DREM: integrating resistance and domain probabilities to prioritize plant immune receptors

Plant resistance (R) genes, key players in the immune system, have traditionally been identified with HMM-based tools relying on sequence homology, whose sensitivity decreases for divergent or non-canonical resistance proteins. Here, DREM (Deep Resistance gene Evaluator Model) is presented as a deep learning framework that learns the compositional properties of resistance proteins from sequence alone rather than searching for predefined motifs. The framework combines two bidirectional long short-term memory networks with attention: a binary resistance classifier and a multi-label classifier of resistance domains. Validation on experimentally characterized proteins from RefPlantNLR recovered 99.8% of sequences, and on 15,000 held-out proteins DREM outperformed the deep learning predictor PRGminer (Matthews correlation coefficient 0.938 vs 0.820), with about two-thirds fewer classification errors. Across five plant proteomes, DREM identified 4,046 High Confidence or Candidate proteins; canonical domain architecture was confirmed in 75.5%, whereas 992 carried strong DREM signals without detectable Pfam domains, and 54 of 55 characterized nucleotide-binding leucine-rich repeat (NLR) loci ranked within the top 1.7% of their proteomes. Unlike domain-gated pipelines such as R-Predictor, DREM also recovered experimentally validated atypical NLRs, RPW8 proteins and non-receptor resistance genes. Attention did not concentrate on canonical ATP-binding motifs, consistent with the model drawing on sequence information beyond conserved domain signatures. Lightweight and applicable to complete proteomes, DREM complements domain-based annotation with a calibrated ranking. It links genome-wide prediction to the selection of resistance genes for functional validation, including atypical receptors missed by profile-based tools.