BindScreen: Protein-Centric Contrastive Learning for Sequence-Based Virtual Screening

Virtual screening involves assessing candidate molecules against a protein target. Sequence-based deep learning techniques for virtual screening bypass the structural requirements of docking models. However, pair-based models have an advantage over these techniques because they require only a single forward pass per protein-molecule pair, which makes them scalable to large libraries.

Contrastive models remove the need for multiple forward passes, but they assume a symmetric, one-to-one correspondence between protein and molecule pairs. This assumption is not valid for protein-molecule binding, which is an asymmetric, many-to-many relationship. The authors introduce BindScreen, a sequence-only dual-encoder screening model, which addresses this issue.

The key design decision in BindScreen is not the contrastive loss, but how the batch is built. The model combines a protein-centric batch construction method with an asymmetric multi-positive InfoNCE loss. The authors perform a factorial ablation study to determine the individual contributions of the loss and the batch construction method.

They find that the loss alone degrades performance when using standard CLIP batching, while the protein-centric batch construction method alone recovers most of the performance gains. The combination of the two approaches yields the best results. Importantly, the effect of BindScreen's design is encoder-agnostic, meaning it works across eight protein language models, which span four architectural families.

By decoupling the number of proteins from the number of molecules in each batch, BindScreen achieves higher validation BEDROC scores than standard CLIP models in significantly less time. Specifically, BindScreen reaches higher validation BEDROC in 86 hours, compared to standard CLIP models that require 460 hours, and needs about seven times fewer forward passes to screen the LIT-PCBA dataset than pair-based models.

The authors make the source code, pretrained checkpoints, and datasets publicly available at https://github.com/pcdslab/BindScreen and https://huggingface.co/collections/SaeedLab/bindscreen.