Pembrolizumab (PD-1 blockade) + AI-derived iTIL/“tumor content” dynamics predict survival in rare-tumor basket trial

Pembrolizumab is an anti–PD-1 immune checkpoint inhibitor evaluated in a Phase II, multi-cohort rare-tumor basket study (n=84) with paired baseline (256) and on-treatment (248) biopsies analyzed on H&E using a deep learning TME tool (Lunit SCOPE IO) to quantify intratumoral TIL density (iTIL) and tumor content (TC). In “high-iTIL” tumors, baseline iTIL ≥60 cells/mm² associated with better PFS (HR 0.49, 95% CI 0.25–0.99; p=0.046), but this was not significant in the overall cohort (HR 0.62; p=0.082); on-treatment iTIL increase trended toward improved PFS (HR 0.64; p=0.084) and correlated with improved OS (HR 0.55; p=0.037), while TC decrease associated with prolonged PFS (HR 0.51; p=0.019) and OS (HR 0.54; p=0.042). The combined “iTIL up + TC down” signature strengthened associations with both PFS and OS (HR 0.36 for each; PFS p=0.009; OS p=0.029), suggesting a dynamic, on-therapy histology biomarker that may outperform baseline-only measures; in the broader parent rare-cancer pembrolizumab program, activity is typically modest overall (eg, NPR at 27 weeks ~28% and grade ≥3 TRAEs ~9%), so an enrichment approach that better identifies long-term beneficiaries could be clinically meaningful.