Cross-gene kinase domain hotspot analysis with AI pathogenicity prediction identifies FGFR2 D650G as a candidate actionable variant

Technical Summary: Cross-gene Kinase Domain Hotspot Analysis with AI Pathogenicity Prediction

Problem Statement
Precision oncology faces a critical bottleneck in the interpretation of Variants of Unknown Significance (VUSs), which constitute approximately 90% of variants detected through clinical next-generation sequencing. This ambiguity hinders the translation of genomic data into effective targeted therapies. Traditional approaches rely on gene-centric analyses that identify recurrent mutations within single genes; however, this perspective may overlook driver mutations that occur at functionally equivalent positions across different genes within the same protein family. Furthermore, distinguishing true oncogenic drivers from neutral passenger mutations within the vast landscape of VUSs remains a central challenge.

Methodology
The authors developed a systematic framework integrating cross-gene mutation mapping with AI-driven pathogenicity prediction to filter the VUS landscape. The methodology proceeded through the following steps:

1. Data Aggregation and Mapping: The study analyzed 32,462 missense mutations and in-frame indels from 41 receptor tyrosine kinase (RTK) genes using data from the AACR Project GENIE database (versions 14.0-public and 14.1-public).
2. Consensus Sequence Generation: Mutations were mapped onto a consensus kinase-domain sequence derived from multiple sequence alignment (MSA) of the 41 RTK genes. This resulted in a unified mutational landscape of 403 aligned residues (ARs).
3. Hotspot Identification: The authors identified cross-gene mutation hotspots by aggregating mutations at specific AR positions. Significant hotspots were defined as positions accumulating >500 total mutations from ≥15 distinct RTK genes.
4. AI-Based Stratification: To distinguish functional drivers from passengers within these hotspots, the study applied two AI-based pathogenicity prediction tools: AlphaMissense (leveraging deep learning on protein structure and conservation) and PrimateAI-3D (trained on primate genetic variation). Variants were classified based on predicted pathogenicity scores.
5. Functional Validation: Candidates prioritized by the computational framework were functionally validated using in vitro assays in HEK293H cells. This included assessing ligand-independent autophosphorylation, downstream ERK activation, and sensitivity to FDA-approved FGFR inhibitors (erdafitinib and futibatinib). Recombinant kinase-domain proteins were also used for biochemical inhibitor efficacy assays.

Key Results

• Identification of Conserved Hotspots: The cross-gene analysis revealed six major mutation hotspots. While some hotspots were dominated by specific gene mutations (e.g., EGFR exon 19 deletions) or known resistance mechanisms (e.g., EGFR T790M), two hotspots—AR 78 (back-loop) and AR 274 (activation loop)—displayed heterogeneous compositions with contributions from over 15 distinct RTK genes.
• AI-Driven Prioritization: AI analysis revealed a distinct functional divergence between the two heterogeneous hotspots. The back-loop hotspot (AR 78) was enriched with VUSs predicted as likely benign (e.g., ALK R1181C), suggesting it tolerates passenger mutations. In contrast, the activation-loop hotspot (AR 274) contained 85% of its VUSs classified as likely pathogenic by AlphaMissense.
• Discovery of FGFR2 D650G: Within the AR 274 hotspot, the VUS FGFR2 D650G received a high pathogenicity score (0.89), comparable to established drivers.
• Functional Confirmation: In vitro assays demonstrated that FGFR2 D650G conferred ligand-independent autophosphorylation and downstream ERK activation levels comparable to the known driver FGFR2 N549K.
• Therapeutic Sensitivity: The FGFR2 D650G mutant exhibited sensitivity to FDA-approved FGFR inhibitors, futibatinib and erdafitinib, with IC₅₀ values comparable to or better than the FGFR2 N549K driver. Clinical data from the GENIE cohort further supported its driver status, showing a significantly elevated variant allele frequency (VAF) consistent with positive selection.

Significance and Claims
The paper claims that this "computation-to-validation" approach successfully identifies overlooked candidate driver mutations that traditional gene-centric analyses might dismiss as background noise. By redefining somatic mutations according to homologous positions within a consensus sequence and overlaying AI-based pathogenicity predictions, the authors demonstrate a method to expand the repertoire of candidate actionable therapeutic targets.

Specifically, the study presents FGFR2 D650G as a previously unrecognized candidate driver mutation with potential clinical utility as a biomarker for FGFR-targeted therapies. The authors assert that integrating evolutionary conservation with AI prediction provides a high-throughput filter to prioritize rare genetic alterations for functional validation, thereby addressing the challenge of VUS interpretation in precision oncology. The authors note that while the framework is demonstrated on RTKs, the principles may be applicable to other enzyme families with conserved catalytic domains, though broader experimental confirmation is required for translation.