Why one in five EXPRESSO patients was structurally unreachable by anti-PD-1

Executive summary

• Cohort. 240 metastatic cutaneous melanoma patients on anti-PD-1 monotherapy, pooled across six published trials (Hugo 2016, Riaz 2017, Gide 2019, Liu 2019, Puch-Solis 2021, GSE160638). Pooled ORR 40.4% [34.4–46.7%].
• Finding. A binary structural classification computed ex ante from a five-gene panel separates the cohort into two classes that respond at 21.4% and 48.2% — a 26.8 percentage point gap, χ² = 14.80, p = 0.0002.
• Calibration. Leave-one-sub-cohort-out cross-validation: predicted ORR for the held class is 22.2% (realised 21.4%, error 0.8 pp); predicted ORR for the past class is 49.4% (realised 48.2%, error 1.2 pp). The structural classification produces accurate population-level rates without per-cohort tuning.
• Concentration. 38% of all non-responders in the cohort sit in the 29% of patients classified as held at Activation. The structural call concentrates the failure mode disproportionately into one identifiable group.
• Comparator. The 26.8pp gap exceeds the published gap for PD-L1 IHC (~8–12pp) and TMB (~10–14pp) on the same source cohorts. The classification uses neither feature and requires no per-cohort calibration.
• Trial design implication for synapse-acting drugs. Past-Activation enrolment raises the comparator-arm baseline from 40% to 48% with similar absolute sample size (171 vs 173 per arm at +15pp). Screening burden increases 1.41×.
• Trial design implication for upstream-acting drugs. Activation-held enrolment selects the population where current monotherapy has the lowest expected efficacy (21%) and where upstream-acting drugs (IL-2 variants, dendritic cell agonists, anti-Treg, costimulatory agonists) have the cleanest comparator. Per-arm sample size at +15pp is 143.
• Status. Retrospective hypothesis-generating analysis. Prospective validation in a structurally-stratified Phase 2 has not been run.