acmg-points

Points-based ACMG/AMP variant classification, side by side with the 2015 combining rules — so you can see exactly where the two systems disagree on your evidence.

ClinGen is moving variant classification from the 2015 combining rules to a points system (Tavtigian 2020, the basis of the forthcoming SVC v4). Most labs still reason in 2015 rules; most calculators implement one system or the other. acmg-points runs both on the same evidence and flags the disagreements, with the provenance of every point.

pip install git+https://github.com/MargoSolo/acmg-points      # no dependencies
acmg-points classify PVS1 PM1

  PVS1  VeryStrong  +8
  PM1   Moderate    +2
  total             +10 → Pathogenic          [tavtigian2020]
  2015 rules        → Likely pathogenic       (2015 combining rule met)
  ⚠️ SCHEMES DISAGREE

Where the systems disagree

Fifteen evidence sets from examples/cases.txt, placed on the points scale; the band gives the points class, and ✕ marks the sets where the 2015 rules say something else:

Full table: examples/compare.md.

Conflicting evidence. Richards 2015 puts a variant in Uncertain significance when "the criteria for benign and pathogenic are contradictory". That sentence has two readings, and implementations differ. acmg-points defaults to the stricter one and lets you switch:

BA1 is stand-alone benign evidence under both readings and is not overridden by supporting pathogenic criteria. The JSON output records which reading was used.

PM2. The 2015 default is Moderate; ClinGen SVI recommended PM2 at Supporting strength in 2020. Apply it as PM2_Supporting — the package does not silently downgrade it, because the strength is your decision.

Use

acmg-points classify PVS1 PM2_Supporting PP3      # modified strengths: CODE_Strength (also CODE:Strength, CODE-Sup)
acmg-points classify PS1 PM1 --json               # per-criterion points and both verdicts, machine-readable
acmg-points compare --file cases.txt              # one evidence set per line, '#' comments → markdown table
acmg-points table                                 # the scheme in force, with citation

Strengths: Supporting Moderate Strong VeryStrong (Sup/Mod/Str/VS), StandAlone for BA1. Applying a criterion twice is an error.

What is implemented — and what deliberately is not

flowchart LR
  E["applied criteria<br/>PVS1 · PM2_Supporting · BS1 …"] --> P["points<br/>Tavtigian 2020<br/>+1 +2 +4 +8 / −1 −2 −4 −8, BA1 = −8"]
  E --> R["2015 combining rules<br/>Richards 2015, Table 5<br/>counted at applied strength"]
  P --> V[both verdicts · disagree flag · JSON]
  R --> V

• Points: Tavtigian et al. 2020, Hum Mutat 41:1734. Bands: Pathogenic ≥ 10 · Likely pathogenic 6–9 · VUS 0–5 · Likely benign −6…−1 · Benign ≤ −7.
• 2015 rules: Richards et al. 2015, Genet Med 17:405, including the conflict rule. Criteria count at their applied strength (PM2_Supporting is a Supporting criterion), as ClinGen SVI modifiers intend.
• Not implemented on purpose: any "SVC v4.0 points table". v4 is in pilot and unpublished. Schemes live in one file, schemes.py; when the official table appears it becomes a second scheme, not a rewrite.

This is a combiner, not a classifier: it assumes you have decided which criteria apply at which strength. It does not assess evidence, query gnomAD, or know VCEP gene-specific specifications. It will also happily sum criteria that should not be co-applied (e.g. PVS1 with PM4) — mutual-exclusion checks are on the roadmap.

Tests

Five pytest cases pin the published bands, the strength modifiers, the 2015 conflict rule and the three divergences above; CI also regenerates the comparison table.

Roadmap

Official SVC v4.0 table as a drop-in scheme · SVI modifier library (PVS1 decision tree, calibrated PP3/BP4) · co-application checks · VCEP-specific schemes · batch from CSV / VCF INFO · JOSS paper.

Cite

DOI (all versions): 10.5281/zenodo.22286276 · this version: 10.5281/zenodo.22286277

Soloshenko M. acmg-points: points-based ACMG/AMP classification alongside the 2015 rules. 2026, v0.1.0. Please also cite Tavtigian et al. 2020 and Richards et al. 2015 — the science is theirs. MIT License.