bio-clinical-databases-acmg-classification

What this skill does

## Version Compatibility

Reference examples tested with: requests 2.31+, pandas 2.2+, AutoPVS1 (Xiang 2020), InterVar 2.2+, GeneBe 1.0+ (Stawinski 2024 *Clin Genet*). ACMG/AMP Bayesian point system is Tavtigian 2018 *Genet Med* / 2020 *Hum Mutat*. Pejaver 2022 *AJHG* PP3/BP4 calibrated thresholds. ClinGen Splicing Subgroup 2023 (Walker *AJHG*). v3.2 ACMG SF list (Miller 2023). The ACMG 2.0 framework is in development as of May 2026; not yet published.

Before using code patterns, verify installed versions match. If versions differ:
- Python: `pip show <package>` then `help(module.function)` to check signatures
- CLI: `<tool> --version`

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying. VCEP-specific CSpec rules override default ACMG application; the authoritative directory is `https://cspec.genome.network/cspec/ui/svi/all`.

# ACMG/AMP Variant Classification Framework

**'Classify this variant per ACMG/AMP'** -> Apply 28-criterion framework using Tavtigian point system; gate on ClinGen SVI specifications and VCEP-specific overrides; assign P / LP / VUS / LB / B classification with evidence trail.

- Python (automated): GeneBe API `https://api.genebe.net/cloud/api-public/v1/variant`
- Python (rule-based): InterVar -> `python InterVar.py -i input.vcf -b hg38 --table_annovar table_annovar.pl`
- Web tools: VarSome (commercial), Franklin/Genoox (commercial), ClinGen VCI (gold standard for SVI)
- Citation: Richards 2015 *Genet Med* 17:405 (original framework); Tavtigian 2020 *Hum Mutat* 41:1734 (point system)

## The Tavtigian Bayesian Point System: The Engine Inside All Modern Classifiers

Richards 2015 specified 28 criteria with strength labels (Supporting / Moderate / Strong / Very Strong); combination rules produced P / LP / VUS / LB / B. **Tavtigian 2018/2020 demonstrated this framework is mathematically a Bayesian classifier and proposed the naturally-scaled point system that every modern automated classifier implements:**

| Strength | Points | Odds of pathogenicity |
|----------|--------|----------------------|
| Supporting | 1 | 2.08:1 |
| Moderate | 2 | 4.33:1 |
| Strong | 4 | 18.7:1 |
| Very Strong | 8 | 350:1 |

Benign codes are negative-signed. Final classification:

| Sum of points | Category |
|---------------|----------|
| >= 10 | **Pathogenic** |
| 6-9 | **Likely Pathogenic** |
| 0-5 | **VUS** |
| -1 to -6 | **Likely Benign** |
| <= -7 | **Benign** |

InterVar / GeneBe / VarSome / Franklin all implement Tavtigian point summation under the hood. Combinations never appearing in the 2015 combining rules (e.g., PVS1_VeryStrong + PM2_Supporting -> LP) emerge naturally from point arithmetic.

## PVS1 Decision Tree (Abou Tayoun 2018 *Hum Mutat* 39:1517)

PVS1 is the most consequential code: pathogenic Very Strong (8 points) for predicted loss-of-function in a gene where LoF is established disease mechanism. The 2018 decision tree refined PVS1 from a binary into a graded code based on:

1. **Variant type**: nonsense / frameshift / canonical +-1,2 splice / initiation codon / single-exon deletion / multi-exon deletion.
2. **NMD prediction**: variant in 5'-most exon OR >50bp upstream of last exon-exon junction -> NMD-triggered. Else truncated protein.
3. **Critical region**: removal of >10% of coding sequence OR removal of a critical functional domain.
4. **Alternative isoform**: does the variant affect a transcript expressed in disease-relevant tissue?

Output strengths:

| Output | Original Strength |
|--------|-------------------|
| PVS1_VeryStrong | Strongest (Very Strong) |
| PVS1_Strong | Strong |
| PVS1_Moderate | Moderate |
| PVS1_Supporting | Supporting |

**Subsumption rule** (Abou Tayoun 2018): PVS1 + PP3 -> only PVS1 counts (PP3 is subsumed). Same for PVS1 + PM4.

**>15 VCEP-specific PVS1 trees exist** as of 2024 (CDH1, ENIGMA BRCA1/2, FH LDLR/APOB/PCSK9, InSiGHT MMR, RASopathies, hearing loss, hypertrophic cardiomyopathy, Rett/Angelman, etc.). The automated implementation is **AutoPVS1** (Xiang 2020).

## Pejaver 2022 PP3/BP4 Calibrated Thresholds (the load-bearing 2024+ calibration)

Pejaver 2022 *AJHG* 109:2163 Bayesian-calibrated 13 missense predictors to PP3/BP4 strength levels using ClinVar P/B variants with leave-one-gene-out cross-validation.

| Predictor | BP4_Strong | BP4_Moderate | BP4_Supporting | PP3_Supporting | PP3_Moderate | PP3_Strong | Fails when |
|-----------|-----------|--------------|----------------|----------------|--------------|------------|-----------|
| **REVEL** | <= 0.003 | <= 0.016 | <= 0.290 | >= 0.644 | >= 0.773 | >= 0.932 | Stacked with BayesDel/VEST4 (training overlap; double-counting) |
| **BayesDel (no AF)** | <= -0.36 | <= -0.18 | <= -0.08 | >= 0.13 | >= 0.27 | >= 0.50 | Combined with AF-aware variant (use no-AF version with PM2_Supporting) |
| **VEST4** | <= 0.302 | <= 0.449 | <= 0.302 | >= 0.764 | >= 0.861 | >= 0.965 | Indels (missense-trained); regulatory variants |
| **MutPred2** | (Pejaver 2022) | -- | -- | -- | -- | -- | Genes with sparse MAVE training data |
| **AlphaMissense** | NOT ClinGen-endorsed | -- | -- | Use as supporting only | -- | NOT ClinGen-endorsed | Developer threshold 0.564 misapplied as PP3 |

**The two numbers to memorize: REVEL >= 0.932 = PP3_Strong; REVEL <= 0.290 = BP4_Strong (or <= 0.003 BP4_VeryStrong).**

**AlphaMissense calibration** (Schmidt 2025 *Genet Med* 27:e101339, originally Pejaver et al. bioRxiv 2024.09.17): AlphaMissense reaches **PP3_Strong** and **BP4_Moderate** at calibrated thresholds. **Critical:** the developer-recommended 0.564 threshold is NOT the Pejaver PP3 threshold. ClinGen has NOT endorsed AlphaMissense PP3 strength as of May 2026; treat as supporting evidence only.

**Do not stack predictors.** REVEL, BayesDel, VEST4 share ClinVar/HGMD training data; using REVEL >= 0.773 AND BayesDel >= 0.27 to claim "two independent moderate hits" is double-counting. Pejaver 2022 explicitly recommends using ONE predictor per variant.

## PM2_Supporting (ClinGen SVI 2020)

The original PM2 ("absent from controls") was over-weighted. SVI 2020 downgraded to **PM2_Supporting** (1 point, not 2). Mechanism: most rare variants are benign. Empirical recalibration showed ~6 variants per gene downgrade from LP to VUS when PM2 -> Supporting. Many 2017-2019 LP curations require re-classification post-SVI 2020 update.

## PS3/BS3 Functional Evidence (Brnich 2020 *Genome Med* 12:3)

OddsPath framework; the four-step SOP:

1. Define disease mechanism for the gene.
2. Evaluate assay class (e.g., MAVE, biochemical, animal model).
3. Evaluate specific assay instance (controls, replicate consistency).
4. Apply per-variant.

OddsPath calibration mapping to ACMG strengths:

| OddsPath | Pathogenic strength | Benign strength |
|----------|--------------------:|----------------:|
| > 18.7 | Very Strong | n/a |
| 4.3 - 18.7 | Strong | -- |
| 2.1 - 4.3 | Moderate | -- |
| 1.2 - 2.1 | Supporting | (mirror) |

**MAVEdb deep-mutational scans** with >=11 controls (>=5 P/LP + >=5 B/LB) can yield up to PS3_Strong/BS3_Strong via OddsPath calibration. This is the entry point for MAVE/saturation-mutagenesis evidence into ACMG.

**Default-Strong PS3 application is increasingly over-strengthening** without OddsPath calibration; ClinGen SVI recommends moving toward PS3_Moderate as default unless OddsPath > 4.3.

## ClinGen SVI Splicing Subgroup 2023 (Walker *AJHG* 110:1046)

**SpliceAI is the recommended primary splicing tool.** Calibrated thresholds:

| SpliceAI DS_max | Strength |
|-----------------|----------|
| >= 0.5 (Jaganathan 2019 default) | Can support PP3_Strong with corroborating evidence |
| >= 0.20 | Minimum threshold for ANY splicing PP3 |
| < 0.1 | BP4_Moderate (weaker than missense BP4 because absence of predicted aberrant splicing is less informative) |

**SpliceVault / 300K-RNA** (Dawes 2023 *Nat Genet* 55:324): does NOT predict whether a variant is splice-altering; pre