PGViS: Personal Genome Variant interpretation Score for lung cancer genomes.

Aggregation into a single per-patient score of three variant-level signals: predicted disruption of transcription factor binding and splice sites by a DNA foundation model, cancer-versus-reference alternate allele frequency shift, and a regulatory interaction term derived from allele frequency ratios.

Inherited lung cancer risk arises from both coding and non-coding germline variants, but the functional non-coding component remains largely uncharacterised, as genome-wide association studies and polygenic risk scores identify tag rather than causal variants. PGViS is a statistical framework quantifying individual non-coding germline regulatory risk in non-small cell lung cancer, integrating three variant-level signals: DNABERT-predicted disruption of transcription factor binding and splice sites, the cancer-versus-reference alternate allele frequency shift, and a regulatory interaction term derived from cancer-to-reference allele frequency ratios; each signal is weighted by cohort prevalence, then aggregated into a single ancestry-matched, reference-normalised score. Applied to germline whole-genome sequencing from 1,102 TCGA and CPTAC patients, with the 1000 Genomes Project (n = 2,504) as the normal population reference, PGViS separated adenocarcinoma and squamous cell carcinoma from controls in European ancestry, and adenocarcinoma in East Asian ancestry. Genes at contributing loci were enriched for PI3K-Akt, Wnt, DNA damage response and epithelial-mesenchymal transition programmes, and smoking-stratified analysis concentrated this signal on canonical non-small cell lung cancer driver pathways.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

The framework is ingenious, but the study design precludes any conclusion about risk: TCGA and CPTAC are tumour series rather than at-risk cohorts, and the 1000 Genomes individuals are matched neither on age nor on smoking, the main determinant of lung cancer. Separating cases from controls is not predicting individual risk, and no absolute risk calibration or quantified discrimination is reported in this preprint. File it as a proof of concept for applying DNA foundation models to the germline non-coding space, with no place in cancer genetics practice today.

Analysis by Dr Thibaut Benquey