ESM-DefenseFinder

DefenseFinder, from Aude Bernheim's Molecular Diversity of Microbes lab at Institut Pasteur, is the standard tool for annotating anti-phage defense systems in bacterial genomes. It works by profile HMM search against experimentally validated systems, which makes it precise and also makes it structurally incapable of finding anything unlike what is already known. Across prokaryotic RefSeq it flags only 0.41% of proteins as defensive, and a published list of 7,472 defense-island-enriched families overlaps its profiles by just 38%.

ESM-DefenseFinder is the same lab's attempt to push past that ceiling with a learned classifier. It fine-tunes ESM-2 at the 35M-parameter scale to output a single logit per protein: how likely this sequence is to be part of an anti-phage defense system. Because it reads amino acid sequence rather than genomic neighborhood, it is complementary to the defense-island heuristics that dominate the field — and to its sibling ALBERT-DefenseFinder, which reads genomic context and nothing else. The work was posted as a preprint in January 2025 and published in Science in 2026.

The critical design choice is in the data split. Rather than partitioning protein families at random, the authors split the 137 defense system types recognized by DefenseFinder into disjoint training, validation, and test sets, so every family belonging to Lamassu lands in the same partition. Test performance therefore measures generalization to entire system families never seen in training.

#Key Features

• Sequence-only scoring: A protein can be scored in isolation, without its genomic neighbors, which is what lets the model reach systems sitting outside defense islands and inside mobile elements.
• Detects homology beyond alignment: Test proteins with no detectable sequence or structural homolog in the training set still score significantly above background, so the model is not merely reimplementing a profile search.
• Calibrated decision thresholds: The released tool reports whether each protein clears the best-F1 threshold and the 1% and 0.1% false discovery rate thresholds, so users pick a precision target rather than an arbitrary cutoff.
• Interpretable attributions: Layer integrated gradients map each residue's contribution onto predicted structures, which is how the authors traced high-scoring calls to nuclease and SIR2 domains.

#Technical Details

The backbone is ESM-2 at 35 million parameters, adapted for sequence classification with LoRA at rank 64 and alpha 128, trained for a single epoch at a 1e-5 learning rate after a grid search over rank, scaling factor, and learning rate. The positive class was 132,567 protein families clustered at 80% amino acid identity and annotated by DefenseFinder; everything else in RefSeq served as negatives. Evaluation used all 57,111 positive test proteins against roughly 12 million sampled negatives, mixed to match the 0.45% defense prevalence of the full database and deduplicated at 80% identity so that over-sampled clades cannot inflate the result.

On a held-out set of Actinomycetota protein families whose system types were withheld from training, ESM-DefenseFinder reached 93.6% AUROC and 12.3% average precision, against 88.4% and 7.7% for the genomic-context ALBERT model, 85.9% and 4.0% for a word2vec-style context baseline, and 80.4% and 4.4% for the conventional family-wise defense score. The two learned models score largely orthogonal populations: principal component analysis separates the sequence-based scores from the two context-based ones, and of 23,344 families predicted positive by at least one method, only 463 are called by all three. That orthogonality is what a third sibling, GeneCLR-DefenseFinder, exploits by fusing both signals in one model, reaching up to 99% precision and 92% recall on curated benchmarks in the published version.

#Applications

The model is distributed as an optional extension to the DefenseFinder command-line tool, selectable with a flag, so a microbiology lab already annotating genomes gets defense-candidate scores in the same run without changing workflow. The tool exposes both a 35M and a 650M backbone, trading throughput for sensitivity. Typical use is prioritizing hypothetical proteins in an accessory genome for cloning and phage challenge: in a 2,559-genome E. coli analysis, 8.4% of pangenome clusters cleared the high-score threshold against 0.43% of persistent-genome clusters, and high scorers were enriched sixty-fold for existing defense annotations and six-fold for hypothetical proteins.

#Impact

The scale result is the argument. Applied with its siblings to more than 32,000 bacterial genomes, the study estimates that roughly 1.5% of a typical bacterial genome is devoted to defense and that over 85% of the protein families predicted as defensive remain uncharacterized, publishing a catalog of more than 19,000 candidate operon families and validating 12 anti-phage systems in Escherichia and Streptomyces. An interactive UMAP of this model's embedding space over nearly six million families makes that diversity browsable. The caveat is precision: ablations show the model leans on homology to training-set defenses, so it finds variants and remote relatives of known systems far more readily than unprecedented folds. That is the gap ALBERT-DefenseFinder was built to fill.