Hello everyone,

I am currently working on a project involving Protein Language Models (PLMs) (similar to EVE, Revel, and AlphaMissense). My goal is to train/fine-tune these models to classify missense variants (Pathogenic vs. Benign) specifically within South Asian populations.

I am using gnomAD and SG10K (Singapore 10K Pilot) datasets, which I have annotated using Ensembl VEP.

The Observation: While inspecting the VEP summary statistics (HTML report) for Chromosome 22, I noticed a massive discrepancy in the counts, which has led to a dilemma regarding data leakage and redundancy:

Consequences (all): ~763,000 missense variants.

Consequences (most severe): ~70,910 missense variants.

The Problem: I understand that "Consequences (all)" counts the variant effects across all overlapping transcripts/isoforms, while "most severe" counts the variant once per genomic position based on its highest impact.

Since Protein Language Models typically operate on a specific amino acid sequence, my concern is that using the "All" set would introduce redundancy and data leakage (e.g., the exact same genomic mutation appearing 5 times because it affects 5 slightly different transcripts of the same gene).

For training PLMs like AlphaMissense or EVE, is the standard practice to filter strictly for the Canonical Transcript (e.g., CANONICAL=YES or MANE_Select in VEP)?

Would including non-canonical isoforms be considered valid data augmentation, or is it strictly considered data leakage/bias in this context?

Implementation: If I filter for CANONICAL=YES inside the VEP CSQ field, does this generally align with the "Most Severe" count, or are there edge cases I should be aware of (e.g., where the canonical transcript has a synonymous change but a non-canonical one has a missense change)?

Any advice or references to how models like these handled this would be greatly appreciated.

Thank you!