BacLLM folds UniProt annotation knowledge into a protein language model, so sequence-only input can call resistance genes and rank phage candidates
Synopsis
BacLLM injects semantic representations of UniProt annotations, encoded by a large language model, into an ESM-2 sequence encoder through cross-modal contrastive learning, so inference reads amino acid sequences alone. Across six resistance-prediction benchmarks, BacLLM (4B) beat Ankh, ESM-2, ProteinBERT and ProTrek in mean F1 under both logistic regression and an RBF-kernel support vector machine; on the SHV-1 β-lactamase P0AD64 it recovered six of eight experimentally supported resistance-attenuating sites among the top 65 ranked positions, while ESM-2 recovered none. The text branch is removed after training, so these gains come from sequence input, not annotation input.
Interpretation
A single encoder that reads only amino acid sequences can now carry resistance-gene identification, resistance class and mechanism calls, single-residue mutation-effect ranking, and strain-level phage–host matching. Earlier routes split in two: similarity-based alignment produces substantial false negatives when query proteins are weakly similar to annotated references, while task-specific deep models such as ARGNet, HMD-ARG and PhageHostLearn depend on specialized architectures or task features and are hard to reuse across surveillance, mechanism study and therapeutic ranking. Under a standardized frozen-encoder protocol, six primary benchmark settings were scored with logistic regression and an RBF-kernel support vector machine, and BacLLM (4B) exceeded Ankh, ESM-2, ProteinBERT and ProTrek in mean F1 under both classifiers; in MMseqs2 five-fold cross-validation at 85%, 90%, 95% and 98% sequence identity, BacLLM (4B) outperformed every baseline at all four thresholds under both classifiers.
Training pulls the sequence encoder toward annotation semantics, and inference keeps only the sequence branch, so annotation knowledge is compressed into the protein representation rather than required as input. General protein language models either learn from amino acid context alone without explicit semantic supervision from functional annotations, or, like ProTrek, align sequence, structure and text without adaptation to the resistance domain. Training used roughly 12,000 UniProt sequence–annotation pairs covering WHO priority pathogens; annotation fields were rewritten into coherent descriptions by MedLLM-Llama3-8B and encoded by a frozen Qwen3, while the sequence side used ESM-2 with two fully connected layers and layer normalization projected into the same space, aligned by a unidirectional InfoNCE loss.
The representation space also retains biological organization beyond classification: single-residue perturbation effects, taxonomic relations among strains, and a TolC-centered efflux functional neighborhood. These structures emerge from unsupervised similarity in pure sequence representations rather than from task-specific protein-interaction or network supervision. On P0AD64, BacLLM ranked Y101L as the substitution with the largest negative probability shift, matching the roughly 500-fold resistance reduction recorded in UniProt; unsupervised hierarchical clustering of ESKAPE strain embeddings separated A. baumannii from P. aeruginosa, consistent with their placement in Moraxellaceae and Pseudomonadaceae; in the TolC subnetwork, three low-identity proteins KPN_00995, KPN_01486 and KPN_03299 aligned to TolC with AlphaFold TM-scores of 0.765, 0.786 and 0.735.
Perspective
The demonstrated scope is retrospective public-dataset evaluation on WHO priority pathogens, covering resistance sequence classification, single-residue mutation analysis, protein functional networks and strain-level phage–host matching; inference uses amino acid sequences only and adds no omics or clinical data.
Training and evaluation both rest on retrospective public datasets, and temporally independent and clinically derived data are needed to establish generalizability; the prioritized TolC-like candidates remain computational predictions requiring direct functional validation; downstream analyses currently rely on protein sequence alone and incorporate no other omics or clinical data.