Individualized treatment risk stratification using histopathology-based genomics prediction in multiple myeloma: A multicenter Study in 1429 participants

Abstract Background: We previously established the first individualized risk prediction model (IRMMa) integrating genomic, clinical, and treatment data to predict clinical outcomes. However, genomic profiling of multiple myeloma (MM) by next-generation sequencing, or FISH testing is time-consuming and resource-intensive, delaying treatment decisions. To address the limited universal access of comprehensive molecular profiling we developed CORAL, a next-generation model to IRMMa that predicts genomic abnormalities from whole slide images (WSI) of routine bone marrow core biopsies and integrates clinical and treatment data to provide individualized and real-time treatment guidance and predict clinical outcomes. Methods We conducted a multicenter international real-world analysis of 1309 consented MM patients and 120 normal controls (i.e. healthy bone marrow samples): University of Miami (UM; n=519), HealthTree Foundation (HT; n=805) and Ministry of Health, Malaysia (MOHM; n=105). We collected Hematoxylin & Eosin (H&E) stained slides from formalin-fixed paraffin embedded (FFPE) tissue block core bone marrow biopsy samples, and scanned them using Aperio AT2 at 40x resolution. Using deep learning methodology, we trained CORAL on the UM cohort to predict seven established genomic subgroups (t(11;14), t(4;14), 1q gain, del 17p, del 13q, del 1p, hyperdiploidy) directly from H&E histopathology images. First, we compared CORAL predictive accuracy (AUC) with existing FISH data in the HT validation cohort. Next, we applied unsupervised clustering to CORAL-derived histomorphologic features from UM and HT to identify novel biological subgroups. Lastly, to assess individualized risks and predict clinical outcomes (based on genomic characterization provided by our model along with available clinical and treatment data), we validated the prognostic performance using concordance-index (c-index) for overall survival (OS) on the MOHM cohort (>3-year follow-up) which lacked genomic testing. Results For the following established genomic subgroups CORAL achieved robust performance (AUC): t(11;14) (0.754), t(4;14) (0.779), 1q gain (0.747), amp 1q (0.836), del 13q (0.767), del 17p (0.759), del 1p (0.723) and hyperdiploidy (0.800). The model identified 12 distinct histomolecular clusters, seven of which were associated with predominant genomic subgroups: t(11;14) , t(4;14), t(4;14) including 1q gain, MAF translocations including t(14;16) and t(14;20), del 1p, del 17p, del 13q, three hyperdiploidy clusters differentiated by secondary alterations, and two clusters without evident genomic alterations. This establishes a histopathology-based molecular classification bridging tissue architecture with genomic complexity. In the MOHM validation cohort, CORAL achieved a high prognostic performance for OS (c-index; 0.7), by filling in the gaps of missing genomics, in conjunction with existing clinical and treatment data. Furthermore, our model revealed five treatment-responsive clusters (TRC) that leverage histomorphologic whole slide image features with therapeutic responses, with significant differences in OS (log-rank p < 0.01) and progression-free survival (PFS) (log-rank p < 0.05) among clusters. TRC1 had the longest overall survival comprising of patients that did not receive autologous stem cell transplantation (ASCT) (mean PFS 26 mo, OS 47.5 mo, ≥complete response rate 33.3%); TRC2 received ASCT and showed the best PFS with resource-limited regimens but a shorter OS (28.9 mo, 32.2 mo, 50%) than TRC1; TRC3 showed limited durable responses (24.6 mo, 25.3 mo, 43.7%); TRC4 and TRC5 showing high risk disease, where TRC4 showed lack of response to intensification resulting in rapid progression (< 6 mo; 21.5 mo, 22.5 mo, and 14.3%), and TRC5 had short benefit from intensification followed by a slower progression (19.9 mo, 25.5 mo, 25%). Conclusions We have developed CORAL, a next-generation model which accurately predicts genomic subgroups from histopathology whole slide images of routine bone marrow core biopsies (H&E staining) using deep learning methodology that integrates clinical and treatment data to provide individualized treatment guidance and predict clinical outcomes. The model's scalable architecture allows for continuous development and dataset expansion, transforming clinical management of patients with MM, and with the potential to expand into other malignancies.