Prediction of HIV sensitivity to monoclonal antibodies using aminoacid sequences and deep learning

Knowing the sensitivity of a viral strain versus a monoclonal antibody is of interest for HIV vaccine development and therapy. The HIV strains vary in their resistance to antibodies, and the accurate prediction of virus-antibody sensitivity can be used to find potent antibody combinations that broadly neutralize multiple and diverse HIV strains. Sensitivity prediction can be combined with other methods such as generative algorithms to design novel antibodies in silico or with feature selection to uncover the sites of interest in the sequence. However, these tools are limited in the absence of in silico accurate prediction methods.

Our method leverages the CATNAP dataset, probably the most comprehensive collection of HIV-antibodies assays, and predicts the antibody-virus sensitivity in the form of binary classification. The methods proposed by others focus primarily on analyzing the virus sequences. However, our article demonstrates the advantages gained by modeling the antibody-virus sensitivity as a function of both virus and antibody sequences. The input is formed by the virus envelope and the antibody variable region aminoacid sequences. No structural features are required, which makes our system very practical, given that sequence data is more common than structures. We compare with two other state-of-the-art methods that leverage the same dataset and use sequence data only. Our approach, based on neuronal networks and transfer learning, measures increased predictive performance as measured on a set of 31 specific broadly neutralizing antibodies.

https://github.com/vlad-danaila/deep_hiv_ab_pred/tree/fc-att-fix.