HitScreen: A Sequence-Based Drug Virtual Screening Approach Using Data Augmentation and Protein Language Models.

Area of Science:

• Computational biology
• Drug discovery
• Bioinformatics

Background:

• Sequence-based drug-target interaction (DTI) prediction is vital for identifying drug candidates without 3D protein structures.
• Existing sequence-based DTI methods struggle with generalization and capturing spatial interactions, leading to a performance gap with structure-based approaches.

Purpose of the Study:

• To develop a robust deep learning framework, HitScreen, for sequence-based DTI prediction in virtual screening.
• To bridge the performance gap between sequence-based and structure-based DTI prediction methods.

Main Methods:

• HitScreen employs a conditional label inversion strategy to mitigate data biases.
• It integrates multiple pretrained protein language models (Ankh, ESM-2, ProtT5) and Uni-Mol for spatial information encoding.
• A cross-attention mechanism captures local intermolecular interactions between drug molecules and protein sequences.

Main Results:

• HitScreen demonstrates performance comparable to or exceeding state-of-the-art structure-based methods on independent datasets (DEKOIS2.0, DUD-E).
• The framework achieves this using only protein sequence information, outperforming previous sequence-based methods.
• Interpretability analyses confirm the model's ability to identify biologically relevant interactions.

Conclusions:

• HitScreen offers a robust, interpretable, and broadly applicable framework for sequence-based DTI prediction.
• The method shows significant potential for enhancing drug virtual screening and providing insights for rational drug design.