Anti-phage defense gene classifier pairing protein language model embeddings with genomic features to find immune systems outside defense islands.
No providers recorded yet. Browse all providers
Bacteria carry an enormous and still largely uncatalogued arsenal of anti-phage defense systems, and the tools that produced CRISPR-Cas and restriction-modification came out of that arsenal. Most of the systems discovered so far were found by guilt-by-association: defense genes tend to cluster in genomic neighborhoods called defense islands, so uncharacterized genes sitting next to known defense genes are good candidates. The method works, but it is structurally blind to systems that live outside islands — and evidence from genomic-fragment screens suggests those may be the majority.
DefensePredictor, from Michael Laub's lab at MIT, replaces the neighborhood heuristic with a learned classifier. It scores every protein-coding gene in a prokaryotic genome for the probability that it participates in anti-phage defense, using embeddings from a protein language model as its dominant signal. The preprint appeared in January 2025 and the peer-reviewed version was published in Science in 2026.
The result that matters is not the benchmark but the wet lab: 106 predicted transcriptional units were cloned and challenged with phage, and 45 conferred measurable protection — a 42% validation rate against a background where random genes essentially never defend.
Each gene is represented by the mean ESM-2 150M embedding of its protein product, concatenated with the same embeddings for the two flanking genes on either side, giving a 3,200-dimensional protein feature block. A further 119 genomic features — nucleotide and dinucleotide frequencies, GC content relative to the host genome, gene length, intergenic distances, and neighbor orientation — bring the input vector to 3,319 dimensions. A LightGBM gradient boosting classifier is fit on top, with 255 leaves and 782 minimum child samples at a 0.01 learning rate and early stopping on validation average precision. Training labels came from searching about 17,000 taxonomically representative genomes with DefenseFinder, yielding roughly 244,000 positive homologs against about 14 million annotated non-defensive genes, both deduplicated by MMseqs2 clustering.
On the held-out test set DefensePredictor reaches an average precision of 0.95, against 0.38 for a guilt-by-association baseline, 0.33 for HMM-based remote homology search, and 0.70 for cosine similarity between language model embeddings. Its weakness is recall at the family level: for 10 of 40 test system clusters, fewer than half of member proteins clear a 0.5 probability threshold. Applied at a stringent log-odds cutoff of 7.2 across 69 E. coli strains, it called about 2,500 of 321,000 genes as defensive where DefenseFinder found 395.
Microbiology and phage-biology labs use DefensePredictor to turn a newly sequenced genome into a ranked shortlist of defense candidates worth cloning, the step that previously required either a defense island to anchor the search or a laborious genomic-fragment screen. Because 45% of its novel calls carry no domain previously seen in defense, it is also a discovery route toward new molecular mechanisms — the validated systems include a metallophosphatase homologous to the human innate immune protein SMPDL3A, a di-adenylate cyclase, a thymidylate synthase, and five Pfam domains of unknown function. Biotechnology groups mining prokaryotic immunity for new enzymes are the downstream beneficiaries.
The scale results argue that the catalogue of bacterial immunity is nowhere near closed. Scanning 3,000 E. coli and Shigella genomes, DefensePredictor averaged 28 systems per genome against 4.5 for DefenseFinder and turned up 1,374 unique protein clusters harboring a novel predicted defense protein, of which the latest DefenseFinder and PADLOC releases recognize only 4% and 3%. A parallel scan of 500 taxonomically diverse bacterial genomes yielded 3,376 such clusters, and in neither case did the discovery curve saturate. The model ships as an MIT-licensed package on PyPI with a five-fold ensemble of weights archived on Zenodo — an unusually complete release for a paper of this kind. The caveat is the false negative rate: a quarter of known defense protein families are missed at the default threshold, so DefensePredictor complements homology-based annotation rather than replacing it.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.