Bridging local-global transmembrane protein contexts with contrastive pretraining for alignment-free pathogenicity...

Area of Science:

• Computational biology
• Genomics
• Biophysics

Background:

• Predicting pathogenicity of protein mutations is crucial for precision medicine.
• Transmembrane proteins (TMPs) are vital drug targets, but predicting their mutation effects is challenging.
• Existing computational methods struggle with TMP-specific biophysical constraints and often require evolutionary or structural data.

Purpose of the Study:

• To develop a robust, alignment-free deep learning framework for predicting the pathogenicity of transmembrane protein variants.
• To overcome limitations of existing methods by learning sequence-encoded biophysical signatures specific to TMPs.
• To enable efficient large-scale screening of TMP variants, especially when evolutionary or structural data are scarce.

Main Methods:

• Introduced Memo-Patho, a deep learning framework utilizing a novel within-protein, label-informed supervised contrastive pretraining strategy.
• Fused sequence-level representations from protein language models with local structural proxies derived from sequence.
• Developed an alignment-free approach, eliminating the need for multiple sequence alignments or experimental protein structures.

Main Results:

• Memo-Patho achieved high accuracy (up to 0.93) in predicting TMP variant pathogenicity across diverse benchmarks.
• Consistently outperformed leading prediction methods, even under stringent protein-level group splits.
• Demonstrated successful transfer learning to an independent KCNQ1 ion-channel cohort without re-training.

Conclusions:

• Memo-Patho provides an accurate, resource-efficient, and alignment-free method for predicting transmembrane protein variant pathogenicity.
• The framework learns discriminative, sequence-anchored signatures relevant to TMP-specific constraints.
• Offers a principled and generalizable foundation for clinical variant triage and proteome-wide mutation-effect modeling in precision medicine.