The first immunotherapy prediction framework derived from structural theory

Immune checkpoint inhibitors produce durable remission in 20–30% of patients. The rest experience side effects — sometimes severe — with no clinical benefit. Identifying responders before treatment would reduce unnecessary toxicity, improve trial design, and direct non-responders toward more effective alternatives earlier.

PD-L1 staining is predictive in only 29% of FDA-approved indications. TMB varies wildly. Gene expression signatures (CYT, IFN-γ, GEP) were developed on specific cohorts and fail pan-cancer. TIDE works for melanoma but not lung or bladder. Every model works somewhere and fails somewhere else.

CYT scores how active the immune response is. IFN-γ signature scores how inflamed the tumour is. GEP scores T cell infiltration. None of them ask how fast the tumour is growing. A strong immune response against a slow-growing tumour is a completely different situation from the same response against an aggressive one. The existing scores cannot distinguish them. That is the structural gap FULCRUM fills.

Predicting response is only useful if it changes what happens next. A score that says "this patient will not respond" is not enough — the clinician needs to know why, and what to try instead. If the bottleneck is T cell exhaustion, checkpoint inhibitors may rescue them. If Tregs are dominating, anti-Treg strategies can shift the balance. If the immune system has not recognised the tumour at all, NK cell therapy or oncolytic viruses may be the starting point. FULCRUM was built to provide that level of insight — not to discover new drugs, but to match existing interventions to the specific structural failure in each patient's immune-tumour interaction.

Every other prediction model returns a score. FULCRUM returns a report. Because each component of the formula maps to a specific structural position in the dissipative framework, the output tells you not just whether the patient will respond, but where the immune-tumour interaction is failing and what that implies.

When FULCRUM-S ratio inverts on a dataset (AUC below 0.5), the report classifies the tumour-immune interaction as abundance-limited: the immune system's total output is too low to distinguish composition. All populations — including Tregs and exhausted T cells — co-vary with total infiltrate. The bottleneck is immune presence, not quality. In this regime, FULCRUM-S+ still predicts because it learns that high suppression + high exhaustion = high infiltrate = respond.

When the ratio works (AUC above 0.5), the regime is quality-limited: infiltrate is present and the effector-to-suppressor balance determines outcome. FULCRUM-S ratio is directly informative.

This classification is part of every FULCRUM-S report. It tells you not just whether the prediction works, but why — and what to measure next if it does not.

TCGA Pan-Cancer Atlas, 9,966 patients, 33 cancer types. Cox proportional hazards, median split. The critical comparison is the last row: remove the MKI67 denominator and the score becomes non-significant. The tumour's dissipative output is the entire structural contribution.

In April 2026, FULCRUM was run on the data from EXPRESSO and TIME_ACT — the two largest transcriptomic prediction frameworks for immunotherapy response, both from NCI's Cancer Data Science Laboratory (Ruppin et al.). Same patients, same labels, same cohort definitions. No parameters were adjusted.

EXPRESSO is a supervised LASSO framework trained on 69 cohorts (3,729 patients) using the whole transcriptome with biological priors. Leave-one-cohort-out cross-validation. The largest supervised transcriptomic resource for drug response prediction to date.

TIME_ACT is an unsupervised 66-gene signature of tumour immune activation. It outperforms 22 established signatures across 15 anti-PD1 cohorts (mean AUC 0.76, OR 6.11). Comparison on the 8 publicly available shared cohorts:

Same patients. Same cross-validation folds. Same random seed. Repeated stratified 5-fold CV, 200 repeats. FULCRUM-S+ v1 uses 6 features, v2 adds host Conservation drain as a 7th. Full ML uses all available cell types per dataset (16 to 51). v2 beats full ML on both quality-limited datasets.