DeepPEP

Knowing which of a bacterium's proteins it cannot live without is what turns a genome sequence into a list of antibiotic targets — but that knowledge comes from transposon mutagenesis and CRISPR screens, which have been run on only a few dozen organisms. The standard way to extend it to a new genome is orthology: Geptop 2.0 and similar tools map an unannotated protein onto a homolog in a reference organism and inherit its essentiality call. That works well for the conserved core, and by construction it cannot work at all for a protein with no homolog in the reference set. In Pseudomonas protegens Pf-5, 528 of the strain's proteins have no detectable homolog among 63 reference genomes; 21 of those are essential, and an orthology tool finds none of them.

DeepPEP, from the College of Information Science and Technology at Beijing University of Chemical Technology, replaces the homology lookup with a learned representation. Each protein sequence is encoded by ESM-2, used frozen, and the resulting embedding is classified as essential or not by a small recurrent head. Because the encoder places a sequence in a space shaped by evolutionary and structural regularities rather than by explicit alignment to a reference set, a protein without an ortholog still gets a position and therefore a prediction — DeepPEP recovered 8 of those 21 Pf-5 species-specific essential proteins, at a cost of 8 false positives.

The name is shared with two unrelated tools: DeepPep, a 2017 peptide-to-protein inference method, and DeepEP, a network-based essential-gene predictor. Neither shares authors, inputs, or task with this model. DeepPEP was posted as a preprint in June 2024 and published in the Journal of Microbiological Methods in 2026.

#Key Features

• Frozen protein language model as the representation layer: the 15-billion-parameter esm2_t48_15B_UR50D checkpoint is run in inference only, producing one 5,120-dimensional mean-pooled vector per protein. No language-model weights are updated.
• Alignment-free essentiality calls: predictions need only a FASTA file, so species-specific proteins with no homolog in any annotated genome are still scored.
• Resampling ensemble for class imbalance: essential proteins are outnumbered roughly eight to one, so the training set is split into balanced subsets, a classifier is trained on each, and the final score is the mean of their outputs.
• Cross-organism evaluation by design: performance is measured by training on one set of genomes and testing on genomes excluded from training, not on a held-out fold of the same organism.
• Released weights and inference pipeline: trained classifier weights are distributed alongside the code, so a new genome can be scored without retraining.

#Technical Details

The training corpus is 66 prokaryotic datasets curated from DEG15 — 25,500 essential and 207,290 non-essential proteins — with non-essential sequences recovered from updated NCBI reference genomes rather than DEG's stale ones. The classifier head is a two-layer bidirectional LSTM with 256 hidden units per direction, followed by adaptive max pooling and a three-layer MLP with a sigmoid output; the ensemble spans five validation splits, each contributing its own set of balanced-subset base models. Across the 66 within-organism test sets, 54 reached AUROC above 0.7. On Staphylococcus aureus (DEG1017) DeepPEP improved on DeepCellEss, the strongest sequence-based baseline, by 0.104 AUROC and 0.183 recall. A 62×62 pairwise transfer matrix showed AUROC falling with 16S evolutionary distance (R = −0.257, p = 5.86e−59), but pooling training genomes beat picking the single closest relative. In the application scenario — train on the 37 reference genomes Geptop 2.0 uses, test on 18 genomes added later — DeepPEP reached 0.850 AUROC with recall 0.754 against Geptop 2.0's 0.594, at lower precision (0.378 versus 0.663). Ablations attribute most of the performance to the encoder: one-hot encoding in place of ESM-2 costs 0.131 AUROC, removing the bi-LSTM costs 0.022.

#Applications

The practical use is annotating a freshly sequenced prokaryote: a biocontrol strain, an environmental isolate, or a pathogen being triaged for antibacterial targets. High recall suits minimal-genome engineering, where missing an essential gene is worse than carrying a few false positives, while orthology tools remain preferable when precision matters more. The authors show the two are complementary and fuse them by soft voting, which improves AUROC over either alone.

#Impact

DeepPEP's contribution is less a new architecture than a careful demonstration that a general-purpose protein language model transfers essentiality annotations between distantly related bacteria, together with an evaluation protocol — train on the reference genomes, test on genomes added to DEG afterwards — that other groups can reuse. The costs are real: the representation step needs the 30 GB ESM-2 checkpoint and upwards of 60 GB of memory (the paper puts it at 60–80 GB, the repository recommends reserving 60–100 GB), precision trails orthology methods, and accuracy tracks the quality of the underlying DEG screens — three noisy datasets were dropped from training. Code, trained weights, and the curated datasets are publicly downloadable, the datasets through a password-gated Baidu Pan share whose password is published in the repository README.