Our Research Focus

A Machine Learning Approach to Prioritizing Secondary Mis-Match Repair Pathway Genes Missed by Current Lynch Syndrome Multi-Gene Sequencing Panels

Lynch syndrome (LS) occupies a paradoxical position in clinical genetics: it is the most comprehensively characterized hereditary cancer syndrome, yet one in three clinically suspected cases leaves standard genetic testing without a molecular diagnosis. This is not a failure of sequencing technology it is a failure of what we choose to sequence. Germline pathogenic variants in MLH1, MSH2, MSH6, PMS2, or EPCAM deletions impair DNA mismatch repair (MMR), generating a mutator phenotype that drives colorectal and endometrial malignancy before age 50 in approximately 1 in 279 individuals. The clinical stakes are exceptional: colonoscopic surveillance reduces colorectal cancer mortality by over 60%, and pembrolizumab achieves response rates exceeding 40% in MMR-deficient metastatic disease. Diagnosis is not merely prognostic it is therapeutically decisive. The gap persists for compounding, addressable reasons. The Amsterdam II and Bethesda criteria still widely used as testing triggers are absent in up to 63% of confirmed carriers, systematically excluding patients who lack classic familial clustering. Multigene panel testing has expanded reach, but a structural problem remains: genes harboring well-documented MMR-pathway pathogenic variants are absent from major commercial panels, leaving high-risk individuals undiagnosed after negative standard testing. Genotype-specific blind spots compound this pseudogene interference confounding PMS2 sequencing, attenuated microsatellite instability masking MSH6 carriers, and deep intronic variants invisible to exon-capture generating false-negatives even when the correct gene is tested. The cumulative consequence is measurable and avoidable: missed cascade testing, delayed surveillance, and forfeited immunotherapy eligibility across thousands of families annually. This project addresses that gap at its structural root. By mining ClinVar, COSMIC, gnomAD, and TCGA for genes harboring ≥3 Lynch-specific pathogenic variants absent from >70% of commercial panels, we quantify the evidence-to-adoption gap systematically for the first time. A multi-dimensional prioritization framework integrating clinical, functional, population, and MMR network biology evidence validated through interpretable regularized logistic regression with SHAP-attributed feature importance identifies which overlooked genes carry the strongest case for clinical inclusion. Multi-cohort validation across TCGA, NCBI GEO/ArrayExpress, and the All of Us Research Program delivers the ancestry-diverse replication, including dedicated Hispanic/Latino subgroup analyses, that prior LS research has largely omitted.

Systematic Computational Discovery of Novel Hereditary Breast and Ovarian Cancer Syndrome Candidate Genes: A Multi-Cohort Bioinformatics Framework with Hispanic/Latino Representation

Hereditary Breast and Ovarian Cancer (HBOC) syndrome is among the most clinically actionable inherited cancer predispositions, yet a substantial fraction of familial breast and ovarian cancer clustering remains genetically unexplained even after comprehensive standard panel testing of established genes including BRCA1, BRCA2, PALB2, ATM, CHEK2, RAD51C, RAD51D, BARD1, BRIP1, and their homologous recombination repair (HDR) pathway partners. This diagnostic gap carries measurable clinical consequences: missed eligibility for PARP inhibitor therapy in HDR-deficient metastatic disease, foregone risk-reducing surgical interventions, and failed cascade testing in at-risk relatives who remain unaware of their inherited risk profile. We propose that secondary HDR susceptibility genes those harboring well-documented HBOC-specific pathogenic variants yet systematically excluded from commercial panels account for a meaningful, quantifiable portion of this unresolved hereditary fraction. This pilot study interrogates that premise through a five-layer computational strategy: exhaustive variant-level curation from ClinVar, COSMIC, LOVD, TCGA, and gnomAD; a multi-dimensional evidence-scoring framework integrating clinical annotation, population allele frequencies, functional genomic signals, and degree-normalized BRCA1/2 interactome centrality; dual regularized machine learning models (logistic regression and random forest) trained on historical panel adoption decisions, reporting precision, recall, ROC-AUC, and SHAP-attributed feature importance; BRCA1/2-centered protein interactome network analysis with triple-filter candidate advancement; and multi-cohort validation across TCGA, NCBI GEO/ArrayExpress, and the NIH All of Us Research Program. Deliverables 20–25 ranked candidate genes, a reproducible Docker-containerized scoring pipeline, and an interpretable ML framework constitute rigorous preliminary data for a subsequent NIH R15 Academic Research Enhancement Award application.