ASCO GU 2024: Predicting Clinical Outcomes in the S1314-COXEN Trial Using a Multimodal Deep Learning Model Integrating Histopathology, Cell Types, and Gene Expression

(UroToday.com) The 2024 American Society of Clinical Oncology Genitourinary (ASCO GU) cancers symposium held in San Francisco, CA between January 25^th and 27^th was host to a urothelial carcinoma oral abstract session. Dr. Bishoy Faltas presented an artificial intelligence-based, multimodal deep learning model integrating histopathology, cell types, and gene expression to predict clinical outcomes in the SWOG S1314-COXEN trial.

Dr. Faltas noted that predicting pathologic complete response (pCR) prior to neoadjuvant chemotherapy remains an unmet clinical need, despite neoadjuvant chemotherapy followed by radical cystectomy being considered standard of care treatment for patients with muscle invasive bladder cancer (MIBC). It is estimated that approximately 40% of patients with MIBC achieve a pCR, and this pathologic outcome has been demonstrated to positively correlate with overall survival.¹ A reliable predictive biomarker of pCR to NAC can identify patients who are potential candidates for bladder preservation.

The multi-modal deep learning model relies on histopathology image analysis, RNA expression, and spatial cell type data to predict pCR outcomes. Data from the SWOG S1314 clinical trial was used to train a deep learning model. SWOG S1314 was a phase II randomized trial of cT2-4aN0M0 patients planned for radical cystectomy who were randomized to 4 cycles of gemcitabine + cisplatin versus dense dose MVAC (ddMVAC).² For the purposes of training the deep learning model, H&E histopathology and transcriptomic analysis of transurethral resection of bladder tumor (TURBT) tissue specimens from 180 patients were used as inputs to predict the pCR outcomes.
Dr. Faltas and colleagues integrated three neural networks into a predictive interpretable architecture. This model relied on three branches:

A gene expression branch (multi-layer perception)
Embeddings branch that relies on whole slide imaging (ResNet50)
Cell type morphology branch (HoverNet)

When these three branches are combined to produce the final model, it is able to:

Predict pCR
Determine the branch contribution of each feature
Evaluate the biologic significance of specific architectural features

Patients from the SWOG S1314 trial were split into pathologic complete responders (30.8%) and non-responders (69.2%). Next, These patients were split into training (80%) and testing (20%) sets. The entire cohort then underwent 5-fold-cross-validation, dividing the dataset into five equal fold, whereby each fold acts as a testing set, while the remainder of the cohort is held as a training set. This ensures that the model is robust and performs reproducibly across different subsets of the data.
Next, the investigators examined the value of integrating multiple inputs and using deep learning to accurately predict neoadjuvant chemotherapy responses. To accomplish this, the investigators compared the respective area under the curves (AUC) with the incorporation of each additional branch:

Integration of all three branches outperformed each branch alone or any dual combination for predicting pCR, with an AUC of 0.72. This was confirmed on 5-fold-cross validation, with a mean AUC of 0.74.
After establishing that the ‘sum of the whole parts’ was greater than each individual branch alone, which branch contributed the most to outcome predictions? To evaluate this, the investigators used a game theory concept with the Shapley Additive exPlanations (SHAP) values, to determine each ‘player’s’ contribution to the final payout (i.e., pCR prediction).
The gene expression branch showed the highest contribution to integrated pCR prediction (SHAP value: 0.13), followed by the neural embeddings branch (SHAP value: 0.12).
With regards to model interpretability, did this model learn bladder cancer biology? The issue with most deep learning models is the ‘Blackbox’ problem, whereby a deep learning model cannot explain how it makes decisions. However, to counteract this, the investigators proactively designed the model as an ‘interpretable’ model that can provide insight into decision-making.
This interpretable model autonomously learned that expression of biologically relevant genes is critical for pCR prediction, in particular TP63.

TP63 is a transcription factor that is a master driver of basal differentiation. When the investigators performed gene set enrichment analysis of genes prioritized by the deep learning model, they found that basal differentiation was significantly associated with a pCR. This finding correlates well with previous analyses that have demonstrated that basal subtype predicts for pCR with neoadjuvant chemotherapy. TP63 is a transcription factor that is a master driver of basal differentiation
Dr. Faltas and colleagues were next able to develop a spatial map of high prediction value patches for pathologic response. Among patients with a pCR, they determined that the tumor-stromal ratio for such patients was enriched in high attention patches. tumor-stromal ratio predicts pCR
Dr. Faltas concluded that:

They were able to develop a deep learning model to integrate different data inputs encoding distinct aspects of tumor biology to predict pCR in SWOG S1314 patients
Multimodal integration of RNA expression, H&E whole slide features, and spatial cell type features improved pCR prediction
Gene expression was the highest contributing branch to the integrated pCR predictions
This interpretable model autonomously learned biology: the contribution of TP63 and basal differentiation, the importance of tumor-stromal ratio for pCR

Presented by: Bishoy Morris Faltas, MD, Assistant Professor of Medicine, Assistant Attending, Genitourinary Oncology Program, Weill Cornell Medicine, New York, NY

Written by: Rashid Sayyid, MD, MSc – Society of Urologic Oncology (SUO) Clinical Fellow at The University of Toronto, @rksayyid on Twitter during the 2024 American Society of Clinical Oncology Genitourinary (ASCO GU) Cancers Symposium, San Francisco, CA, January 25^th – January 27^th, 2024

References:

Login