AI-assisted iTIL scoring in melanoma lymph node metastases associates with neoadjuvant immunotherapy response

Neoadjuvant immune checkpoint blockade is becoming the standard of care for stage III melanoma, and biomarkers to refine these regimens are urgently needed. Intratumoral tumor-infiltrating lymphocytes (iTIL) are essential for response, but their clinical scoring is hampered by high inter-observer variability, particularly in lymph node biopsies. Here, we present a pathologist-in-the-loop deep learning workflow, to quantify iTIL in hematoxylin and eosin-stained biopsies from melanoma lymph node metastases. Our model was trained on 127,278 pathologist-annotated cells, and externally validated on 159 whole-slide images from three pioneering phase I/II trials testing neoadjuvant ipilimumab and nivolumab in stage III melanoma, with exceptionally long-term follow up (>5 years). AI-assisted iTIL% correlated strongly with pathologist-based iTIL% (Spearman ρ = 0.76, P < 0.0001) and was predictive of major pathological response ( P = 0.0008). A high AI-assisted iTIL% was associated with longer event-free survival (HR = 0.40, P = 0.010), especially when combined with a high tumor mutational burden (TMB; HR = 0.18, P = 0.0056); in the double high subgroup ( n = 33), only one melanoma-related event occurred in 5 years of follow up. This work provides a roadmap towards standardised, AI-assisted iTIL scoring in melanoma lymph node metastases. Prospective studies are indicated to confirm that iTIL metrics, particularly with the TMB, could inform standard of care protocols and future trials.