PEARL

PEARL (Protein Epitope Avoidance by Reinforcement Learning) addresses a failure mode that sits downstream of most protein design pipelines: a designed protein can fold correctly and function well in vitro and still fail clinically because the human immune system recognizes it. Immune recognition through the MHC Class I pathway is a cascade — proteasomal cleavage of the source protein, elution of the resulting peptides, and binding of those peptides to highly polymorphic MHC Class I molecules for display to CD8+ T cells. Prior computational work has typically modeled one stage of this cascade in isolation, and usually across only a handful of HLA alleles, which limits how well those predictions can steer a generative design process.

Developed at MIT by Manvitha Ponnapati, Sapna Sinha, Brian Lynch, Edward S. Boyden, and Joseph Jacobson, and posted to bioRxiv in September 2025, PEARL couples the whole cascade into a single reward signal and uses it to align a protein language model with reinforcement learning. The authors first train their own evidential deep learning predictors for cleavage, peptide elution, and peptide–MHC binding affinity, each of which returns a calibrated uncertainty alongside its prediction. They then supervised-fine-tune ESM Cambrian on the human proteome and align it with group relative policy optimization (GRPO) under a curriculum that progressively raises the number of masked epitope regions and the number of HLA alleles considered.

The result is a generative policy that emits redesigned sequences with reduced predicted epitope load, rather than a scoring function bolted onto an existing design tool. It complements structure-conditioned designers such as ProteinMPNN and backbone generators such as RFdiffusion, which optimize for foldability rather than immune compatibility.

#Key Features

• Whole-pathway reward: A single multi-objective reward integrates proteasomal cleavage probability, peptide elution likelihood, and peptide–MHC binding affinity, instead of optimizing against one stage of antigen presentation.
• Uncertainty-weighted optimization: Every predictor is an evidential deep learning model returning aleatoric and epistemic uncertainty, so confident predictions exert proportionally more influence on the policy gradient and overconfident-but-unreliable signals are damped.
• Human tolerance scoring: A position-specific scoring matrix built from the human proteome measures how closely candidate 9-mers resemble self-peptides, so the reward does not penalize fragments that central tolerance would ignore.
• Curriculum-scheduled GRPO: The CL-GRPO variant ramps up both the number of masked epitope regions and the number of alleles across training, stabilizing optimization as the multi-allele objective becomes harder.
• Allele-specific redesign: Because alleles enter the reward explicitly, the framework can in principle be pointed at a specific patient's HLA genotype rather than only at population-frequent alleles.

#Technical Details

The policy is ESM Cambrian adapted with Low-Rank Adaptation (LoRA), first supervised-fine-tuned on human proteins with random masking of 9–14-residue spans, then aligned by GRPO against the composite reward with nucleus (top-p) sampling and a temperature modifier to preserve candidate diversity. The elution classifier is an evidential Dirichlet model over one-hot peptide and HLA encodings trained on 651,237 unique peptides and evaluated on 41,725 positive HLA peptides against 208,625 sampled decoys across 24 alleles, reaching AUROC 0.963. The binding affinity regressor uses a Normal–Inverse–Gamma evidential head trained on 182,203 peptide–MHC pairs with 9,590 held out, reaching per-allele AUROCs from roughly 0.79 to 1.0. Two cleavage classifiers cover N- and C-terminal contexts, trained on ~1.28 million samples each; only the C-terminal model, at AUROC 0.862, feeds the reward. As a design test case the authors redesigned channelrhodopsin C1C2, a 350-residue light-gated ion channel (PDB 3UG9), generating 64 variants. Scored independently with NetMHCpan across HLA-A02:01, A01:01, A03:01, A24:02, and B07:02, nearly all variants showed fewer strong and weak binders than the parent, with a mean predicted pLDDT of 84.78 indicating the fold was broadly preserved.

#Applications

The immediate use case is deimmunizing therapeutic proteins of non-human origin — the channelrhodopsin case study is a direct proxy for optogenetic tools, whose clinical translation has been limited by immune responses, and the same procedure applies to microbial enzymes, gene editors, and other biologics lacking human sequence analogs. Because the reward is defined over an explicit allele set, the framework also supports population-tailored or patient-specific redesign, and the authors note it extends to MHC Class II presentation and T-cell receptor recognition. The cleavage, elution, and binding predictors are usable on their own as fast, uncertainty-aware epitope screens.

#Impact

PEARL demonstrates that GRPO — the group-relative policy optimization method popularized by reasoning LLMs — transfers to protein sequence design, where it optimizes non-differentiable biological objectives without the pairwise preference datasets that direct preference optimization requires. That framing matters beyond immunogenicity: any composite, non-differentiable protein property could be substituted into the reward. The work carries real limitations. It remains a preprint that has not been peer-reviewed, and all results are computational, with epitope reduction assessed by NetMHCpan and structural integrity by predicted pLDDT rather than by experiment; the authors state that experimental validation is needed to show reduced epitope counts translate into lower immune response in vivo. No code or model weights have been released.