ProtAff: Protein Binding Affinity Prediction via LoRA-Finetuned ESM-2.

Area of Science:

• Computational biology
• Structural biology
• Bioinformatics

Background:

• Predicting protein-protein interaction (PPI) binding affinity is crucial but challenging.
• Current structure prediction models like AlphaFold3 (AF3) and Boltz-2 provide good poses but their confidence scores do not correlate with binding affinity.
• Existing machine learning predictors also struggle with accurate affinity ranking.

Purpose of the Study:

• To develop a novel, sequence-only computational model for predicting protein-protein interaction binding affinity.
• To evaluate the performance of the new model against state-of-the-art structure-based and machine learning methods.

Main Methods:

• Developed ProtAff, a sequence-only model using ESM-2 (650M parameters) with low-rank adaptation (LoRA) and a cross-attention module.
• Trained ProtAff on 362,567 affinity measurements from 20 diverse data sources using a margin ranking loss.
• Ensured robust evaluation by removing training samples with >50% sequence similarity to test targets (e.g., EGFR).

Main Results:

• ProtAff achieved a Spearman correlation coefficient (ρ) of 0.413 on the AdaptyvBio EGFR benchmark, significantly outperforming AF3 (ρ=0.054), Boltz-2 (ρ=-0.046), MINT (ρ=0.242), and CrossAffinity (ρ=0.216).
• In a binder design competition for Nipah virus, ProtAff-guided design yielded a binder with K_D=0.132 nM, a 2.8-fold improvement over the winner.
• ProtAff showed limitations in cross-target discrimination, underperforming structural methods for distinguishing cognate from non-cognate pairings.

Conclusions:

• Sequence-based models like ProtAff are highly effective for ranking binding affinity within a specific target.
• While ProtAff excels at affinity prediction, structural methods remain superior for assessing cross-target specificity.
• ProtAff represents a significant advancement in computational prediction of protein-protein interaction binding affinity, with potential applications in drug discovery and protein engineering.