Structure-Agnostic Protein-Ligand Binding Affinity Prediction via Hierarchical Representation Alignment

Area of Science:

• Computational Biology
• Drug Discovery
• Structural Bioinformatics

Background:

• Real-world protein-ligand affinity prediction requires models robust to variable structural data.
• Out-of-distribution generalization is crucial for reliable predictions in diverse scenarios.

Purpose of the Study:

• To develop a novel framework for robust and generalizable protein-ligand affinity prediction.
• To address challenges posed by intra- and inter-molecular heterogeneity in structural data.

Main Methods:

• Introduced AlignNet, a hierarchical representation alignment framework.
• Employed intra-molecular module for feature projection and fusion.
• Utilized inter-molecular module for cross-molecular embedding alignment.

Main Results:

• Achieved competitive performance with up to 20.4% gain in SCC on challenging datasets.
• Demonstrated improved out-of-distribution generalization using sequence-only inputs.
• Learned well-separated, affinity-related clusters supporting structure-independent prediction.

Conclusions:

• AlignNet provides robust protein-ligand embeddings for generalizable affinity prediction.
• The framework enables reliable predictions even from sequence-level data, mitigating structural variability.
• AlignNet advances structure-agnostic approaches in computational drug discovery.