How AI‑CURA Uses Large Language Models to Automate ACMG Variant Classification

Problem

Interpretation of thousands of variants from whole‑genome sequencing in clinical genetics requires manual literature review for ACMG/AMP classification, averaging about four hours per case and representing a major bottleneck.

AI‑CURA Architecture

AI‑CURA implements an automated workflow composed of two layers:

Rule‑engine layer : Executes 13 ACMG criteria that do not need literature evidence (≈65% of the rules), such as allele‑frequency thresholds and computational prediction scores.

LLM inference layer : Handles the remaining seven literature‑dependent criteria (PVS1‑RNA, PS2/PM6, PS3, PS4, PM3, PP1, PP4). Two large language models— DeepSeek‑R1 and o3‑mini‑high —are used to generate literature summaries and extract supporting evidence. Each rule has a dedicated prompt, and an auxiliary knowledge base incorporates ClinGen recommendations and known gene‑phenotype associations. A prototype module employing Claude‑4.0‑Sonnet converts family‑tree figures and complex tables into text for downstream processing.

Performance on Literature‑Dependent Rules

Both models achieved 100% specificity (no false positives when evidence was absent). DeepSeek‑R1 consistently showed higher sensitivity than o3‑mini‑high, capturing nearly all required evidence without introducing errors.

ClinGen Validation

AI‑CURA was evaluated on 150 ClinGen‑annotated variants. Results:

Rule‑application consistency for DeepSeek‑R1: 84.0%.

Overall classification agreement: 96.0%.

When pathogenic and likely‑pathogenic categories are merged (clinical‑equivalence), agreement rises to 99.3%, exceeding typical inter‑lab variability.

Re‑analysis of ClinVar Conflict Variants

In a simulated re‑analysis of 150 ClinVar variants with conflicting interpretations, AI‑CURA provided definitive classifications for 70% of cases. The final classification concordance with expert manual review was 98.7%.

Limitations

Supports only Mendelian single‑nucleotide variants and short insertions/deletions.

Limited ability to extract information from figures or complex tables; rules that rely on pedigree graphics have reduced accuracy.

Key Attributes

Transparent, step‑by‑step reasoning process.

Open‑source and locally deployable, satisfying medical data security requirements.

Extensible framework that treats genomic data as “living” information, enabling continuous reinterpretation as knowledge evolves.