Completing the Chen–Mellman Cancer-Immunity Cycle
The Chen–Mellman cancer-immunity cycle has organised tumour immunology for thirteen years. Alongside its successes, clinical observations have accumulated that the cycle as drawn cannot resolve: PD-L1 predicts at AUC 0.63, response rates vary from 70% to 2% across cancer types with the same drug, cold tumours stay cold for reasons the cycle cannot name. This paper proposes that the seven Mellman steps are a true description of six of sixteen positions, and that the missing ten are where the unanswered questions live. Across 33 cancer types, only 4 stall at the synapse where anti-PD-1 acts.
The cycle as drawn
In 2013, Chen and Mellman published a diagram in Immunity that has organised tumour immunology ever since. The cancer-immunity cycle described seven steps from antigen release through killing, closing back on itself as killed tumour cells release more antigen. The diagram's contribution was the insight that the anti-tumour immune response is a cycle, that each step can be rate-limiting, and that the rate-limiting step determines which intervention will work.
The framework holds where it applies. It is cited over six thousand times and continues to be the scaffolding on which the field explains its successes.
Seven questions the cycle leaves open
It is also the scaffolding on which the field registers its failures. Seven questions sit on top of the cycle that the cycle cannot answer on its own terms.
Why is PD-L1 a weak biomarker when it measures the brake on the synapse where anti-PD-1 acts? A meta-analysis across 18,792 patients found PD-L1 staining predicts at AUC 0.63 — marginally above chance.
Why do response rates vary so widely across cancer types with the same drug? Anti-PD-1 achieves objective response rates of approximately 40% in melanoma, 70% in Hodgkin lymphoma, 50% in MSI-high colorectal, less than 5% in MSS colorectal, 2% in pancreatic, and 8% in glioblastoma. Same drug, same target.
Why do cold tumours stay cold? The cycle has a mechanism for why T cells are disabled in a hot tumour. It has no mechanism for why T cells fail to arrive in the first place.
Why does CAR-T succeed where checkpoint blockade fails? The cycle cannot explain why replacement succeeds where release fails, because the entire build pathway is compressed into the arrow between step 3 and step 4.
Why do patients respond and then stop responding? Acquired resistance is a loop that has broken at a location the loop does not name.
Why did five independent laboratories converge on CXCL13 as the response marker without a structural explanation for that particular chemokine?
Why did the 2023 Mellman update extend outward into stratification axes rather than inward into new steps? A signal from the field itself that the seven-step sequence has reached the limit of what it can explain as drawn.
The completion
The argument is that all seven have the same answer: the Chen–Mellman cycle describes six of sixteen positions in the complete antigen-specific immune response, and the missing ten are where the unanswered questions live.
The completion is additive. The seven steps Mellman drew remain the seven steps he drew, unmodified, occupying their original meaning in the completed map. What the completion adds is the naming and empirical validation of the positions the original diagram compressed into arrows.
The cycle traverses four regimes — Activation, Construction, Encounter, Conservation — separated by four irreversible barriers: the Activation barrier (commitment to build a response against this antigen), the Egress barrier (built effectors release from the lymph node into circulation), the Drain barrier (sustained killing resolves into stable clearance), and the Suppression barrier (the maintained state generates the next antigen-cycle).
Four empirical results
Position 10 — the cytolytic synapse — is locked by a structurally derived biomarker, measured on two independent axes within one cohort and at two measurement resolutions across cohorts. Within Bassez breast cancer (n=29 pretreatment): CXCL13+ CD8 fraction predicts anti-PD-1 response at CV AUC 0.9746 ± 0.065 as a single feature. The Bassez authors' own CD8_EX label — selected without reference to the framework — predicts at CV AUC 0.9322 ± 0.037. Two independent labelings, one position. Bulk-RNA validation on GSE207422 NSCLC (n=24): the LCAM-T module predicts at CV AUC 0.74. Five additional published cohorts converge on the same marker in the same direction.
Position 8 — the Egress barrier — is locked across two cancer types and confirmed rare in a third. A fingerprint specified before any data was inspected — build complete, egress program absent, reserve intact, exhaustion absent, CXCL13 absent — distinguishes ICI-failing MMRp tumours from ICI-responding MMRd tumours at Fisher OR = 4.64, p = 0.0028 in Pelka 2021 colorectal (76,965 cells). The same fingerprint, applied without modification to the Bassez breast cohort, identifies 12 of 20 (60%) non-responders versus 0 of 9 (0%) responders, one-sided p = 0.0024. Pre-registered before data inspection at <25% match in melanoma, the test on Sade-Feldman 2018 yields 0 of 11 patients matching — the strongest version of the prediction.
The single largest patient bottleneck is in the Construction phase — the lymph-node and tumour-bed work that converts an antigen-aware system into deployment-ready effectors. A blind k-means rebuild of 887 cancer patients across 12 tumour types places 43% of patients in this phase, with two distinct mechanisms: myeloid substrate hijack in the tumour bed preventing effector readiness (n=234), and Treg IL-2 sink in the lymph node interrupting clonal expansion during priming (n=147). Neither anti-PD-1 nor any synapse-acting drug can help these patients.
In most cancer types, the response stalls upstream of where anti-PD-1 acts. A regime classification of 11,373 TCGA samples across 33 cancer types finds that in only 4 of 33 cancer types does the cycle reach the Encounter phase. In the other 29, the response stalls upstream, and PD-1 release does nothing because there is no engaged-but-braked cell to release.
The single most consequential claim
Cancer is not primarily a failure of Encounter. Across 33 cancer types, only 4 stall at the synapse phase where anti-PD-1 acts. The remaining 29 stall earlier in the cycle (8 in Activation, 8 in Construction) or in Treg over-control (13). The decade of immunotherapy effort directed at the killing step has worked precisely in the cancers where the cycle reaches the synapse with the brake on, and has failed precisely where the response stalls upstream.
The next decade should be directed not only at the killing step but at the build step — and in particular at the previously unnamed handshake at Position 8 immediately upstream of the Egress barrier, where built effector cells fail to receive the licensing signal that would let them ship into Encounter at all.
The drugs that act on the upstream regimes mostly already exist. CD40 agonists, IL-2 muteins, anti-VEGF, anti-CSF1R, STING agonists, CAR-T, anti-CCR8. What has been missing is the map from patient to drug across the full cycle.
Read the full paper
The complete paper — including the sixteen-position scaffold, detailed treatment of each position, pre-registered tests, limitations, and the bidirectional ipilimumab/belatacept consistency check from kidney transplant — is deposited on Zenodo.