BGPT: Paper Review: T cell receptor specificity landscape revealed through de novo peptide design

Explore by Goal

Quick Explanation Copied

Core claim (paper): A structure-based, 3D-equivariant model (HERMES) can score TCR–peptide–MHC-I interactions and drive de novo peptide design that activates T cells in three TCR–MHC systems, and can be used to quantify how TCRs constrain the peptide recognition “landscape” (degeneracy via peptide-entropy).

Evidence presented includes (i) affinity correlation for two SPR datasets (Spearman up to 0.72) and (ii) activity correlations for multiple peptide mutational scans, plus (iii) de novo design validation with success rates up to ~50% in the best system.

Primary source: PNAS: 10.1073/pnas.2504783122

Long Explanation

Paper Review (evidence-based, skeptical): “T cell receptor specificity landscape revealed through de novo peptide design”

PNAS (published Oct 16, 2025). DOI: 10.1073/pnas.2504783122

Key mechanism (as stated): HERMES predicts per-residue amino-acid “energies/logits” from local 3D structural environments and uses them to compute “peptide energy” for a given TCR–MHC complex; design uses likelihood matrices (HERMES-fixed) or repeated local peptide packing/relaxation with PyRosetta and MCMC-style acceptance (HERMES-relaxed).

1) Visual synthesis: what the paper measured

Correlation of HERMES peptide energy with SPR affinity (TCR–pMHC-I) in two systems (1G4/NY-ESO, A6/Tax variants).
Correlation of HERMES peptide energy with T cell activity (EC50 / reporter readouts) across multiple peptide mutational scans (viral CMV, melanoma gp100/self, p53 R175H neoepitope via bispecific Fab context).
De novo peptide design validated experimentally in three TCR–MHC systems using Jurkat T cells expressing the paired TCR and measuring NFAT-eGFP activation; reported validation success rates up to ~50% in the best system (NY-ESO/1G4/HLA-A*02:01).
“Specificity landscape” quantified via peptide entropy from predicted amino-acid likelihood matrices (HERMES-fixed) to estimate peptide degeneracy per TCR–MHC and compare to MHC-only motif degeneracy.

2) Plotly figure A — Affinity prediction: HERMES-relaxed vs alternatives (summary)

Data points correspond to the paper’s reported Spearman correlations for affinity tasks: 1G4–NY-ESO shows HERMES-relaxed ρ=0.72 and HERMES-fixed ρ=0.29; and A6–Tax shows HERMES-relaxed ρ=0.56 and HERMES-fixed ρ=0.63, with top baselines reported as 0.83 and 0.88 respectively in the paper’s benchmarking table.

3) Plotly figure B — De novo design validation accuracy vs edit distance constraints (reported)

The paper reports an overall ~50% accuracy for NY-ESO designs; it also states HERMES-fixed achieves ~70% accuracy within a smaller subspace (≤5 substitutions), while HERMES-relaxed explores farther (4–7 substitutions) with reduced accuracy (~24%).

4) Plotly figure C — Specificity landscape: peptide entropy (TCR–MHC vs MHC-only; reported medians)

The paper reports median peptide entropy ~10 bits for typical TCR–MHC recognition (estimating ~10^3 peptides) and median ~29–31 bits for MHC-only presentation motif degeneracy (estimating ~2×10^9 peptides in humans and ~5×10^8 in mice).

5) What is well-supported (and why it matters)

5.1 Evidence chain from structure→score→correlation

Strength: The paper explicitly compares HERMES-fixed vs HERMES-relaxed and reports that HERMES-relaxed can better track SPR affinity in at least one affinity benchmark, while HERMES-fixed can track several activity readouts.

Why it matters: If a structure-based model transfers across diverse epitopes without being trained directly on TCR–pMHC labeled data, that supports a “physical inductive bias” hypothesis. The paper argues this via performance despite not being directly trained on TCR–pMHC data (as stated in the abstract).

5.2 De novo peptide design hits a biologically relevant regime (edit distance)

Strength: The experimental design validation uses multiple systems spanning viral and cancer contexts and reports peptide substitutions up to five positions from wild-type that can still activate T cells (best system).

6) Skeptical critique: main limitations / failure modes

6.1 Template anchoring + limited sampling can bias peptide-entropy “degeneracy”

The peptide-entropy approach is explicitly tied to HERMES-fixed likelihood matrices generated from a TCR–MHC structural template and assumes peptide backbone conformations match the template (HERMES-fixed); the paper itself calls this a lower bound on true degeneracy and notes backbone sampling limitations can under-estimate degeneracy.

6.2 AF3/TCRdock filtering introduces selection bias (what you look at can constrain what you find)

The paper uses PAE-based thresholds from structure prediction/docking to filter designs; it also reports that using AF3 structures can degrade performance when the template is poor (e.g., specific bispecific antibody misdocking). This means failure modes may be dominated by the structural filter rather than the learned scoring itself.

6.3 Generalization beyond the small set of studied TCR–MHC scaffolds remains uncertain

The study’s wet validation focuses on three design systems and a limited set of affinity/activity panels, and it explicitly states that broader datasets are needed to draw definitive generalizability conclusions.

7) Practical takeaways for a computational immunology user

Design strategy: Use HERMES-fixed when a good template exists and you expect backbone pose similarity; use HERMES-relaxed to expand sequence exploration but expect lower accuracy as edit distance grows.
Model evaluation: Treat PAE/docking filters as “model-dependent selection mechanisms,” and verify whether your conclusions persist under alternative structural filtering settings.
Specificity quantification: The entropy-based degeneracy results are best interpreted as conservative lower bounds due to backbone sampling constraints.

8) Suggested follow-up analyses you can run in BGPT

Author reviews (bespoke BGPT links)

Feedback:

Updated: April 10, 2026