LigityScore: A CNN-Based Method for Binding Affinity Predictions

Abstract

Scoring functions are the heart of structure based drug design, where they are used to estimate how strongly the docked pose of a ligand binds to the target. Seeking a scoring function that can accurately predict the binding affinity is key for successful virtual screening methods. Deep learning approaches have recently seen a rise in popularity as a means to improve the scoring function having the key advantage to automatically extract features and create a complex representation of the data without feature engineering and expert knowledge.

In this study we present the LigityScore scoring functions – LigityScore1D and LigityScore3D. The LigityScore models are rotationally invariant scoring functions based on convolutional neural networks (CNN). LigityScore descriptors are extracted directly from the structural and interacting properties of the protein-ligand complex which are input to a CNN for automatic feature extraction and binding affinity prediction. This representation uses the spatial distribution of Pharmacophoric Interaction Points (PIPs), derived from interaction features from the protein-ligand complex based on pharmacophoric features conformant to specific family types (Hydrogen Bond Acceptor (HBA), Hydrogen Bond Donor (HBD), etc.) and distance thresholds. The data representation component and the CNN architecture together, constitute the LigityScore scoring function.

LigityScore1D considers a single distance from each combination of two PIPs from the extracted ligand and protein PIP pools, and generates a feature vector for each pharmacophoric family pair (example HBA-HBD) based on the distribution of the PIP pair distances in discretised space. The different pharmacophoric family pair combinations are grouped to construct a matrix representation of the complex. Similarly, LigityScore3D considers 3-PIP combinations to create a triangular structure with three distances between the PIPs. These are discretised to represent binning coordinates for a 3D feature cube for each pharmacophoric family set (example HBA-HBD-HBD) that describe the distribution of distances in 3D space. The cube from each pharmacophoric family set are grouped together to from a feature cube collection.

The main contribution for this study is to present a novel rotationally invariant protein-ligand representation for use as a CNN based scoring function for binding affinity prediction. The CNN model is used for automatic feature extraction from the LigityScore representations. LigityScore models are evaluated for scoring power on the latest two CASF benchmarks. The Pearson correlation coefficient, and the standard deviation in linear regression were used to compare LigityScore with the benchmark model, and also other models in literature published in recent years. LigityScore3D has achieved better results than LigityScore1D and showed similar performance in both CASF benchmarks. LigityScore3D ranked 5$^{\textrm{th}}$ place in the CASF-2013 benchmark, and 8$^{\textrm{th}}$ in CASF-2016, with an average Pearson correlation coefficient (R) performance of 0.713 and 0.725 respectively. LigityScore1D obtained best results when trained using the PBDbind v2018 dataset, and ranked 8$^{\textrm{th}}$ place in the CASF-2013 and 7$^{\textrm{th}}$ place in CASF-2016 with an R performance of 0.635 and 0.741 respectively. Our methods show relatively good performance that exceed the Pafnucy performance, as one of the best performing CNN based scoring function, on the CASF-2013 benchmark, using a less computationally complex model that can be trained 16 times faster.