Paul Francoeur

Knowledge of the bound protein-ligand structure is critical to many drug discovery tasks. One tool for in silico bound structure elucidation is molecular docking, which samples and scores ligand binding conformations. Recent work has demonstrated that convolutional neural networks (CNNs) for protein-ligand pose scoring outperform conventional scoring functions. Scoring performance can be further increased by taking the average of multiple CNN models, termed ensembles. However, ensembles of large parameter models require significant computational resources and therefore are difficult to apply to high-throughput molecular docking for virtual screening. We investigate knowledge distillation as a framework to condense the knowledge of large, powerful CNN model ensembles into a single reduced CNN model for a significant reduction in computational cost. Ensemble KD produces single models that outperform non-KD trained single models.

Deep learning methods that predict protein-ligand binding have recently been used for structure-based virtual screening. Many such models have been trained using protein-ligand complexes with known crystal structures and activities from the PDBBind data set. However, because PDBbind only includes 20K complexes, models typically fail to generalize to new targets, and model performance is on par with models trained with only ligand information. Conversely, the ChEMBL database contains a wealth of chemical activity information but includes no information about binding poses. We introduce BigBind, a data set that maps ChEMBL activity data to proteins from the CrossDocked data set. BigBind comprises 583 K ligand activities and includes 3D structures of the protein binding pockets. Additionally, we augmented the data by adding an equal number of putative inactives for each target. Using this data, we developed Banana (basic neural network for binding affinity), a neural network-based model to classify active from inactive compounds, defined by a 10 μM cutoff. Our model achieved an AUC of 0.72 on BigBind's test set, while a ligand-only model achieved an AUC of 0.59. Furthermore, Banana achieved competitive performance on the LIT-PCBA benchmark (median EF1% 1.81) while running 16,000 times faster than molecular docking with Gnina. We suggest that Banana, as well as other models trained on this data set, will significantly improve the outcomes of prospective virtual screening tasks.

Deep learning methods that predict protein-ligand binding have recently been used for structure-based virtual screening. Many such models have been trained using protein-ligand complexes with known crystal structures and activities from the PDBBind dataset. However, because PDBbind only includes 20K complexes, models typically fail to generalize to new targets, and model performance is on par with models trained with only ligand information. Conversely, the ChEMBL database contains a wealth of chemical activity information but includes no information about binding poses. We introduce BigBind, a dataset that maps ChEMBL activity data to proteins from the CrossDocked dataset. BigBind comprises 583K ligand activities and includes 3D structures of the protein binding pockets. Additionally, we augmented the data by adding an equal number of putative inactives for each target. Using this data, we developed BANANA (BAsic NeurAl Network for binding Affinity), a neural network-based model to classify active from inactive compounds, defined by a 10 μM cutoff. Our model achieved an AUC of 0.72 on BigBind’s test set, while a ligand-only model achieved an AUC of 0.59. Furthermore, BANANA achieved competitive performance on the LIT-PCBA benchmark (median EF1\% 1.81) while running 16,000 times faster than molecular docking with GNINA. We suggest that BANANA, as well as other models trained on this dataset, will significantly improve the outcomes of prospective virtual screening tasks.

The success of machine learning is, in part, due to a large volume of data available to train models. However, the amount of training data for structure-based molecular property prediction remains limited. The previously described CrossDocked2020 data set expanded the available training data for binding pose classification in a molecular docking setting but did not address expanding the amount of receptor-ligand binding affinity data. We present experiments demonstrating that imputing binding affinity labels for complexes without experimentally determined binding affinities is a viable approach to expanding training data for structure-based models of receptor-ligand binding affinity. In particular, we demonstrate that utilizing imputed labels from a convolutional neural network trained only on the affinity data present in CrossDocked2020 results in a small improvement in the binding affinity regression performance, despite the additional sources of noise that such imputed labels add to the training data. The code, data splits, and imputation labels utilized in this paper are freely available at https://github.com/francoep/ImputationPaper.

The immense size of chemical space, the relative scarcity of high quality data, and the cost of running experiments to accurately measure molecular properties makes active learning (AL) an attractive approach to efficiently explore the space and train high-quality models for molecular property prediction. While AL is traditionally successful at classification, there have been recent advances in using AL for regression tasks. Recently, regressing to a normal inverse gamma distribution has been shown to be effective at predicting molecular properties in the QM9 dataset. However, we present a series of experiments demonstrating that various state of the art AL regression techniques are indistinguishable from random selection for small molecule pKa prediction. Source code for this paper is available at https://github.com/francoep/pKa_activelearning.

While accurate prediction of aqueous solubility remains a challenge in drug discovery, machine learning (ML) approaches have become increasingly popular for this task. For instance, in the Second Challenge to Predict Aqueous Solubility (SC2), all groups utilized machine learning methods in their submissions. We present SolTranNet, a molecule attention transformer to predict aqueous solubility from a molecule’s SMILES representation. Atypically, we demonstrate that larger models perform worse at this task, with SolTranNet’s final architecture having 3,393 parameters while outperforming linear ML approaches. SolTranNet has a 3-fold scaffold split cross-validation root-mean-square error (RMSE) of 1.459 on AqSolDB and an RMSE of 1.711 on a withheld test set. We also demonstrate that, when used as a classifier to filter out insoluble compounds, SolTranNet achieves a sensitivity of 94.8% on the SC2 data set and is competitive with the other methods submitted to the competition. SolTranNet is distributed via pip, and its source code is available at https://github.com/gnina/SolTranNet.

One of the main challenges in drug discovery is predicting protein-ligand binding affinity. Recently, machine learning approaches have made substantial progress on this task. However, current methods of model evaluation are overly optimistic in measuring generalization to new targets, and there does not exist a standard dataset of sufficient size to compare performance between models. We present a new dataset for structure-based machine learning, the CrossDocked2020 set, with 22.5 million poses of ligands docked into multiple similar binding pockets across the Protein Data Bank and perform a comprehensive evaluation of grid-based convolutional neural network models on this dataset. We also demonstrate how the partitioning of the training data and test data can impact the results of models trained with the PDBbind dataset, how performance improves by adding more, lower-quality training data, and how training with docked poses imparts pose sensitivity to the predicted affinity of a complex. Our best performing model, an ensemble of 5 densely connected convolutional newtworks, achieves a root mean squared error of 1.42 and Pearson R of 0.612 on the affinity prediction task, an AUC of 0.956 at binding pose classification, and a 68.4% accuracy at pose selection on the CrossDocked2020 set. By providing data splits for clustered cross-validation and the raw data for the CrossDocked2020 set, we establish the first standardized dataset for training machine learning models to recognize ligands in non-cognate target structures while also greatly expanding the number of poses available for training. In order to facilitate community adoption of this dataset for benchmarking protein-ligand binding affinity prediction, we provide our models, weights, and the CrossDocked2020 set at https://github.com/gnina/models.

One of the main challenges in drug discovery is predicting protein-ligand binding affinity. Recently, machine learning approaches have made substantial progress on this task. However, current methods of model evaluation are overly optimistic in measuring generalization to new targets, and there does not exist a standard dataset of sufficient size to compare performance between models. We present a new dataset for structure-based machine learning, the CrossDocked2020 set, with 22.5 million poses of ligands docked into multiple similar binding pockets across the Protein Data Bank and perform a comprehensive evaluation of grid-based convolutional neural network models on this dataset. We also demonstrate how the partitioning of the training data and test data can impact the results of models trained with the PDBbind dataset, how performance improves by adding more, lower-quality training data, and how training with docked poses imparts pose sensitivity to the predicted affinity of a complex. Our best performing model, an ensemble of 5 densely connected convolutional newtworks, achieves a root mean squared error of 1.42 and Pearson R of 0.612 on the affinity prediction task, an AUC of 0.956 at binding pose classification, and a 68.4% accuracy at pose selection on the CrossDocked2020 set. By providing data splits for clustered cross-validation and the raw data for the CrossDocked2020 set, we establish the first standardized dataset for training machine learning models to recognize ligands in non-cognate target structures while also greatly expanding the number of poses available for training. In order to facilitate community adoption of this dataset for benchmarking protein-ligand binding affinity prediction, we provide our models, weights, and the CrossDocked2020 set at https://github.com/gnina/models.