Software picks the 34 neoantigens in Moderna and Merck's cancer vaccine

Start with the arithmetic of the cut. A patient's tumor hands the software thousands of candidate mutations, and 34 of them get encoded into that person's dose ^[5]. At a thousand candidates that is 3.4 percent; at two thousand, 1.7 percent ^[9]. Between roughly 97 and 98 percent of the material is discarded by ranking rules nobody outside the two companies has described. Alex Rubinsteyn, a cancer vaccine researcher at the University of North Carolina, told STAT he has no idea what is in the algorithm, and that everyone in the field calls their own version proprietary intelligence ^[3].

That discard rule is where the therapeutic hypothesis lives: 34 bets on which mutations a given immune system can be taught to see.

Which is why the reported design matters more than the enthusiasm. STAT's two items report the late-stage success without describing the comparator arm or the size of the benefit ^[4], and two explanations remain compatible with a win. Some researchers think the selection algorithm is the treatment's secret sauce ^[11]. Some skeptical Wall Street investors think the personalization is not doing the work at all, and that mRNA itself primes tumors to respond better to Keytruda ^[12]. Separating those requires an arm that holds the platform constant and varies the selection. A trial of the whole pipeline, however large, cannot do it.

The same STAT briefing carried a conventional file for contrast: the FDA's approval of Revolution Medicines' Rasonque in pancreatic cancer, one molecule, with a pivotal trial showing median survival of 13.2 months against 6.7 on chemotherapy in previously treated patients ^[7], a gain of 6.5 months ^[10], priced at $39,800 a month ^[8]. One substance, one dossier, one number to argue about. Intismeran autogene, if cleared, ships a different molecule to every patient ^[1]. The object under review becomes the procedure that writes the sequence: code, model weights, training data, and some rule for when retraining the ranker counts as a manufacturing change that needs fresh clinical evidence.

The thing this doesn't tell you is whether the ranking is any good patient by patient, because a trial-level result cannot show whether the responders got better-chosen peptides. My reading, held loosely: 34 chosen from thousands is too aggressive a filter to be incidental, so the algorithm probably matters. Nothing reported so far distinguishes its contribution from the platform's, and the test that would costs one extra arm.