Using Machine Learning to Predict Immunotherapy Response in Advanced Melanoma

Clin Cancer Res 2021 AI 6 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Page 1
Predicting Who Will Respond to Immunotherapy in Advanced Melanoma

Clinical problem: Immune checkpoint inhibitors (ICI) such as anti-PD-1 and anti-CTLA-4 drugs transform outcomes for a subset of advanced melanoma patients, but only 30-50% respond, and there is currently no reliable pre-treatment biomarker to identify responders versus non-responders.

Study approach: Johannet et al. (2021, Clinical Cancer Research) trained deep convolutional neural networks (DCNN) on hematoxylin and eosin (H&E) stained histology slides from advanced melanoma patients treated with ICIs to predict progression-free survival (PFS) outcomes.

Multivariate classifier: Beyond histology alone, the study combined DCNN image features with clinical variables (tumor stage, LDH level, prior treatment) in a multivariable logistic regression model to create a composite predictor of ICI response.

Key result: The multivariable classifier achieved AUC 0.800 on one scanner platform and 0.805 on another, substantially above clinical-variable-only models, demonstrating that H&E image information contains ICI response-predictive signal not captured by standard clinical staging.

TL;DR: A deep CNN trained on H&E melanoma histology slides combined with clinical variables predicts immunotherapy response with AUC 0.80, capturing tumor microenvironment features invisible to standard clinical staging biomarkers.
Pages 1-2
Study Design, Cohort, and AI Pipeline

Patient cohort: The study included advanced melanoma patients who received ICI therapy (anti-PD-1 or anti-CTLA-4) at NYU Langone Health. Whole-slide images were digitized from pre-treatment biopsy or resection specimens using two different scanner platforms: Aperio AT2 and Leica SCN400.

DCNN architecture: The deep convolutional neural network extracted tile-level image features from H&E slides at multiple magnifications. Features from tiles were aggregated to generate a slide-level representation summarizing the tumor histological phenotype across the entire specimen.

Outcome definition: Progression-free survival was the primary endpoint. Patients were classified as high or low PFS risk using the machine learning model outputs, with Kaplan-Meier analysis used to demonstrate survival stratification in the test cohort.

Cross-platform validation: A key methodological strength was training and testing on images scanned on two different scanner types, demonstrating that the predictive signal was robust to the technical variability introduced by different digital pathology platforms.

TL;DR: The study used pre-treatment H&E whole-slide images from advanced melanoma ICI patients, digitized on two scanner platforms, training a DCNN feature extractor combined with clinical variables to predict progression-free survival.
Pages 2-3
Prediction Performance and Survival Stratification

Classifier AUC: The multivariable classifier (DCNN histology + clinical variables) achieved AUC 0.800 (Aperio AT2 scanner) and AUC 0.805 (Leica SCN400 scanner) for predicting ICI response, substantially outperforming clinical variables alone (AUC approximately 0.63).

Survival curve separation: High-risk and low-risk patient groups identified by the classifier showed statistically significant divergence in Kaplan-Meier progression-free survival curves, with high-risk patients having median PFS substantially shorter than low-risk patients.

Cross-platform reproducibility: The near-identical AUC values across two scanner platforms (0.800 vs. 0.805) confirmed that the predictive features extracted by the DCNN were scanner-independent, a critical requirement for multi-institutional clinical deployment.

Contribution of histology features: Adding DCNN-derived histology features to a clinical-variable model provided the largest single improvement in AUC, suggesting that H&E images contain substantial tumor microenvironment information - including immune infiltrate patterns and stromal architecture - that predicts ICI response.

TL;DR: The combined DCNN-clinical model achieved AUC 0.80 on both scanner platforms, with the histology features providing the dominant predictive contribution above clinical staging alone, and separated patients into significantly different survival outcomes.
Page 3
What the AI Is Detecting in the Histology

Tumor-infiltrating lymphocytes: Histological quantification of TIL density is a known predictor of ICI response. The DCNN likely captures TIL-related features - immune cell spatial distribution, density, and proximity to tumor cells - that are difficult to quantify manually but predictive of immune activation status.

Stromal architecture: Desmoplastic stroma and the physical barriers it creates to T-cell infiltration are visible on H&E slides and may contribute to the predictive signal captured by the DCNN, as stroma-rich tumors tend to be immunologically excluded.

Tumor heterogeneity: The tile-based aggregation approach may capture intra-tumoral heterogeneity patterns - the spatial variation in tumor phenotype across the slide - which correlate with immune escape mechanisms and ICI resistance.

Explainability gap: While the DCNN predicts ICI response, identifying which specific histological features drive the prediction remains challenging; gradient-weighted class activation maps (Grad-CAM) and similar methods are needed to translate predictive tiles into interpretable pathological features.

TL;DR: The DCNN likely captures TIL density, stromal architecture, and tumor heterogeneity patterns from H&E images - all biologically relevant to ICI response - but precise feature identification requires explainability methods beyond simple AUC reporting.
Pages 3-4
Potential Clinical Utility for ICI Treatment Selection

Treatment selection biomarker: A validated H&E-based ICI response predictor could guide treatment decisions: low-risk patients would proceed directly to ICI, while high-risk patients might be prioritized for clinical trials of combination approaches, alternative targets, or molecular biomarker screening.

Routine specimen utility: H&E slides are already generated from virtually every melanoma biopsy, meaning this AI approach adds predictive value with no additional specimen collection, processing, or patient burden - a key practical advantage over dedicated biomarker assays.

Integration with molecular markers: BRAF mutation status, PD-L1 expression, and tumor mutational burden are current biomarkers. Combining H&E AI features with these molecular markers in an integrated prediction model could further improve stratification accuracy.

Validation requirements: Multi-institution prospective validation on geographically diverse patient cohorts with different ICI regimens (anti-PD-1 monotherapy vs. combination) is required before clinical implementation, as the current study cohort is from a single center.

TL;DR: H&E-based ICI response prediction could guide treatment selection using specimens already generated in standard care, and combining these AI features with PD-L1, BRAF, and TMB markers could create a comprehensive clinical-molecular-histological response predictor.
Page 4
Validation Gaps and Next Steps

Single-center cohort: The training and test data came from a single institution (NYU), raising questions about generalizability to other centers with different patient demographics, biopsy practices, and tissue processing protocols.

ICI regimen heterogeneity: The cohort included patients treated with anti-PD-1, anti-CTLA-4, and combination regimens; these drugs have different mechanisms and response patterns, and separate models for each regimen may be needed for maximal predictive accuracy.

Biopsy site variability: Pre-treatment specimens included primary tumors, metastases, and lymph node biopsies - each with potentially different immune microenvironments. Site-stratified analysis or site-aware model training could improve performance.

Next steps: Multi-center prospective studies, explainability analyses to identify the specific pathological features driving prediction, and integration with molecular biomarkers in clinical trial designs are the priority directions for this line of research.

TL;DR: Single-center origin, mixed ICI regimen cohort, and biopsy site heterogeneity require multi-center prospective validation and site-stratified analysis before this H&E-based ICI response predictor can be considered for clinical guideline integration.
Citation: Open Access, 2021. Available at: PMC7785656.