BACKGROUND
Homologous recombination deficiency (HRD) is an increasingly important molecular phenotype
given the development of targeted treatments for HRD positive tumors. While HRD is routinely
assessed in breast and ovarian cancer, prostate cancer patients with homologous
recombination repair (HRR) gene loss may also benefit from targeted therapy. Here, we applied
weakly supervised deep learning to predict HRD status from hematoxylin and eosin (H&E)
stained whole-slide images (WSIs) in prostate cancer as a potential screening assay for
confirmatory sequencing.

DESIGN
Real world prostate cancer tumor biopsy and resection specimens were collected; each sample
additionally included: 1) clinical characteristics, 2) molecular profiles via DNA/RNA sequencing,
and 3) digitized WSI data. The ground truth HRD status for each WSI was generated from an
analytically validated commercial assay that relies on RNA expression. Attention-based
multiple-instance-learning networks were trained to predict HRD status from WSIs. All WSIs
(N=3210, HRD+ 8.1%) were split into training (N=2038, HRD+ 7.8%), tuning (N=455, HRD+
9.8%), and test sets (N=332, HRD+ 8.4%). The test set consisted of needle core biopsies
scanned on Leica GT450, while the training/tuning sets included biopsies and resections
scanned on Leica GT450 or Philips UFS instruments.

RESULTS
The cohort skewed towards high gleason scores (GS) and these high GS samples had a higher
prevalence of HRD (Table 1). Predicting HRD from WSI resulted in an area under receiving
operating characteristic (AUROC) of 0.73 [0.64-0.82 95% CI] on the test set (Figure 1A). At
70% sensitivity, the predictor had a 21.3% PPV across all samples compared to an underlying
prevalence of 8.4%. In additional analysis, we found that the model is able to effectively stratify
patients with BRCA double hit mutations (AUROC: 0.77 [0.68-0.84 95% CI], Figure 1B).
Performance remained robust in GS7-8 (AUROC 0.81, PPV 23.5% at 70% sensitivity) and
GS9-10 (AUROC 0.68, PPV 16.7% at 70% sensitivity). Visualization of high attention
HRD+/HRD- tiles suggest the model relies on dense tumor regions in making its predictions
(Figure 2).

CONCLUSION
An imaging-based model effectively predicts HRD status and BRCA double-hit mutations from
routine H&E images. This low-cost and rapid detection capability could help to prioritize tissue
for confirmatory molecular testing and better identify populations that may benefit from existing
therapy or clinical trial enrollment eligibility.

VIEW THE PUBLICATION

VIEW THE POSTER