Pik-1 HMA Protein Language Model

Rice blast, caused by the fungus Magnaporthe oryzae, destroys enough rice each year to feed roughly 60 million people. The rice immune receptor Pik-1 detects the fungal effector Avr-Pik through an integrated heavy metal-associated (HMA) domain, but the pathogen has evolved effector variants — Avr-PikC and Avr-PikF among them — that no characterized Pik-1 allele recognizes. Finding rice varieties whose Pik-1 alleles already bind these escape variants would be valuable for breeding, but there are thousands of sequenced rice genomes and testing each candidate allele in the lab is slow.

This work, from Pamela Ronald's group at UC Davis with the Joint BioEnergy Institute and the Innovative Genomics Institute, closes that gap by training a model on directed evolution data rather than on curated variant annotations. A yeast-surface-display library of roughly 2×10⁷ Pikh-1 HMA variants — averaging 2.1 substitutions across the 78-residue domain — was sorted by fluorescence for binding to Avr-PikC or Avr-PikF and sequenced. The change in each variant's abundance before and after sorting gives an enrichment score, and ESM-2 is fine-tuned end-to-end to regress that score from sequence.

The point of the exercise is transfer. The fine-tuned models were pointed at Pik-1 alleles mined from the 3,000 Rice Genomes Project — sequences the library never contained — and two of their top-scoring predictions were confirmed by in vitro binding assays.

#Key Features

• Trained on selection data, not annotations: Enrichment scores from a sorted display library supply hundreds of thousands of quantitative labels for a single domain, which is what makes fine-tuning a language model viable at this scale of locus-specific data.
• Full fine-tuning beats frozen embeddings: Backpropagating into the language model weights outperformed CatBoost, ElasticNet, and support-vector regressors trained on the same model's fixed embeddings.
• Generalizes to mutation types absent from training: The library contained no insertions or deletions, yet the models consistently and correctly scored natural Pik-1 alleles carrying an HMA-domain insertion as strong Avr-PikC binders.
• One model per ligand: Separate regressors for Avr-PikC and Avr-PikF give a per-effector recognition profile rather than a single binding score.
• Portable recipe: The identical pipeline applied to a deep mutational scan of the human enzyme NUDT15 predicted thiopurine cytotoxicity risk, including for indel variants the original assay could not measure.

#Technical Details

The backbone is the 8-million-parameter, six-layer ESM-2 checkpoint, loaded with a single-output sequence-regression head and fine-tuned for 20 epochs at a 1e-5 learning rate with 0.01 weight decay, selecting the checkpoint with the best validation Spearman correlation. Data splits were made both randomly and by clustering in the embedding latent space. Both the Avr-PikC and Avr-PikF models exceeded a Spearman correlation of 0.85 on held-out validation sequences.

Applied to the 3,000 Rice Genomes Project, read alignment identified 119 rice varieties with full coverage across the HMA domain, yielding 13 unique variants of which 11 had no published binding phenotype. Ten scored positively for Avr-PikC binding and none for Avr-PikF; the Pikh-1 and Pikp-1 alleles were scored negatively for both, matching prior experimental characterization. Two insertion-carrying variants, from the Vellai Kolomban and Sanhuangzhan-2 cultivars, were displayed on yeast and assayed by microscopy and flow cytometry: both bound Avr-PikC more strongly than wild-type Pikh-1, with Sanhuangzhan-2 the strongest. The NUDT15 model reached a Spearman correlation of 0.76 and correctly called all three clinically cytotoxic indel variants as nonfunctional.

#Applications

For rice breeding, the model converts a genomic database into a ranked list of germplasm worth phenotyping for blast resistance — Sanhuangzhan-2 is already used as a resistance donor, and this analysis suggests its Pik-1 allele may be part of why. More generally, the approach applies wherever a display or deep-mutational-scanning assay can be run on a single domain: the resulting model extrapolates from an engineered mutational library to the natural variation sitting in population-scale sequencing datasets, which is the direction most clinical and agricultural variant interpretation actually needs.

#Impact

The contribution is methodological and deliberately narrow. Each model is fit to one domain and one ligand, and the paper does not claim a general-purpose binding predictor; what it demonstrates is that a small language model, fine-tuned on selection data from an engineered library, extrapolates accurately to naturally evolved sequences and to mutation classes it never saw. The wet-lab confirmation is in vitro binding, not immune activation in planta, so whether the improved affinity translates into blast resistance in the field remains untested. The full analysis pipeline — data formatting, fine-tuning, variant calling, prediction, and the NUDT15 replication — is published as a GitHub repository, though it carries no license file and the trained weights themselves are not distributed. The preprint has not been peer reviewed.