BaCNet

Most compound-protein interaction models were built for human targets. Bacterial proteomes differ enough in post-translational chemistry and sequence composition that repurposing those models for anti-infective discovery is unreliable, and the shortfall matters: two-thirds of the antibiotics in clinical use were developed between the 1940s and the 1960s, while resistance in the ESKAPEE pathogens keeps narrowing the options that remain.

BaCNet — Bacterial Compound-protein interaction Network — is a binding-affinity predictor trained specifically on bacterial targets. Developed by Yuki Otani, Daisuke Koga, Yasunari Wakizaka, and Hideyuki Shimizu at the Institute of Science Tokyo and posted as a preprint in March 2025, it takes a protein's amino acid sequence and a compound's SMILES string and returns a single scalar, the "BaCNet score," with no structure, docking, or assay data required at inference.

Its distinguishing design choice is the evaluation protocol. Two entire species — the Escherichia coli K-12 strain and Pseudomonas aeruginosa — were removed from the training corpus and reserved as an external test set, simulating the arrival of a pathogen the model has never met. Because no other Pseudomonas species appear in training, P. aeruginosa is novel at the genus level. A single trained checkpoint is applied to both without retraining, and the same fixed scorer drives all downstream screening.

#Key Features

• Sequence-and-SMILES-only input: Predictions need nothing beyond a protein sequence and a compound string, so a target can be screened as soon as its gene is annotated, before any structure is solved.
• Held-out-species evaluation: Two clinically important species were excluded from training entirely, and the frozen model retained its level of predictive performance on both.
• Three-way compound representation: Morgan fingerprints, Chemical Checker signatures, and a chemical language model encode local substructure, bioactivity similarity, and global molecular pattern respectively. Combining all three beat every single embedding and every pair.
• Rediscovery of known antibiotics: Scoring documented drug-target pairs from species absent from training — Streptococcus pneumoniae, Haemophilus influenzae, and E. coli K-12 — recovered high scores for antimicrobials with diverse mechanisms of action.
• Generative extension: An "Optimization Flow" loop generates analogs of hit compounds with a variational autoencoder, filters them for drug-likeness, and re-scores them with the same fixed BaCNet model.

#Technical Details

Training used roughly 16 million compound-protein interaction records drawn from STITCH v5.0 for 23 ESKAPEE-associated bacterial species, with STITCH combined scores Box-Cox transformed and normalized to [0, 1]. After removing the two external-test species, the remainder was split 80/10/10 with no protein sequence shared across partitions. Proteins are embedded with a frozen ESM-2 15-billion-parameter model (esm2_t48_15B_UR50D), column-averaged to a 5,120-dimensional vector. Compounds are embedded as a 1,024-bit Morgan fingerprint, a 1,280-dimensional Chemical Checker signature, and a 384-dimensional ChemBERTa-2 embedding, concatenated to 2,688 dimensions. The resulting 7,808-dimensional vector feeds a three-layer perceptron of widths 1,024, 256, and 32 with ReLU and batch normalization, trained with MSE loss and AdamW on eight A100 GPUs. On the held-out test set the model reached a Pearson correlation of 0.9382 (95% CI 0.9378-0.9387), a Spearman correlation of 0.9325, MSE 0.00975, and MAE 0.0577.

Applied to penicillin-binding proteins, BaCNet ranked 2,616 DrugBank compounds against E. coli K-12 PBP and 186,036 filtered natural products from COCONUT against P. aeruginosa PBP. Molecular dynamics on the top 300 COCONUT hits found one compound binding the transpeptidase pocket at -9.14 kcal/mol without a beta-lactam ring. Three rounds of Optimization Flow produced a generated compound scoring above every molecule in the source library.

#Applications

BaCNet targets the screening stage of antibacterial discovery, where the practical bottleneck is deciding which of hundreds of thousands of compounds justify synthesis and assay. Because it needs only a sequence, it fits outbreak scenarios in which a newly sequenced pathogen has no solved structures and no assay history: a candidate target can be scored against a repurposing library the same day. The non-beta-lactam PBP binders it surfaced are of particular interest to groups working on beta-lactamase-resistant infections, and the Optimization Flow loop extends the same scorer into lead optimization.

#Impact

BaCNet is a preprint and has not been peer reviewed; no code or model weights have been released, and the preprint carries no data or code availability statement, so independent reproduction is not currently possible. Its evaluation is entirely computational — the compound predictions were assessed by molecular dynamics rather than by microbiological assay — and the authors are explicit that the score reflects predicted binding, not measured inhibition, since the model encodes no three-dimensional structure and no pharmacokinetic or toxicity information. What it does establish is that a protein language model trained on all of life transfers usefully to bacterial targets, and that a fixed compound-protein scorer can hold accuracy on a genus it has never seen — the property that any screening tool for emerging pathogens has to demonstrate.