Docking is a classic computational biology problem that is widely used to predict binding conformations and to virtually screen binding molecules. Recently, a deep learning-based method, DeepDock was proposed to address the docking problem. The method shows great performance on conformation prediction. One major limitation of the method is that the interaction of ligands and targets is too simple. Here, we introduce caDeepDock, which is a geometric deep learning model for protein-ligand binding conformation prediction. Inspired by DeepDock, caDeepDock has two major advantages over DeepDock. First, cross-attention is employed to enable communications between the molecule and the protein binding pocket. Second, a positional embedding based on node degrees is used to incorporate both data-dependent and position-dependent communications. Experiments on the CASF-2016 benchmark have shown that the potential function learned by caDeepDock is able to pick a near-native conformation (with RMSD$\leqslant$2Å) from a set of decoy conformations with a success rate of 91.6%. This result outperforms not only DeepDock by 4.6% but also AutoDock by 1.4%. Consequently, the potential function learned by caDeepDock yields 5.96% more near-native conformations in conformation optimization experiments.

Predicting binding affinities of peptide antigens presented on major histocompatibility complex (MHC) is of great importance in T-cell immune response research. Accurate prediction of peptide-MHC binding affinities is essential for vaccine design and disease treatment. Recent deep learning-based prediction methods have shown that effective sequence embedding is critical to accurately predict binding affinities. One common neural network layer shared by these methods is the global average pooling layer that aggregates features. However, can we design a better global pooling layer? Here, we introduce a novel cross attention pooling (caPool) layer to aggregate features. As our initial application of caPool, a novel end-to-end transformer model, called capTransformer, is proposed for peptide-MHC class I binding prediction. In our model, caPool jointly aligns peptide-MHC residual pairs and aggregates residual features. Thus, instead of treating all residues equally and independently, caPool focuses more on correlated residue pairs that are potentially contact pairs contributing major forces to stabilize the complex structure. Using a five-fold cross-validation experiment, we found that caPool achieved the highest PCC value of 0.845, which was 0.139 higher than a global average pooling. Here, the global pooling layer was the only difference between the two tested models, and this observation indicated that global average pooling was not always the best choice. Importantly, our capTransformer model achieved a SRCC value of 0.614 (i.e., 6.4% higher than the best-performing method) when applied to the IEDB dataset.

Predicting Protein-ligand binding affinities using Deep Learning can significantly shorten the drug development cycle. Recently, Graph neural network models have been developed, and are successfully used to accelerate the development of potential drugs. One major limitation of these GNN models is that they focus on node features (i.e., atom features), as these nodes carry the most important information of molecules. However, atoms are connected via different bonds in molecules, and we argue that such chemical bonds carry critical information for assessing how atomic features should be aggregated. To overcome the lack of bond-related information in earlier models, we here proposed a novel edge-gated graph neural network (egGNN) that predict the binding affinities between proteins and ligands. Specifically, our model treats chemical bonds as gates that control how information is extracted between atoms, this modification enables our model to extract more accurate information for different bonds. We tested our model using the CASF-2016 dataset, and found that the Pearson’s correlation coefficient (R) of egGNN is three percent higher than that of the best tested method (i.e., 0.86 vs 0.83), and the Root Mean Square Error (RMSE) is significantly lower than that of the best tested method (i.e., 1.12 vs 1.23) when compared to state-of-the-art-approaches. In ablation experiments, we demonstrate that our edge-gated feature extraction (EGFE) consoderably improves the performance of GNNs. These results indicate that egGNN represents a promising tool applicable for virtual screening, and should greatly assist in accelerating drug development.

Computational algorithms are being successfully used to speed up drug development processes, primarily by way of turning biochemical problems into data problems. Recently, with increasing amounts of available biological data generated by biochemical methods measured affinity, some computational algorithms based deep learning for predicting drug-target affinity (DTA) become promising directions for accelerating the process of drug development. These deep learning models first attempts focus on representation learning of individual amino acids or atoms. Next global average pooling layers in those models are used to to combine such individual features to global features, and finally simple feed forward networks are adopted to yield affinity predictions. Notably, research which has been undertaken on global feature aggregations (e.g., the global pooling and the feed forward layers) for the drug-target affinity problem is still lacked currently. To address this issue, we propose a new rzMLP block featured newly designed global feature aggregations. This rzMLP block is based on two recent technologies in deep learning research: the gMLP model and the ReZero layer. We use gMLP model to aggregate input features with a constant size, while the ReZero layer is used to smooth the training process of this block. Our rzMLP is capable of learning complicated global features while overcoming the problems caused by the model being too deep. Importantly, when we compared a model contained rzMLP block to others, the mean squared error(MSE) decreases by 33%. Comparing to state-of-the-art methods for predicting affinity, rzMLP-DTA achieves the lowest MSE and highest CI on two benchmarks, Davis and KIBA datasets, respectively.

Discovery of new drugs heavily relies on predicting the binding affinities of drug molecules to suitable drug targets. To accelerate this process of identifying accurate affinities, computer-aided methods need to be applied in drug discovery pipeline. While various computational methods have been developed in the past ten years, the most successful methods to date use 3D convolutional neural networks (3D-CNNs). These 3D-CNN networks are based on deep learning models, and are both faster and more accurate than machine learning methods. However, currently used CNN is difficult to learn global and spatial features, while we hypothesis that spacial features should be critical for structure-based binding affinity predictions. Here we propose an end-to-end 3D-CNN with spatial attention mechanisms, called saCNN, to encourage spatial feature learning. When visualizing the learned spacial attentions in our experiments, it can be observed that saCNN model focuses more on the voxels near interaction centers. This key observation well supports our hypothesis that spacial features are critical for binding affinity predictions. In additions, we show that our model improves the Root Mean Square Error (RMSE) of the predicted binding affinities by 11.5% (with an absolute value of 1.117) and the Pearson Correlation Coefficient (R) by 3.2% (with an absolute value of 0.865) compared to currently used models on the PDBbind v.2016 core set. Importantly, the generalization abilities of our model is further demonstrated on CASF-2013 and CASF-2007 datasets.