Helix Insight

Documentation / Computational Predictors

Computational Predictors

Role in Classification

In ACMG variant classification, two criteria depend on computational predictions: PP3 (computational evidence supports a deleterious effect) and BP4 (computational evidence suggests no impact). These are evidence criteria -- they contribute to the overall classification but do not determine it on their own.

Helix Insight uses two computational tools that directly influence ACMG classification: BayesDel_noAF for missense variant pathogenicity assessment, and SpliceAI for splice impact prediction. Both were selected based on ClinGen Sequence Variant Interpretation (SVI) Working Group recommendations and are calibrated against clinical truth sets.

Classification Tools vs. Displayed Predictors

Not all computational predictions shown in the results interface contribute to ACMG classification. The platform distinguishes between tools that drive classification decisions and predictors displayed for clinical reference.

ToolRoleACMG CriteriaWhat It Measures
BayesDel_noAFClassificationPP3, BP4Missense pathogenicity (calibrated meta-predictor)
SpliceAIClassificationPP3_splice, BP4/BP7 guardSplice site creation or disruption
SIFTDisplayed--Amino acid substitution tolerance
AlphaMissenseDisplayed--Protein structure-based pathogenicity
MetaSVMDisplayed--Ensemble of multiple prediction methods
DANNDisplayed--Deep neural network pathogenicity
PhyloPDisplayed--Evolutionary conservation (100 vertebrates)
GERPDisplayed--Evolutionary constraint at genomic position

Why BayesDel_noAF

The ClinGen SVI Working Group evaluated multiple computational tools against clinical truth sets and published calibrated evidence strength thresholds for BayesDel_noAF (Pejaver et al. 2022). This tool was selected for three reasons: it has ClinGen-calibrated thresholds that map directly to ACMG evidence strength levels (Supporting, Moderate, Strong); it explicitly excludes allele frequency from its model, avoiding circular reasoning with the frequency-based criteria PM2, BA1, and BS1; and it is precomputed in the dbNSFP database, requiring no external API calls during processing.

Unlike approaches that rely on a single damaging/benign threshold, BayesDel_noAF with ClinGen SVI calibration provides evidence strength modulation -- the same tool can contribute Supporting, Moderate, or Strong evidence depending on the score magnitude. This reflects the clinical reality that a variant with a very high pathogenicity score provides stronger evidence than one just above the threshold.

Why Displayed Predictors Are Shown

SIFT, AlphaMissense, MetaSVM, DANN, PhyloP, and GERP do not contribute to ACMG criteria, but they are shown in the results interface because experienced geneticists use them as additional clinical context. A geneticist reviewing a VUS may find it informative that AlphaMissense predicts the variant as pathogenic based on protein structure, even though this does not change the formal ACMG classification. These predictions help clinicians form their independent assessment alongside the automated classification.

Important

Computational predictions are supporting evidence. They contribute PP3 or BP4 criteria to the ACMG framework, but they do not determine classification on their own. A variant is never classified as Pathogenic based solely on computational evidence, and a variant is never classified as Benign based solely on the absence of computational predictions.

In This Section

Reference: Pejaver V, et al. Am J Hum Genet. 2022;109(12):2163-2177. PMID: 36413997