1,566 immunotherapy predictors tested. They perform like coin flips.

The literature failed its own test

H. Jiang and colleagues did something the immunotherapy field has needed for years: they re-tested the literature. Collecting 1,566 published response signatures from 1,235 papers, they found the set predicts immunotherapy response with an AUC of 0.515 — statistically indistinguishable from a coin flip. That is 1,566 signatures from 1,235 papers carrying essentially no signal. m:…5764

The preprint, posted to bioRxiv on October 22, also names a mechanism for the failure. TIDE, one of the most-used predictors, misclassifies immune-infiltrated tumors that do respond to therapy as non-responders by misinterpreting exhausted T cells. The field was not just unlucky; its scores were systematically mis-oriented. m:…5764

What the fix looks like

The authors' answer is ResponseVerdict. In 2,000 paired comparisons against TIDE and GEP18, it wins 1,178 and loses none. In blind cohorts it adds 16.2 responders per 100 enrollment slots over existing signatures and 10.2 over TIDE. m:…5764

Those are the authors' own numbers, from a preprint, and they will need independent replication. But the enrollment math is the point a trial designer should notice: if prediction has been coin-flip, every trial that selected patients on signatures has been enrolling noise. A predictor that actually separates responders changes trial economics, not just biology. m:…5764

Prediction is plan A. Control is plan B

Two other bioRxiv preprints from the same day skip prediction entirely and go straight to controlling the cancer cell. A. H. Joechner and colleagues built BRiDGE CAR, a modular peptide-tag adaptor CAR T platform. The design rules they report: a medium-length hinge plus a full-length, bivalent adaptor gives potent activity across multiple antigens. m:…0346

The control features are the interesting part. IL-18 'armouring' lets the cells control tumors at reduced doses; CD33-directed BRiDGE cells retain less-differentiated phenotypes than direct CD33 CAR T cells under repeated challenge; and the cells persist during adaptor withdrawal, regaining anti-tumor activity when the adaptor is re-administered. That is an off-switch and an on-switch — the kind of controllability that prediction-based selection was supposed to make unnecessary. m:…0346

Find the cell's load-bearing gene

A. Navaraj, W. S. El-Deiry and colleagues report five small molecules (NSC667251, NSC636948, NSC21683, NSC281383, NSC136513) that induce cell death in BRCA1-deficient human mammary epithelial cells. m:…6096

Two of the five, NSC636948 and NSC667251, act as inverse agonists of PPAR beta/delta — and high PPARD expression is associated with significantly worse overall survival in breast cancer patients (p=0.008). The result links PPAR signaling to BRCA1-deficient tumor survival and nominates it as a therapeutic target. No in vivo efficacy data is reported, and the mechanisms of the other three compounds are unresolved — both gaps the authors leave open — but the target has human survival data behind it, which is more than most screening hits can say. m:…6096

What changes for trial design

Step back and the three papers tell one story about where oncology is moving. The prediction literature turns out to have been selecting patients on noise; the fix is to orient scores on measured response. But the deeper move is away from prediction toward control — engineered T cells with switches, small molecules aimed at genetic vulnerabilities. For anyone running or funding trials, the question is no longer just which signature to use. It is whether the trial's selection logic was built on the 1,566 signatures that just failed their re-test. m:…5764 m:…0346 m:…6096