Selective representation learning boosts intrinsically disordered protein prediction from pretrained language models

Background Intrinsically disordered proteins and intrinsically disordered regions lack stable tertiary structures yet play indispensable roles in cellular signaling, regulation, and disease processes. Predicting disorder from sequence alone remains a fundamental challenge in computational biology. Recent advances in pretrained protein language models have shown great potential in capturing rich contextual representations. However, hierarchical pretrained protein language model features suffer from token-level noise and multi-layer redundancy, which limits their efficacy in downstream tasks. Methods In this work, we present SRLIDP, a selective representation learning framework designed to extract disorder-relevant signals from PPLMs. SRLIDP employs a dual-granularity selection strategy: at the token level, a Depth-Adaptive Retention Schedule (DARS) with Gumbel-Sigmoid differentiable relaxation filters uninformative residue noise; at the layer level, a dynamic cross-layer attention mechanism aggregates contextualized representations, followed by a bidirectional GRU module to restore global sequence context. Furthermore, to minimize potential homology-related data leakage, a strict multi-stage homology filtering pipeline (<25% sequence identity) was established across all evaluation benchmarks (Clean MXD494, SL329, DISORDER723, CAID3_NOX, and CAID3_PDB). Results Extensive evaluations on these strictly filtered benchmarks indicate that SRLIDP exhibits highly competitive performance compared to several representative baseline predictors, achieving robust cross-dataset generalization while maintaining high computational efficiency. Conclusion Our findings establish SRLIDP as a robust, rigorously homology-filtered, and effective solution for protein disorder prediction, with broad implications for functional analysis and therapeutic discovery.