The algorithms behind the vaccine: which machine-learning models pick a tumour's 34 targets

In our previous post we described how an individualized mRNA vaccine teaches the immune system to recognise one patient's tumour. One sentence in it deserves an article of its own: out of the hundreds or thousands of mutations found in a tumour, algorithms pick up to 34 neoantigens to go into the vaccine. That choice decides whether the whole therapy works — and no human makes it at a desk. Machine-learning models do.

This article takes them apart one by one. But it starts with an honest caveat, because without it the rest would be misleading.

From a tumour sample to a finished mRNA recipe. At the heart of that path sits a computational model that reduces a crowd of candidates to a handful of real targets.

What we don't know — and why we say so upfront

The exact algorithm used in intismeran autogene (formerly mRNA-4157) has not been published. Moderna describes the selection step as "a proprietary algorithm developed in collaboration with Merck", and in its annual report to the U.S. Securities and Exchange Commission (Form 10-K for 2025) the company writes of "machine-learning based algorithms that design specific therapy for each individual patient and optimize timely manufacture and delivery". That is the whole of what is officially confirmed: machine learning is inside, and the code is closed.

What we do know a great deal about is the class of models that solves exactly this problem — because those are published, peer-reviewed and widely used in neoantigen research. They are what this article is about. Where we describe Moderna's specific product we say so explicitly; everywhere else we describe the methods the whole field rests on.

Step 1: find the typos — variant calling

Everything begins with two sequencing runs: the tumour and healthy tissue from the same patient (usually DNA from blood). A somatic mutation is, by definition, a difference between them.

Sequencing reads are stacked one under another. Where some tumour reads carry a different letter than normal tissue, there is a candidate somatic mutation.

It sounds simple; it isn't. Sequencing makes errors at a measurable rate, and a tumour is a mixture of clones, so a mutation may be present in only a fraction of cells. Variant callers (most commonly Mutect2 from the GATK suite) settle this with a statistical model: they compute the probability that an observed difference is a real mutation rather than a read error. This is not yet machine learning in the everyday sense — it is Bayesian inference — but model quality already decides everything, because whatever is missed here cannot be recovered later.

A second filter is applied to the list: expression. RNA sequencing shows which mutated genes the cell actually transcribes. A mutation in a gene that is silent in this tumour will produce no protein — and therefore no target.

Step 2: learn the patient's locks — HLA genotyping

For a T cell to see a protein fragment, the cell must display it on its surface in an HLA molecule (the human name for the major histocompatibility complex, MHC). HLA is not one shape but a groove whose geometry depends on genotype — and there are several genes (HLA-A, HLA-B, HLA-C), each in two copies, occurring in thousands of variants across the population.

A peptide lies in the groove formed by two helices of the HLA molecule. The groove's shape depends on the allele — which is why the same peptide is presented in one person and not in another.

This is why the KEYNOTE-942 trial read HLA genotypes from exome sequencing of DNA from the patient's blood cells. Without that information, presentation prediction is meaningless: the question is not "is this peptide any good" but "does this peptide fit this patient's locks".

Step 3: the core — antigen presentation models

Here machine learning enters in full. The task: given a peptide (a dozen or so letters) and a patient's HLA molecule, estimate whether that pair will appear on the cell surface at all.

The model takes TWO inputs — the peptide sequence and the sequence of the MHC groove — and returns a score. That is what lets it judge alleles it never saw during training.

The most widely used publicly available model of this class is NetMHCpan (version 4.1). It is an artificial neural network built on the NNAlign_MA framework, trained on 13,245,212 data points covering 250 distinct MHC class I molecules. Two of its properties matter most.

First, pan-specificity. Instead of one model per allele (unworkable with thousands of variants), the model also receives the sequence of the HLA groove itself as input. It therefore learns the general rule of "which groove shape fits which peptide" and can score an allele for which it had not a single training example. That is how it covers more than 11,000 MHC molecules — in practice any patient, including one with a rare genotype.

Second, the kind of data. Older models were trained on test-tube binding affinity measurements. NetMHCpan-4.1 combines those with mass-spectrometry eluted-ligand data: peptides actually stripped from the surface of living cells. That is the difference between "this peptide can bind" and "this peptide was really there".

Step 4: deep learning on what the cell actually displayed

The direction opened by mass-spectrometry data led to models trained almost entirely on it. The flagship example is EDGE (Bulik-Sullivan et al., Nature Biotechnology 2019): a deep network trained on HLA peptides collected from the tumours of 74 patients together with their genomic data.

Peptides are eluted from HLA molecules and measured by mass spectrometry. The result is a training set of antigens genuinely presented, not merely predicted.

The gain is measurable and worth remembering: the model raised positive predictive value up to ninefold over the earlier approach. In practice that means far fewer false alarms on the candidate list handed downstream — and with only 34 slots in the vaccine, every wasted one is a real loss.

The field did not stop there. In use today are MHCflurry, MHCnuggets, MixMHCpred and others, including models built on the transformer architecture familiar from large language models (TransPHLA, MHCRoBERTa) — because an amino-acid sequence is, formally speaking, a sentence in a twenty-letter alphabet.

Step 5: what else has to line up

Presentation is not the only condition. Before a peptide reaches the HLA groove it must be cut out of the protein by the proteasome — separate models predict that cleavage (NetChop, ProteaSMM). Then it must be transported into the endoplasmic reticulum. And finally — the hardest part — it must be recognised by a T-cell receptor, one that this particular patient has in their repertoire at all.

That last step, immunogenicity prediction, is far less mature than presentation prediction. Such models exist and are actively developed, but nobody claims today that the problem is solved.

Step 6: ranking and the cut at 34

The outputs of all these models must collapse into one decision. Candidates are scored — presentation score, expression level, the fraction of cells carrying the mutation, predicted immunogenicity, sometimes similarity to self-proteins (so as not to aim at something also present in healthy tissue) — then sorted and cut at a threshold.

Every candidate gets a score. The top of the list goes into the vaccine — for intismeran, up to 34 entries encoded in a single mRNA molecule.

The scale of the narrowing. Successive sieves — expression, presentation, immunogenicity — leave a handful out of thousands of starting mutations.

The number 34 is neither arbitrary nor purely biological: it is a compromise between breadth of attack and the length of an mRNA molecule that can actually be manufactured, quality-controlled and delivered in reasonable time. That is precisely what the quoted passage from Moderna's report refers to — algorithms that "optimize timely manufacture and delivery". For a patient after surgery, weeks carry clinical weight, so production time is part of the optimisation problem, not a logistics footnote.

Where these models get it wrong

An honest article about algorithms must also say this: predictions are hypotheses, not facts. The literature has repeatedly shown that tools like NetMHCpan and MHCflurry tend to overestimate peptide–HLA binding affinity, so a share of the selected neoantigens elicits no response at all. That is why research work verifies candidate lists experimentally, on the patient's own cells.

Out of a large set of predicted targets, only some provoke a real T-cell response. The model narrows the search — it does not replace the evidence.

This is not an indictment of the method but a description of its role. Without models one would have to test thousands of peptides in every single patient — meaning individualized therapy simply would not exist. Models turn an impossible problem into a feasible one; biology and the clinical trial decide the rest.

What follows from this

The phase 3 success announced on 19 August 2026 is the result of a clinical trial in patients, not a triumph of software — efficacy was settled by patients and statistics, not by a neural network. But without that network there would be nothing to trial: choosing 34 targets from thousands of candidates, repeated separately for every patient and completed within weeks, is a task that cannot be done by hand.

That is why this story is interesting beyond oncology. It shows a mode of collaboration in which the algorithm makes no diagnosis and no clinical decision, but does something humbler and very concrete: it narrows the search space enough for the experiment to become possible.

Sources

1. Reynisson B. et al., NetMHCpan-4.1 and NetMHCIIpan-4.0: improved predictions of MHC antigen presentation by concurrent motif deconvolution and integration of MS MHC eluted ligand data, Nucleic Acids Research 2020 — academic.oup.com
2. Bulik-Sullivan B. et al., Deep learning using tumor HLA peptide mass spectrometry datasets improves neoantigen identification, Nature Biotechnology 2019 — nature.com
3. Moderna, Inc., Annual Report on Form 10-K for fiscal year 2025 (description of machine-learning algorithms in designing and manufacturing individualized therapies) — sec.gov
4. Weber J. et al., Individualized neoantigen therapy mRNA-4157 (V940) plus pembrolizumab in resected melanoma: 3-year update from the mRNA-4157-P201 (KEYNOTE-942) trial, Journal of Clinical Oncology 2024 — ascopubs.org
5. Beyond MHC binding: immunogenicity prediction tools to refine neoantigen selection in cancer patients, Exploration of Immunology 2024 — explorationpub.com
6. Next-generation neoantigen mRNA vaccines: immuno-engineering strategies for precision cancer immunotherapy, Cellular Oncology 2026 — link.springer.com

This article is informational and does not constitute medical advice. Intismeran autogene remains an investigational therapy — treatment decisions are made with your oncology team.