AttentionScore: A Target-Specific, Bias-Aware Scoring Function for Structure-Based Virtual Screening: A Case Study on...

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning in drug discovery

Background:

• Generic scoring functions in virtual screening can be biased and prone to errors.
• Target-specific scoring functions are needed to improve the accuracy and reliability of structure-based virtual screening.
• METTL3 is a key target in various diseases, making its inhibitors a focus for drug discovery.

Purpose of the Study:

• To introduce AttentionScore, a deep learning-based scoring function for METTL3.
• To integrate ligand-only and protein-ligand interaction information for improved screening.
• To provide a bias-aware evaluation framework for virtual screening methods.

Main Methods:

• Developed an end-to-end deep learning architecture (AttentionScore) using multihead attention encoders and joint latent representations.
• Integrated protein-ligand interaction (PLEC) fingerprints with ligand chemotype fingerprints (Avalon/ECFP4).
• Constructed a similarity-constrained (SC) split and an extrapolative test set (Set 2) for rigorous, bias-aware evaluation.

Main Results:

• AttentionScore achieved high performance on the SC test set (Set 1) with PR-AUC = 0.9609.
• The model demonstrated robust performance on the stricter Set 2, outperforming generic scoring functions and machine-learning baselines.
• Statistical analyses confirmed the robustness and reliability of the observed performance gains.

Conclusions:

• AttentionScore offers a significant advancement in target-specific virtual screening for METTL3.
• The developed bias-aware evaluation framework minimizes analogue leakage and ensures reliable performance assessment.
• Publicly available data, code, and a user-friendly interface promote transparency and accessibility for researchers.