The Cross-Channel Generalisation Gap (CCGG) Benchmark for Protein Language Models

The Cross-Channel Generalization Gap (CCGG) is a diagnostic framework designed to quantify assay-conditioned geometric bias in Protein Language Models (PLMs). It evaluates whether zero-shot representations from models like ESM-2 and ProtBERT encode universal biological interaction physics, or if they silently overfit to the statistical artifacts and selection biases of specific high-throughput laboratory assays (AP-MS, Y2H, and BioID).