Artificial Intelligence Algorithm Predicts Response to Immune Checkpoint Inhibitors.

Area of Science:

• Oncology
• Computational Pathology
• Immunotherapy

Background:

• Immune checkpoint inhibitors (ICIs) have transformed cancer care, but patient response varies significantly.
• Predicting ICI response and adverse events remains challenging due to a lack of generalizable biomarkers.
• Previous work established a supervised machine learning (ML) model for ICI response in metastatic melanoma.

Purpose of the Study:

• To validate and expand the generalizability of a supervised ML algorithm for predicting ICI response in larger melanoma cohorts.
• To develop a self-supervised ML model to identify histologic features associated with patient survival after ICI treatment.
• To investigate ICI response and survival in both adjuvant and metastatic melanoma settings.

Main Methods:

• Analysis of pretreatment hematoxylin and eosin slides from 639 stage III/IV melanoma patients treated with ICIs (anti-CTLA-4, anti-PD-1, or combination).
• Testing the generalizability of a supervised ML algorithm on a metastatic melanoma cohort.
• Developing a self-supervised ML model to correlate histologic morphologies with progression-free and overall survival.

Main Results:

• The supervised ML algorithm achieved an AUC of 0.72 in predicting ICI treatment response.
• A deep convolutional neural network classified patients into high and low risk groups for progression-free survival (P < 0.0001).
• Novel associations were found between epithelioid histology, low tumor-stroma ratio, and improved survival following ICI treatment.

Conclusions:

• The developed ML algorithm demonstrates generalizability in predicting ICI treatment response for metastatic melanoma.
• This study provides the first identification of specific tumor histologic features associated with overall survival in patients receiving ICIs.
• These findings pave the way for integrating ML-based biomarkers into clinical practice for personalized melanoma treatment.