Abstract A057: Multimodal AI Modeling of Clinical Panel-Based Sequencing for HRD Detection in Ovarian Cancer
Bibliographic record
Abstract
Abstract Background: Identifying homologous recombination deficient (HRD) versus proficient (HRP) patients is essential for treatment selection in high-grade serous ovarian cancer (HGSOC), as HRD patients respond favorably to PARP inhibitors. There remains a pressing need for accurate and cost-effective alternatives to reliably determine HRD status and guide therapeutic decisions. Methods: We integrated genomic and digital pathology data from 708 HGSOC patients treated at Memorial Sloan Kettering Cancer Center (MSK). HRD/HRP status was defined by BRCA1/2 mutation status or Myriad MyChoice genomic instability scores. Tumors were profiled using MSK-IMPACT, a targeted next-generation sequencing assay covering 505 cancer-associated genes. Genomic features included mutations (excluding BRCA1/2), copy number alterations, structural variants, pathway activations, HRD-scar metrics, and chromosomal instability signatures. Morphological features were extracted from H&E-stained whole-slide images using GigaPath, a foundation model for histopathology. We trained unimodal and multimodal AI models to predict DNA repair deficiency. Results: Our genomic model built using MSK-IMPACT data achieved exceptionally high performance (AUC = 0.98) in identifying HRD status, highlighting its potential as a reliable and cost-effective alternative to specialized HRD assays. While combining genomic and image features did not improve classification performance, the multimodal model exhibited more compact and separable latent feature spaces, suggesting enhanced representation learning. In survival analysis, predictions from the genomics and multimodal models significantly stratified overall survival (log-rank p = 0.038 and 0.079, respectively), closely aligning with true clinical outcomes (log-rank p = 0.056). The image-only model did not significantly stratify survival (p = 0.39). Morphological features showed added predictive value when combined with specific genomic subsets—particularly structural variants (+23% AUC), mutations (+18%), and copy number alterations (+9%), demonstrating complementary information for specific genomic signals. Conclusion: MSK-IMPACT alone is highly accurate in identifying DNA repair deficiency in HGSOC, achieving performance comparable to commercial HRD assays while offering a more accessible and scalable solution. Although multimodal integration did not enhance classification accuracy, it improved the quality of learned representations and revealed complementary image-based signals in specific genomic contexts. These findings support the role of AI-driven multimodal approaches in enriching interpretability and biological insight in ovarian cancer stratification. Citation Format: Areej Alsaafin, Tom Pollard, Andrew T. Aukerman, Kevin Boehm, Arfath Pasha, Darin Moore, Nicole Rusk, Anika Begum, Mackenzie W. Sullivan, Stephen Graves, Britta Weigelt, Daniel Muldoon, Chaitanya Bandlamudi, Michael F. Berger, Adam Price, Mark Donoghue, Ying L. Liu, Rachel Grisham, Nikolaus Schultz, Francisco Sanchez Vega, Sohrab P. Shah. Multimodal AI Modeling of Clinical Panel-Based Sequencing for HRD Detection in Ovarian Cancer [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr A057.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.002 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.000 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.002 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".