Computational predictors in ACMG variant classification

Accessible text version

Edited video transcript

Repeated spoken phrases and verbal corrections have been removed for reading. The sequence follows all 20 slides. The SpliceAI section includes the current production strength modulation described above. The video includes optional English captions generated directly from the supplied SRT timings.

1. 01

   Computational predictors

   Computational predictors are among the most searched topics in variant interpretation. Names such as BayesDel, AlphaMissense, MetaSVM, DANN, SIFT and SpliceAI lead to a practical question: what should any score change in an ACMG classification? This briefing separates predictors that provide formal computational evidence from predictors shown as clinical context. It also separates missense and splice mechanisms and explains calibration, evidence strength and double counting. The goal is a disciplined workflow: define the biological question, choose the appropriate predictor, apply a validated threshold, assign only the permitted evidence strength and combine the result with independent evidence.

2. 02

   Evidence, not a verdict

   A computational predictor does not classify a variant. It estimates a property that may be relevant to classification. SIFT estimates whether an amino acid substitution is tolerated at a conserved protein position. SpliceAI estimates whether sequence context may create or disrupt splice sites. BayesDel combines several signals into a missense pathogenicity score. ACMG classification asks whether the total evidence supports pathogenicity or benignity for a particular disease and inheritance model. That record can include population frequency, functional studies, segregation, phenotype, allelic data and computational evidence.

3. 03

   The ACMG evidence frame

   The ACMG and AMP framework organises evidence by direction and strength. Pathogenic evidence codes begin with P and benign evidence codes begin with B. Computational evidence enters through PP3 on the pathogenic side and BP4 on the benign side. Quantitative calibration maps defined score intervals to defined evidence strengths. A predictor is not admitted because it is popular or technically sophisticated. It is admitted because a specified range has been evaluated against an appropriate truth set under a documented method.

4. 04

   One stream in the complete record

   Population evidence asks whether an allele frequency is compatible with the disease mechanism. Functional and case evidence asks what has been observed in experiments, families and affected individuals. Computational evidence infers biological impact from sequence patterns and annotations. It is fast, reproducible and available for many variants, including variants that lack a patient series. Its limitation is that it remains an inference rather than a direct observation of causality. A rare variant with a high predictor score may therefore remain a VUS.

5. 05

   PP3 and BP4

   PP3 records computational support for a deleterious effect. BP4 records computational support for no impact. Running several tools and taking a majority vote can repeat correlated information because predictors often share conservation features, training variants and annotations. A defensible process chooses one predefined tool for a formal evidence path, maps its score to a calibrated strength and leaves an indeterminate interval where neither criterion applies. A missing damaging prediction is not automatically benign evidence.

6. 06

   Scope, calibration and independence

   Three rules make computational evidence defensible. Choose the tool before seeing the answer. Use validated score intervals rather than developer labels alone. Protect independence from evidence already counted elsewhere. A predictor that includes allele frequency can overlap with population criteria. A splice prediction can describe the same mechanism already captured by PVS1. These checks determine whether a score can enter an auditable clinical workflow or should remain contextual information.

7. 07

   Two formal paths in Folklore

   Folklore uses BayesDel_noAF for the formal missense path and SpliceAI for the formal splice path. BayesDel_noAF excludes allele frequency and reduces circularity with PM2, BA1 and BS1. SpliceAI evaluates local sequence context for donor or acceptor gain and loss. A variant may raise a missense question, a splice question or both. The scores are not averaged. Each path has its own thresholds, consequences and safeguards.

8. 08

   Why BayesDel_noAF

   BayesDel is a meta-predictor that combines information from multiple annotations and prediction methods. The noAF model removes allele frequency, which matters because frequency is already assessed through independent ACMG population criteria. ClinGen SVI calibration also gives the score a graded interpretation. A score just above a pathogenic boundary does not carry the same weight as an extreme score, and an indeterminate region receives no computational criterion.

9. 09

   BayesDel_noAF thresholds

   The current Folklore intervals are 0.130 to 0.289 for PP3 Supporting, 0.290 to 0.517 for PP3 Moderate and 0.518 or higher for PP3 Strong. Scores from minus 0.360 to minus 0.181 support BP4, while minus 0.361 or lower supports Moderate benign evidence. Scores from minus 0.180 through 0.129 are indeterminate. The missense path also applies relevance and disease-association guards. PM1 and PP3 are capped at a combined Strong equivalent, and benign missense evidence requires no concerning splice signal.

10. 10

    SpliceAI thresholds and current safeguards

    SpliceAI produces acceptor gain, acceptor loss, donor gain and donor loss delta scores, and Folklore reads their maximum. The current production intervals are 0.2 to below 0.5 for PP3_splice Supporting, 0.5 to below 0.8 for Moderate and 0.8 or higher for Strong. These strengths remain subject to consequence, disease-association and transcript safeguards. If PVS1 already captures the same loss-of-function mechanism, the splice criterion is suppressed. A score at or below 0.1 participates only in defined benign guards whose eligibility depends on the variant consequence.

11. 11

    Do not count one signal twice

    ACMG codes are combined as evidence contributions, so dependence matters. Two predictors can share training sets or component annotations. A splice model and a loss-of-function rule can describe the same transcript disruption. A predictor containing frequency can repeat rarity evidence. More displayed signals do not necessarily mean more independent information. Calibration controls the strength of one tool, while dependency guards control its relationship with the rest of the record.

12. 12

    Formal evidence and displayed context

    Several predictor scores can appear together on a result screen, but not all of them contribute to automated classification. BayesDel_noAF and SpliceAI connect to explicit formal evidence paths. SIFT, AlphaMissense, MetaSVM, DANN, PhyloP and GERP are displayed for clinical context. This separation lets a reviewer inspect useful signals without turning every output into another ACMG vote.

13. 13

    SIFT, PhyloP and GERP

    SIFT evaluates a missense substitution using homologous protein sequences. PhyloP measures whether a genomic position evolves more slowly or quickly than expected under a neutral model. GERP estimates evolutionary constraint from substitutions rejected over time. Conservation can identify an important position, but it does not prove that every possible substitution at that position is damaging. These scores remain contextual in Folklore.

14. 14

    AlphaMissense, MetaSVM and DANN

    AlphaMissense combines protein sequence modelling with structural context. MetaSVM combines multiple annotations through a support vector machine. DANN uses a neural network over genomic annotations and can score single nucleotide variants beyond coding missense positions. Their architectures and scopes differ, but difference does not guarantee statistical independence. Folklore displays them for review while the formal missense path uses BayesDel_noAF.

15. 15

    Agreement, disagreement and missing scores

    Agreement can provide useful context, but it does not multiply formal evidence. If BayesDel crosses a calibrated threshold and other displayed tools point in the same direction, PP3 is still recorded once through the predefined path. Disagreement may reflect different transcripts, variant scopes or mechanisms and should trigger inspection. A missing output means that evidence is unavailable. It does not mean that the predictor returned a benign result.

16. 16

    Synthetic worked example

    Consider a synthetic missense variant with BayesDel_noAF 0.35 and SpliceAI 0.02. AlphaMissense predicts likely pathogenic and SIFT predicts deleterious. The missense mechanism is relevant, and BayesDel 0.35 falls in the Moderate pathogenic interval, so the formal computational contribution is PP3_Moderate. AlphaMissense and SIFT do not add more PP3 points. The reviewer returns to population frequency, gene-disease validity, phenotype, segregation, functional data and other independent evidence.

17. 17

    Review checklist

    Before a predictor changes classification, confirm scope, calibration and independence. Scope asks whether the tool supports the variant type, mechanism and clinically relevant transcript. Calibration asks whether the exact score interval maps to a documented evidence strength. Independence asks whether the same biological or statistical signal has already entered the record. A score that fails scope is not applicable. A score that lacks calibration may remain contextual. Dependent evidence should be capped or omitted.

18. 18

    Prediction narrows attention

    Computational predictors can highlight a residue, suggest a splice mechanism, identify a constrained position and focus deeper review. They do not establish the disease relationship of a gene, prove that a molecular effect explains the patient or replace functional and segregation data. Even an accurate model can fail for a specific transcript, mechanism or gene. The final interpretation returns to the complete observed evidence and the clinical context.

19. 19

    Primary sources

    Richards and colleagues defined the ACMG and AMP sequence variant interpretation framework. Pejaver and colleagues provided quantitative calibration recommendations for computational missense predictors. Walker and colleagues specified how predicted and observed splice effects should enter the ACMG framework, including protection against double counting. Jaganathan and colleagues described SpliceAI. Folklore documentation records the current operational thresholds and safeguards used by the platform.

20. 20

    Continue with the documentation

    The Folklore documentation provides a page for each predictor. The BayesDel page records missense score intervals, evidence strengths and dependency guards. The SpliceAI page records the four delta scores, current production intervals and PVS1 safeguards. Separate pages cover SIFT, AlphaMissense, MetaSVM, DANN, PhyloP and GERP. Read the model, threshold, mechanism and ACMG role together rather than reading a prediction in isolation.