1024MO - Interpretable AI-identified spatial tumor microenvironment (TME) neighborhoods associate with severe immune-related adverse events (irAE) and progressive disease (PD) in patients with melanoma treated with anti-PD-1-based therapy

Background

Among patients with advanced melanoma who receive standard-of-care immune checkpoint inhibitors (ICI; e.g., anti-PD-1 +/- anti-CTLA-4), only ∼50% experience tumor response, and ∼10-55% experience grade ≥3 irAEs. Pre-treatment biomarkers that can help predict efficacy and toxicity outcomes are highly desirable.

Methods

Pre-treatment tumor specimens from 53 patients with advanced melanoma who received ICI were stained with two multiplex immunofluorescence (mIF) panels: Panel 1 (PD-L1, CD8, FOXP3, PD1, CD163, Sox10/S100); Panel 2 (PD1, CD8, GZMB, TBET, EOMES, and LAG3). We used NaroNet, a deep learning artificial intelligence (AI) framework, to identify spatial cellular neighborhoods (125x125μm² image-patches) within the TME. Every morphology, marker expression texture, and interaction was extracted from the images. Toxicity outcomes (grade ≥3 irAEs) and objective response data per RECIST v1.1 were incorporated into the AI algorithm. The AI-identified neighborhoods were visualized and interpreted using the AstroPath platform and tested for an association with clinical covariates.

Results

For each mIF panel, we identified a single neighborhood associated with grade 3/4 irAEs. The first (Panel 1) contained intratumoral CD163+, PDL1(-) cells and lacked FOXP3, PDL1, PD1, and CD8 expression (p=0.008). The second (Panel 2) had PD1+, EOMES+ intratumoral cells without TBET, LAG3, CD8, and PDL1 expression (p=0.04). Notably, these neighborhoods were also associated with PD (p=0.009; p=0.04, respectively). When combined, these two neighborhoods had an AUC of 0.85 (n=8 with PD and severe irAE vs n=45 other), which was irrespective of age, sex, and therapy type (i.e., anti-PD-1 monotherapy or anti-PD-1+CTLA-4).

Conclusions

Our findings suggest that AI-identified, distinct TME neighborhoods could serve as potential biomarkers to identify the subset of patients likely to experience serious irAE and PD after anti-PD-1-based therapy, allowing for consideration of an alternate therapeutic regimen. Additional studies are needed to validate these findings.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Janis M. Taube, M.D., Johns Hopkins.

Funding

The Mark Foundation for Cancer Research (JMT, AS, DJS, EJL); Melanoma Research Alliance (JMT, AS, EW); National Cancer Institute R01 CA142779 (JMT); The Marilyn and Michael Glosserman Fund for Basal Cell Carcinoma and Melanoma Research (EJL), The Barney Family Foundation (EJL), The Laverna Hahn Charitable Trust (EJL), and The Bloomberg∼Kimmel Institute for Cancer Immunotherapy.

Disclosure

E.J. Lipson: Financial Interests, Advisory Board: Bristol Myers Squibb, Novartis, Merck, Genentech, Eisai, Natera, Instil Bio, Nektar, OncoSec, Pfizer, Rain Therapeutics, Regeneron, CareDX, Immunocore, Replimune, HUYA Bioscience International, Sanofi. A.S. Szalay: Financial Interests, Advisory Board: Akoya Biosciences. J. Taube: Financial Interests, Advisory Board: Bristol Myers Squibb, Merck, Genentech, AstraZeneca, Compugen, Regeneron, Akoya Biosciences. All other authors have declared no conflicts of interest.