deep cross-domain compound-protein afinity prediction through LSTM and convolutional neural networks. Bioinformatics 2020, 36 (17), 4633–4642. DOI: 10.1093/bioinformatics/btaa544. [DOI] [PubMed] [Google Scholar]
  • (36).He T.; Heidemeyer M.; Ban F.; Cherkasov A.; Ester M. SimBoost: a read-across approach for predicting drug-target binding afinities using gradient boosting machines. J Cheminform 2017, 9 (1), 24. DOI: 10.1186/s13321-017-0209-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (37).Stein R. M.; Kang H. J.; McCorvy J. D.; Glatfelter G. C.; Jones A. J.; Che T.; Slocum S.; Huang X. P.; Savych O.; Moroz Y. S.; et al. Virtual discovery of melatonin receptor ligands to modulate circadian rhythms. Nature 2020, 579 (7800), 609–614. DOI: 10.1038/s41586-020-2027-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (38).Schuller M.; Correy G. J.; Gahbauer S.; Fearon D.; Wu T.; Díaz R. E.; Young I. D.; Carvalho Martins L.; Smith D. H.; Schulze-Gahmen U.; et al. Fragment binding to the Nsp3 macrodomain of SARS-CoV-2 identified through crystallographic screening and computational docking. Sci Adv 2021, 7 (16). DOI: 10.1126/sciadv.abf8711. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (39).Alon A.; Lyu J.; Braz J. M.; Tummino T. A.; Craik V.; O’Meara M. J.; Webb C. M.; Radchenko D. S.; Moroz Y. S.; Huang X. P.; et al. Structures of the σ2 receptor enable docking for bioactive ligand discovery. Nature 2021, 600 (7890), 759–764. DOI: 10.1038/s41586-021-04175-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (40).Heid E.; Greenman K. P.; Chung Y.; Li S. C.; Graf D. E.; Vermeire F. H.; Wu H.; Green W. H.; McGill C. J. Chemprop: A Machine Learning Package for Chemical Property Prediction. J Chem Inf Model 2024, 64 (1), 9–17. DOI: 10.1021/acs.jcim.3c01250. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (41).Hall B. W.; Keiser M. J. Retrieval Augmented Docking Using Hierarchical Navigable Small Worlds. J Chem Inf Model 2024, 64 (19), 7398–7408. DOI: 10.1021/acs.jcim.4c00683. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (42).Pettersen E. F.; Goddard T. D.; Huang C. C.; Couch G. S.; Greenblatt D. M.; Meng E. C.; Ferrin T. E. UCSF Chimera--a visualization system for exploratory research and analysis. J Comput Chem 2004, 25 (13), 1605–1612. DOI: 10.1002/jcc.20084. [DOI] [PubMed] [Google Scholar]
  • (43).Schrödinger, LLC. The PyMOL Molecular Graphics System, Version 3.0.
  • (44).Schrödinger, LLC. Schrödinger Release 2025–1 : Maestro. 2025.
  • (45).Yang K.; Swanson K.; Jin W.; Coley C.; Eiden P.; Gao H.; Guzman-Perez A.; Hopper T.; Kelley B.; Mathea M.; et al. Analyzing Learned Molecular Representations for Property Prediction. J Chem Inf Model 2019, 59 (8), 3370–3388. DOI: 10.1021/acs.jcim.9b00237. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (46).Mysinger M. M.; Shoichet B. K. Rapid context-dependent ligand desolvation in molecular docking. J Chem Inf Model 2010, 50 (9), 1561–1573. DOI: 10.1021/ci100214a. [DOI] [PubMed] [Google Scholar]
  • (47).Mysinger M. M.; Carchia M.; Irwin J. J.; Shoichet B. K. Directory of useful decoys, enhanced (DUD-E): better ligands and decoys for better benchmarking. J Med Chem 2012, 55 (14), 6582–6594. DOI: 10.1021/jm300687e. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (48).Stein R. M.; Yang Y.; Balius T. E.; O’Meara M. J.; Lyu J.; Young J.; Tang K.; Shoichet B. K.; Irwin J. J. Property-Unmatched Decoys in Docking Benchmarks. J Chem Inf Model 2021, 61 (2), 699–714. DOI: 10.1021/acs.jcim.0c00598. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (49).Malkov Y. A.; Yashunin D. A. Eficient and Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs. IEEE Trans Pattern Anal Mach Intell 2020, 42 (4), 824–836. DOI: 10.1109/TPAMI.2018.2889473. [DOI] [PubMed] [Google Scholar]
  • (50).Tossou P.; Wognum C.; Craig M.; Mary H.; Noutahi E. Real-World Molecular Out-Of-Distribution: Specification and Investigation. J Chem Inf Model 2024, 64 (3), 697–711. DOI: 10.1021/acs.jcim.3c01774. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • (51).Volkov M.; Turk J. A.; Drizard N.; Martin N.; Hofmann B.; Gaston-Mathé Y.; Rognan D. On the Frustration to Predict Binding Afinities from Protein-Ligand Structures with Deep Neural Networks. J Med Chem 2022, 65 (11), 7946–7958. DOI: 10.1021/acs.jmedchem.2c00487. [DOI] [PubMed] [Google Scholar]
  • (52).van Tilborg D.; Alenicheva A.; Grisoni F. Exposing the Limitations of Molecular Machine Learning with Activity Clifs. J Chem Inf Model 2022, 62 (23), 5938–5951. DOI: 10.1021/acs.jcim.2c01073. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplement 1

Table S1: Full breakdown of docking screens available on lsd.docking.org

Figure S1: Effects of increasing the training set size on 5HT2A Chemprop models

Figure S2: Effects of increasing the training set size on Sigma2 Chemprop models

Figure S3: Variability of Retrieval Augmented Docking with Chemprop models

Figure S4: Retrieval Augmented Docking of the AmpC library with Chemprop models

Figure S5: Retrieval Augmented Docking of the Sigma2 library with Chemprop models

media-1.pdf (1.5MB, pdf)

Data Availability Statement

All docking results, docking poses, and experimental data are made freely available at lsd.docking.org. Open-source code for performing Retrieval Augmented Docking is available at https://github.com/keiserlab/rad. Open-source code for replicating the Chemprop model training, inference, and integration with RAD is available at https://github.com/bwhall61/lsd.

Articles from bioRxiv are provided here courtesy of Cold Spring Harbor Laboratory Preprints

RESOURCES