INDELVAR model predicts pathogenicity of in-frame insertions and deletions using AlphaFold structural data
A random forest classifier integrating protein structural context, evolutionary conservation, and gene constraint offers calibrated PP3/BP4 thresholds for variant classification.
Researchers have posted a preprint describing INDELVAR, a machine-learning model for assessing the pathogenicity of in-frame insertions and deletions (indels) of one to ten amino acids — a variant class that has historically been difficult to interpret under standard frameworks.
The model integrates 37 features drawn from AlphaFold-predicted wild-type protein structures, evolutionary conservation scores, local sequence change, gene constraint metrics, and curated protein annotations. According to the preprint, pathogenic variants more frequently affected protein regions characterised by high AlphaFold confidence, low solvent exposure, dense local packing, and strong evolutionary conservation — findings consistent with the expectation that buried, structurally critical residues are less tolerant of insertions or deletions.
A notable aspect of the work is the calibration of PP3 and BP4 evidence codes, which are used within the ACMG/AMP variant interpretation framework to weight computational evidence in favour of or against pathogenicity. The authors report calibrated thresholds intended to bring in-frame indel interpretation into closer alignment with the evidence-strength levels those codes are meant to represent.
This work is relevant to clinical genetics laboratories and genetic counsellors who routinely encounter variants of uncertain significance (VUS) in this structural category. The preprint has not yet been peer-reviewed.
Plain-language version
For patients, families, and general readers. Educational only — not medical advice.
When geneticists look at someone's DNA, they sometimes find small changes where a short stretch of DNA — encoding a few amino acids, the building blocks of proteins — has been inserted or deleted. Working out whether these changes cause disease or are harmless is often genuinely difficult.
Researchers have posted early findings describing a computer model called INDELVAR, designed to help predict whether these changes are likely to be harmful. The model uses detailed three-dimensional protein structures, generated by a separate AI tool called AlphaFold, alongside information about how much a stretch of protein has stayed the same across species over millions of years. Changes in regions that are tightly packed and evolutionarily conserved were more often found to be harmful.
This is a preprint — early research that has not yet been checked by independent scientists. It describes a tool intended to assist laboratory scientists and genetics professionals, not to provide individual results to patients.
This is an educational summary, not medical advice. If anything here raises questions for you, please speak with your GP or a clinical professional.
Sources
Read the original reporting — these are the public sources this summary draws from.
-
Primary sourcePreprint bioRxiv (Cold Spring Harbor Laboratory) · 2026-08-18INDELVAR: structure-informed prediction of in-frame indel pathogenicity with calibrated PP3/BP4 thresholds