Memo-Patho: Bridging Local-Global Transmembrane Protein Contexts with Contrastive Pretraining for Alignment-Free Pathogenicity Prediction

Understanding the pathogenic impact of protein mutations remains a fundamental challenge in genomic medicine, particularly for transmembrane proteins (TMPs), a functionally critical yet structurally under-annotated class that includes numerous drug targets. Current variant effect predictors often fall short in this domain due to computational burdens, reliance on multiple sequence alignments (MSAs), and limited ability to generalize to TMP-specific constraints. Here we present Memo-Patho, an alignment-free deep learning framework specifically optimized for TMP mutation pathogenicity prediction. Memo-Patho integrates global and local representations from protein language models (PLMs) and predicted structural features. Furthermore, it introduces a contrastive pre-training strategy that learns discriminative features by comparing benign and pathogenic mutations within the same protein backbone. This design enables the model to capture functional disruptions in challenging contexts without structural or evolutionary alignments. Memo-Patho outperforms state-of-the-art predictors on two types of TMP benchmark datasets (accuracy up to 0.93) and unseen-protein generalization tasks, showing strong agreement with evolutionary conservation profiles. Its predictions are further validated on a manually curated KCNQ1 dataset of 55 ion channel variants, achieving an accuracy of 0.84 and outperforming existing tools. The model also supports high-throughput scanning, efficiently analyzing over 15000 variants in under 80 minutes. Our results establish Memo-Patho as a fast, generalizable, and biologically grounded approach for variant effect prediction in TMPs. Its alignment-free design, contrastive learning paradigm, and clinical robustness mark a step forward towards scalable interpretation of TMP variation for both basic research and precision medicine applications.