What a prediction contributes
Computational tools can estimate whether a change may affect a gene product. Their value depends on calibration, the variant type, the gene, and the threshold used for that evidence framework.
The ClinGen FH expert panel published specific use and thresholds for in silico prediction in LDLR interpretation. That is a reminder that the model name alone is not enough. Reviewers need to know how the output is being used.
Four questions for a computational result
What was predicted?
Separate a protein-impact prediction from a clinical classification. They answer different questions.
Was the tool calibrated?
Record the model, version, applicable variant class, and threshold used in the chosen framework.
Is the evidence independent?
Avoid counting several correlated models as if they were unrelated evidence sources.
What evidence can overrule it?
Functional results, population data, segregation, and case evidence may support or challenge the prediction.
Product boundary
CardioGen AI uses an AlphaMissense pathogenicity value as one input in its current four-gene model. The resulting score is internal, non-clinical model evidence and requires qualified review alongside independent evidence.
Make uncertainty inspectable
A review-ready result should show which computational signal was used, what it contributed, and why it did not settle the classification alone. That makes disagreement easier to review and later evidence easier to add.
Primary references
- ClinGen FH Variant Curation Expert Panel consensus guidelines for LDLR variant classification, Genetics in Medicine, 2022.
- ClinGen FH Expert Panel specifications to the ACMG/AMP guidelines, current criteria registry.
- ACMG/AMP standards and guidelines for interpretation of sequence variants, Genetics in Medicine, 2015.