mamp-ml

Plants sense pathogens through surface-localized leucine-rich repeat pattern-recognition receptors (LRR-PRRs), which bind short microbe-associated molecular pattern (MAMP) epitopes such as flg22 and csp22 and trigger immune signaling. Comparative genomics has exposed enormous sequence variation on both sides of that interaction — thousands of receptor alleles across crop germplasm, and epitope variants that pathogens use to evade detection — while the assays that establish whether a particular pair actually triggers immunity remain slow and low-throughput. mamp-ml addresses that bottleneck by predicting the immunogenic outcome of a receptor-epitope pair directly from the two sequences.

The model was developed by the Krasileva Lab at UC Berkeley with the Coaker Lab at UC Davis and posted to bioRxiv in July 2025. Rather than pretraining a new backbone, it repurposes the smallest ESM-2 protein language model as a feature extractor and trains a compact classifier on two decades of published functional data, curated by hand into a single labeled corpus.

That framing distinguishes it from structure-based approaches to the same question. Co-folding predictors including AlphaFold3 can generate a receptor-ligand complex, but complex confidence did not separate immunogenic from non-immunogenic pairs in the authors' benchmarking: on a zero-shot CORE-csp22 evaluation, an AlphaFold3 ipTM cutoff recovered 36% of outcomes against mamp-ml's 73%.

#Key Features

• Three-way immunogenicity call: Each receptor-epitope pair is classified as immunogenic, weakly immunogenic, or non-immunogenic, capturing the intermediate responses that a binary binder/non-binder framing loses.
• Structure-aware residue weighting: Receptor residues are up- or down-weighted using B-factor values derived from an AlphaFold2 model of the ectodomain processed through LRR-Annotation, emphasizing the inner concave ligand-binding surface (weight 2.0) over the outer surface (0.5).
• Chemical conditioning: Per-sequence bulkiness, charge, and hydrophobicity are injected into the embedding stream through a FiLM layer, the variant that outperformed sequence-only and naive-concatenation baselines.
• No experimental structure required: Predictions hold up for receptors with no solved structure, since the structural signal comes from predicted models rather than crystallography.
• k-shot adaptation: For receptor families far outside the training distribution, fine-tuning on as few as 32 labeled examples raised accuracy by 22-56 percentage points.

#Technical Details

The backbone is the 8M-parameter ESM-2 checkpoint (esm2_t6_8M_UR50D), with the embedding layer and the first five of six transformer layers frozen so that only the final layer is tuned. Receptor and epitope sequences are embedded jointly, modulated by the B-factor weights and FiLM-conditioned chemical features, then passed to a two-layer MLP head with LayerNorm and 0.2 dropout that emits three logits. Training used AdamW for 15-20 epochs at batch size 8-12 with L2 regularization and label-smoothed cross-entropy.

The corpus comprises 1,339 receptor-epitope interactions extracted from 74 publications spanning 1999-2025, covering 11 LRR-PRRs, epitopes of 10-55 amino acids, and 91 plant species — 684 immunogenic, 454 non-immunogenic, and 207 weakly immunogenic. An 80/20 stratified split gave 1,071 training and 268 test pairs. On that held-out set mamp-ml reaches 73% accuracy, weighted F1 of 0.74, and AUROC of 0.83, with per-class AUPRC of 0.84 (immunogenic), 0.82 (non-immunogenic), and 0.40 (weakly immunogenic). Predictions were checked experimentally with reactive oxygen species assays across Solanaceous species against 65 csp22 variants, where 73% of 120 receptor-epitope combinations were called correctly.

#Applications

The intended use is high-throughput triage: screening natural receptor allelic variation against pathogen epitope diversity to nominate the small subset of combinations worth testing at the bench. That serves plant pathologists mapping the recognition specificity of a receptor family, breeders searching germplasm for alleles that detect a locally prevalent pathogen strain, and synthetic-biology groups engineering ectodomains toward broadened or retargeted recognition. The released pipeline runs receptor sequences through ColabFold and LRR-Annotation before prediction, and is packaged for both Google Colab and local GPU execution.

#Impact

mamp-ml is one of the first sequence-level predictors for plant innate immune recognition, and its main contribution may be the curated corpus itself — a machine-readable consolidation of results scattered across two decades of literature. The limitations are explicit. The weakly immunogenic class remains poorly separated (AUPRC 0.40), zero-shot transfer to receptor-ligand families absent from training degrades sharply (21% on PEPR-pep), and scope is narrow: a single supervised task built on an external pretrained language model rather than its own broad pretraining. The work is a preprint awaiting peer review, and both the manuscript and the code, including bundled trained weights, are released under CC BY-NC, restricting commercial use without a separate agreement.