TopoScorer: a light, interpretable predictor for protein-protein binding affinity

ICLR 2026 Conference Submission16194 Authors

Keywords: Topological Deep Learning; Protein Design; Binding Affinity;

Abstract: Protein-protein binding affinity underlies complex stability, selectivity, and therapeutic action, yet experimental measurement is low-throughput and existing deep learning based approaches lack interpretability and a differentiable path from affinity back to the interface. We present TopoScorer, a lightweight, interpretable, end-to-end–trainable affinity scorer that can act as a loss or reward to steer generative and discriminative protein models; across protein and mutation affinity benchmarks, it delivers performance comparable to state-of-the-art methods and, when integrated into a modern antibody-design workflow, improves affinity-related metrics of generated candidates. The core component of TopoScorer is Specter(Spectral Topology Encoder), a topology-driven, multi-channel, multi-scale differentiable feature extractor for protein–protein interfaces that converts full-atom coordinates into topo-spectral representations via Persistent Topological Hyperdigraph Laplacians (PTHLs) and differentiable spectral descriptors, preserving physicochemical-role–aware cues alongside 3D topological structure to yield compact, interpretable features suitable for learning.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 16194