A Sensitive Machine Learning-Based Approach to Assess Multiple Myeloma t(11;14) Genetic Subtype from Histopathology Images

Background: Multiple myeloma (MM) encompasses several cytogenetically distinct plasma cell neoplasms, in which the overexpression of anti-apoptotic BCL-2 family proteins facilitate the survival and accumulation of myeloma cells in the bone marrow. Venetoclax (Ven) is a highly selective, potent oral BCL-2 inhibitor that is currently being evaluated as a targeted therapy for the treatment of t(11;14)-positive relapsed/refractory MM (RRMM). t(11;14) is present in 15%-20% of individuals with MM and has recently been shown to be associated with lower overall survival and response rates to current therapies compared with other standard risk markers, indicating a need for additional therapeutic options (Bal et al. Br J Haematol. 2021;195:e113-e116). t(11;14) MM has been shown to retain unique features of B-cell biology; however, pathologist evaluation of t(11;14) has indicated that only 50% of t(11;14)-positive cases have distinctive lymphoplasmacytic morphology (Hoyer et al. Am J Clin Path. 2000;113:831-837; Garand et al. Leukemia. 2003;17:2032-2035). Here, we report a machine learning (ML)-based approach that is highly sensitive in determining t(11;14) status in digitized biopsy slides from a Ven clinical trial dataset.

Methods: Bone marrow core biopsies from patients with RRMM who had known t(11;14) status were stained with hematoxylin and eosin (H&E) and the immunohistochemistry (IHC) marker CD138, a protein expressed on the plasma cell membrane. Slides were digitized and scanned at 40× (Aperio). For IHC model development, 213 slides (61 positive, 152 negative) passed quality control for sufficient tumor regions free of artifacts. First, a color-thresholding model identified regions of CD138 staining; these regions were then used to train a minimally supervised end-to-end (E2E) model for prediction of t(11;14) status. For H&E model development, 218 slides (64 positive, 154 negative) passed quality control. An initial tissue segmentation model was trained from pathologist annotations of regions of myeloma; these regions were used in E2E training for status prediction. The distribution of positive and negative cases was balanced across training, validation, and test sets. To reduce the number of false- negative cases, a combination approach was used in which cases that were predicted by both the IHC and H&E model as negative for t(11;14) were recorded as negative, and cases predicted as positive by either model were recorded as positive. Model performance was assessed using F1 score, which is a weighted average of the precision and recall values; Cohen's Kappa statistic, which measures agreement between actual and assigned classes, taking chance agreement into account; and area under the ROC (receiver operating characteristics) curve (AUC), which measures separability of classification.

Results: The sensitivity for predicting t(11;14)-positive cases was robust for both models, with values of 0.75 for the IHC model, 0.67 for the H&E model, and 0.75 for the combination of IHC and H&E models (Table). Both the validation of IHC and H&E models showed a high predictive value for detecting t(11;14)-negative samples. In the validation and test sets, respectively, the IHC model yielded values of 0.87 and 0.88, and the H&E model yielded values of 0.86 and 0.84. Combination of IHC and H&E model predictions resulted in the best performance in the validation set, a predictive value 0.96 for t(11;14)-negative cases. Positive predictive values were not as high, with values of 0.56 and 0.53 for the IHC and H&E models, respectively.

Conclusions: This ML-based model shows potential in classifying t(11;14) MM, with high predictive value for t(11;14)-negative cases and high sensitivity for detecting t(11;14)-positive cases. Both the IHC and H&E models can accurately predict the majority of positive and negative t(11;14) cases. These results suggest that this ML algorithm has high sensitivity for use as a screening tool. Additional optimization to improve specificity of the ML-based model could enable use as a diagnostic for t(11;14) status. This approach further demonstrates that t(11;14) MM possesses unique biological and histopathological features and warrants further development to potentially identify patients for targeted therapies like Ven that has been shown to be effective in this subgroup of patients. Clinical trials evaluating Ven in t(11:14)-positive RRMM are ongoing.

Stanford-Moore:PathAI: Current Employment, Current equity holder in private company. Liu:PathAI: Current Employment, Current equity holder in private company. Akiti:PathAI: Current Employment, Current equity holder in private company. Behrooz:PathAI: Current Employment, Current equity holder in private company. Rahsepar:PathAI: Current Employment, Current equity holder in private company. Glass:PathAI: Current Employment, Current equity holder in private company, Current holder of stock options in a privately-held company. Shrotre:AbbVie: Current Employment, Current equity holder in publicly-traded company. Ross:AbbVie: Current Employment, Current equity holder in publicly-traded company. French:AbbVie: Current Employment, Current equity holder in publicly-traded company.