What if we stopped asking whether a peptide binds HLA—and started modeling the entire chain from tumor mutation to T-cell recognition?
Cancer immunotherapy has a frustrating computational bottleneck.
A tumor can contain thousands of somatic mutations. Only a fraction are expressed. An even smaller fraction produces peptides that are processed and presented by a patient's HLA molecules. Fewer still are recognized by T-cell receptors (TCRs), and only a subset ultimately produces a meaningful immune response.
The problem is therefore not simply:
Which peptides bind MHC?
The real question is:
Which tumor-derived mutations are most likely to become functional, patient-specific immune targets?
That distinction motivates a different kind of model.
The Idea: A Neoantigen Foundation Model
I propose a Patient-Specific Neoantigen Immunology Foundation Model: a multimodal model designed to learn the complete biological path from tumor genotype to immune recognition.
Instead of treating neoantigen discovery as a collection of disconnected prediction tasks, the model would represent the patient as an integrated immunome.
Its inputs could include:
- Tumor DNA and somatic mutations
- Tumor RNA / transcriptomic data
- Proteomic information
- HLA genotype
- Candidate peptide sequences
- Peptide–HLA structural information
- TCR repertoire, when available
- Potentially clinical and tumor-context metadata
The model would then produce a ranked set of candidate neoantigens with calibrated confidence and interpretable biological evidence.
Tumor DNA
│
▼
Mutation Representation
│
├──────────────► Tumor RNA
│ │
▼ ▼
Candidate Peptides ◄── Expression
│
▼
HLA Representation
│
▼
Peptide–HLA Presentation
│
▼
pMHC Structural Representation
│
▼
TCR Recognition
│
▼
T-cell Immunogenicity
│
▼
Patient-Specific Neoantigen Ranking
Why Existing Approaches Are Not Enough
This is not an empty research area.
Recent systems already demonstrate that deep learning can improve individual parts of the problem.
TransNRank, published in August 2026, uses a Transformer for neoantigen ranking and reports improved recall on NCI, TESLA and HiTIDE datasets. (arXiv)
NeoaPred incorporates structural and surface features of peptide–HLA complexes rather than relying exclusively on sequence information. Its authors reported strong structural prediction performance and improved neoantigen prediction compared with existing approaches. (PubMed Central (PMC))
ImmunoStruct, published in Nature Machine Intelligence, goes further into multimodal modeling by combining sequence, structural and biochemical information for peptide–MHC immunogenicity prediction. (Nature)
Meanwhile, unified models have begun explicitly modeling the relationship between antigen, HLA and TCR. A 2025 Nature Machine Intelligence study introduced a cross-attention architecture for predicting antigen binding specificity to both HLA and TCR molecules. (Nature)
These developments are important.
But they also reveal something interesting:
The field is solving pieces of the same biological pipeline independently.
The opportunity is to connect those pieces.
The Missing Layer: A Patient-Level Representation
Imagine that instead of running a long sequence of independent tools:
Variant Calling
↓
Expression Filtering
↓
Peptide Generation
↓
HLA Binding
↓
Antigen Presentation
↓
Structural Analysis
↓
TCR Prediction
↓
Immunogenicity Ranking
we build a shared representation:
PATIENT IMMUNOME
│
┌───────────────┼───────────────┐
│ │ │
Tumor HLA TCR
Genome Genotype Repertoire
│ │ │
└───────────────┼───────────────┘
▼
IMMUNOLOGY FOUNDATION
MODEL
│
┌─────────────┼─────────────┐
▼ ▼ ▼
Sequence Structure Context
│ │ │
└─────────────┼─────────────┘
▼
Immunogenicity
│
▼
Neoantigen Ranking
This is the conceptual shift.
The model is not merely learning peptide features.
It is learning the relationship between a patient's tumor and immune system.
A Multimodal Architecture
A possible architecture could contain specialized encoders connected to a shared transformer backbone.
1. Genomic Encoder
Encode:
- reference sequence
- tumor mutations
- SNVs
- indels
- fusion events
- variant context
The objective is to learn how genomic alterations can create novel coding sequences.
2. Transcriptomic Encoder
Genomic existence does not necessarily mean biological relevance.
RNA information could provide evidence about:
- whether a mutation is expressed
- transcript abundance
- allele-specific expression
- alternative splicing
- tumor-specific transcription
This allows the model to distinguish:
mutation present
from
mutation biologically active.
3. HLA Encoder
HLA diversity is one of the fundamental challenges.
The model should represent HLA-I and potentially HLA-II alleles in a sequence-aware manner rather than treating HLA as a categorical label.
This could enable better generalization to rare or poorly represented alleles.
4. Peptide–HLA Structural Encoder
Sequence alone is not enough.
The geometry of the peptide inside the HLA binding groove matters.
NeoaPred demonstrated the value of explicitly modeling structural and surface properties of peptide–HLA complexes. (PubMed Central (PMC))
A next-generation architecture could therefore combine:
sequence embeddings + contact maps + geometric representations + biochemical features.
A geometric GNN or SE(3)-equivariant network could be particularly interesting here.
5. TCR Encoder
This may be the most important extension.
The immune system does not recognize an HLA-bound peptide in isolation.
T cells recognize the pMHC surface through TCRs.
Therefore, if patient-specific TCR repertoire data are available, the model could ask:
Which TCR populations are computationally compatible with this neoantigen?
This transforms the problem from:
neoantigen prediction
into:
neoantigen–immune repertoire matching.
Recent research already demonstrates the importance of jointly modeling antigen, HLA and TCR interactions. (Nature)
Multi-Task Learning Instead of One Prediction Head
The foundation model should not have a single output.
It could be trained through multiple related objectives:
Foundation Model
│
┌──────────────────┼──────────────────┐
▼ ▼ ▼
HLA Binding Antigen Presentation pMHC Stability
│ │ │
└──────────────────┼──────────────────┘
▼
TCR Recognition
│
▼
Immunogenicity
│
▼
Clinical Prioritization
This creates a powerful inductive bias.
The model learns that these biological events are not independent.
They form a causal-like chain.
From Classification to Ranking
A practical system should ultimately answer:
Which 10 neoantigens should we investigate first?
Not:
Is this peptide immunogenic: yes/no?
The output could therefore look like:
| Rank | Neoantigen | HLA | Expression | Presentation | TCR compatibility | Immunogenicity |
|---|---|---|---|---|---|---|
| 1 | Candidate A | HLA-A*02:01 | High | High | High | 0.94 |
| 2 | Candidate B | HLA-B*07:02 | High | High | Medium | 0.88 |
| 3 | Candidate C | HLA-A*03:01 | Medium | High | High | 0.84 |
The critical point is that the final score should not be a black-box number.
The model should explain why a candidate ranks highly.
Uncertainty Must Be a First-Class Output
There is another problem with applying foundation-model thinking to biology:
confidence is not truth.
A clinically useful system should distinguish between:
- high-confidence prediction
- uncertain prediction
- out-of-distribution prediction
- insufficient experimental evidence
For example:
Neoantigen A
Immunogenicity: 0.91
Confidence: 0.94
Evidence:
✓ High expression
✓ Strong HLA presentation
✓ Structural compatibility
✓ TCR repertoire support
Neoantigen B
Immunogenicity: 0.87
Confidence: 0.41
Warning:
⚠ Rare HLA allele
⚠ Limited training representation
⚠ No matching TCR evidence
This is essential if the system ever moves beyond computational research.
The Real Foundation-Model Question
Calling something a "Foundation Model" should mean more than putting a Transformer behind a biological dataset.
A real foundation model should ideally support:
Pretraining → adaptation → multiple downstream tasks
For example:
Large-scale biological pretraining
│
▼
Neoantigen Foundation
Model
│
┌──────────┼──────────┐
▼ ▼ ▼
Cancer Vaccine TCR
Targets Design Discovery
│ │ │
└──────────┼──────────┘
▼
Personalized
Immunotherapy
The same representation could eventually support:
- neoantigen vaccine design
- TCR discovery
- adoptive T-cell therapy
- cancer immunotherapy research
- infectious disease antigen discovery
- personalized immune profiling
What Would Make This Scientifically Interesting?
The novelty should not be claimed as:
"We invented AI for neoantigen prediction."
That is already false.
A stronger research hypothesis is:
Can a unified multimodal representation of tumor genotype, expression, HLA presentation, pMHC structure and TCR repertoire improve patient-specific neoantigen prioritization compared with independent prediction pipelines?
That is a much more interesting question.
And it is experimentally testable.
The Benchmark I Would Build
A serious project needs a benchmark designed around the complete pipeline.
Instead of randomly splitting peptides, I would evaluate difficult generalization settings:
Patient-level split
No leakage between patients.
Tumor-type split
Train on some cancer types and evaluate on unseen ones.
HLA-allele split
Test generalization to rare or unseen HLA alleles.
TCR split
Evaluate whether the model can generalize to unseen TCR sequences.
Temporal validation
Train on historical datasets and evaluate on later experimental discoveries.
And most importantly:
Experimental validation
The final metric should not simply be AUROC.
The question is:
How many experimentally validated immunogenic neoantigens appear in the top K predictions?
That is the metric that matters.
The Bigger Vision
AlphaFold changed structural biology by showing what happens when a difficult biological inference problem is transformed into a scalable machine-learning problem.
The next opportunity may not be another model that predicts a single biological interaction.
It may be a model that learns the entire information flow of tumor–immune recognition.
From:
DNA
to:
RNA
to:
protein
to:
peptide
to:
HLA presentation
to:
pMHC structure
to:
TCR recognition
to:
T-cell immunogenicity.
That is the vision behind a Neoantigen Immunology Foundation Model.
Not another predictor.
Not another ranking algorithm.
A computational representation of the patient's tumor–immune interface.
And perhaps the most interesting question is not:
Can AI find neoantigens?
It is:
Can AI learn why a patient's immune system recognizes one mutation as a threat—and ignores another?
That is the problem worth solving.
Research Direction
A practical first MVP would deliberately avoid trying to solve everything at once:
Phase 1 — Neoantigen Foundation Backbone
DNA/RNA + HLA → candidate peptide representation.
Phase 2 — Multimodal Immunogenicity
Add peptide–HLA structural and biochemical information.
Phase 3 — TCR Integration
Add TCR sequence/repertoire information.
Phase 4 — Patient Immunome
Create a unified patient-level representation.
Phase 5 — Experimental Loop
Prediction → laboratory validation → feedback → model improvement.
The end goal is not to replace experimental immunology.
It is to make the search space dramatically smaller.
Millions of possible biological interactions → hundreds of computational candidates → a small number of experimentally testable targets.
That is where foundation models could become genuinely useful in cancer immunology.
This is a research concept, not a clinical diagnostic or treatment system. Any clinical application would require rigorous experimental and prospective validation.
created by Seyed Alireza Alhosseini Almodarresieh
Top comments (0)