Thikra A.

Bioinformatics Specialization · Cohort 2026

Prediction Errors in Human Mitochondrial tRNAs are Structure-dependent: MitoTIP Over-calls in Base-paired Stems and Under-calls in Unpaired Loops

A tRNA is not uniform. Base-paired stems hold the fold together; unpaired loops carry the recognition and modification chemistry. Yet MitoTIP, the predictor clinicians consult for mitochondrial tRNA variants, is read through one pathogenicity threshold applied identically everywhere. I built a position-level structural atlas of all 22 human mt-tRNAs by parsing the published Mamit-tRNA numbering tables, so every coordinate carries a region label taken from a published source rather than inferred from arithmetic. Auditing 622 of MitoTIP's own labelled variants against that atlas, 23.0% of base-paired variants are classified wrongly versus 1.6% of unpaired ones, a 14-fold difference, and every stem error is a false positive while every loop error is a false negative, with no exception in either direction. This is a calibration failure, not a scoring failure. Region-specific thresholds, derived by two different rules because the two distributions differ, raise detection from 75.9% to 89.7% on variants confirmed after MitoTIP was published, at a measured specificity cost. All code and data are openly available and rebuild from raw sources with one command.

Read the final report (PDF)