BCCoE

Microbial genomes encode their secondary metabolites in biosynthetic gene clusters (BGCs) — contiguous runs of genes whose enzymes assemble antibiotics, antifungals, and other natural products. Genome mining tools such as antiSMASH can find these clusters and assign them to broad chemical families, but they operate in one modality at a time. Given a cluster you can guess a product class; given a compound you have no principled way to search a strain collection for the genome that makes it. That second, inverse direction is exactly the question a natural-product chemist asks after isolating an interesting molecule.

BCCoE (BGC-Chemical Co-Embedding), developed at A*STAR in Singapore, closes that loop by learning a single vector space that both modalities live in. Frozen embeddings from two pretrained models — BiGCARP for BGC Pfam-domain sequences and MoLFormer-XL for compound SMILES — are projected through two trainable encoders into a shared 64-dimensional space, where a metric-learning objective pulls known BGC-compound pairs together and pushes unrelated pairs apart. Retrieval in either direction then reduces to a cosine-similarity nearest-neighbour query.

The approach is complementary to the existing catalog of BGC models. BGC-Finder annotates gene function within clusters and BGC-Transformer treats domains as a generative vocabulary; neither relates a cluster to the chemical structure of its product. BCCoE is a preprint that has not yet been peer reviewed, and no code or trained weights have been released.

#Key Features

• Bidirectional retrieval: A single co-embedding supports both forward queries (rank likely products for a newly sequenced cluster) and inverse queries (rank clusters likely to biosynthesize a target compound).
• Frozen foundation encoders: BiGCARP and MoLFormer-XL weights are held fixed and only the two projection encoders are trained, which preserves the pretrained representations and limits overfitting on the small paired dataset.
• N-pair contrastive objective: Each positive pair is contrasted against multiple in-batch negatives rather than a single negative, giving stronger supervision than triplet or binary cross-entropy losses.
• Temporal holdout evaluation: Models trained on MIBiG 3.1 were used to predict the 473 BGC-compound pairs that appeared only in the later MIBiG 4.0 release, testing forward prediction rather than interpolation.
• Strain prioritization: Retrieved MIBiG clusters are mapped back to sequenced strains by tblastn, turning an embedding query into a ranked shortlist for wet-lab screening.

#Technical Details

BiGCARP produces 256-dimensional BGC embeddings from HMMER-derived Pfam domain sequences; MoLFormer-XL produces 768-dimensional embeddings from canonical SMILES. Each modality-specific encoder linearly projects its input to 64 dimensions, passes it through a two-layer transformer encoder with feed-forward width 512 and a 128-token input limit, mean-pools over the sequence, concatenates the pooled vector with the mean of the initial embedding sequence, and finishes with batch normalization and a two-layer MLP. The two encoders share architecture but not weights. Training uses active BGC-product pairs from MIBiG 3.1 and 4.0. Across ten-fold cross-validation the model recovered roughly 58.9% of ground-truth clusters within the top 5 when queried by compound, against 0.2% for random ranking. Holding out entire BGC product classes — the hardest setting — still gave lifts of 17.0 and 20.2 at top-10 over random, 74.5% and 89.2% better than an unaligned two-hop nearest-neighbour baseline. On the MIBiG 3.1→4.0 temporal split, BCCoE recovered 188 of 473 new pairs (39.7%) at top-10 when retrieving from clusters already present in 3.1.

#Applications

The practical target is genome mining triage in natural-product discovery. A team holding a strain library and a compound of interest can embed the SMILES, retrieve candidate clusters, and rank their own genomes by homology to those clusters — reducing a screening campaign to a handful of strains. The authors demonstrated this retrospectively on BE-54476-A/B, two anti-Acinetobacter baumannii tetramic acids originally isolated from Streptomyces sp. A58051 during a 54-strain screen. Starting only from the compounds' structures, the true producer was ranked 4th of 54 strains, ahead of the 6th-place ranking from the unaligned baseline. The forward direction supports the converse task: proposing plausible chemical products for clusters in newly sequenced genomes.

#Impact

BCCoE is a compact demonstration that two independently pretrained biological and chemical foundation models can be stitched into a useful joint space with a small amount of paired supervision, without retraining either. Its main constraint is data: MIBiG's curated pairs number in the low thousands, which bounds how much chemical and biosynthetic diversity the alignment can cover, and performance drops sharply on unseen product classes even though it remains well above chance. The absence of released code or weights limits independent reproduction, and the wet-lab evidence is a single retrospective case rather than a prospective discovery. The authors point toward using the aligned space to condition generative models — designing molecules from genetic input, or proposing synthetic clusters for a target compound — which is the more ambitious version of the same idea.