Genomic language model finetuned with discrete diffusion to redesign biosynthetic gene clusters, conditioned on antiSMASH catalytic domain order.
No providers recorded yet. Browse all providers
A type I polyketide synthase is an assembly line built into a single polypeptide of several thousand residues: a ketosynthase, an acyltransferase, and a reduction loop of dehydratase, enoylreductase and ketoreductase domains each act in turn on an intermediate tethered to a carrier protein. The standard route to a new molecule is to swap one domain for a counterpart from another cluster, and it usually underperforms, because the borrowed domain must cooperate with neighbors it never evolved beside. The constraints that make those interfaces work are distributed across the enzyme rather than concentrated in a handful of catalytic residues, so there is no short list of positions to mutate.
gLM2 BGC Decoder generates an entire multidomain region at once, with the rest of the assembly line held fixed as context. It is a 650M-parameter finetune of gLM2, Tatta Bio's mixed-modality genomic language model, converted from a fixed-rate masked encoder into a masked discrete diffusion decoder that iteratively denoises sequence at any masking level. Generation is steerable through domain-level conditioning: the tokenizer carries 1,999 antiSMASH HMM domain annotations as discrete special tokens prefixed to each coding sequence, so a caller can stipulate the catalytic domain order and have the model produce sequence consistent with it.
The model was developed by Tatta Bio with the Joint BioEnergy Institute, Lawrence Berkeley National Laboratory, UC Berkeley, UCSF and MIT, and introduced in a September 2026 bioRxiv preprint (Lanclos, Ibrahim, Cornman, and colleagues). Its demonstration is a chimeric PKS engineered to produce δ-valerolactam, a lactam no polyketide synthase makes naturally.
Finetuning proceeded in two stages from the released gLM2 650M checkpoint. Stage one adapted it to generation on the Open Genome (OG) dataset — 6.2 million microbial contigs comprising roughly 400 million proteins — using contiguous same-strand coding sequences with intergenic regions removed, a 4,096-token context, per-sequence masking rates drawn uniformly and a cross-entropy loss reweighted by (1 − p). Contigs were downweighted by MMseqs2 cluster size at 70% identity to offset overrepresented protein families; one epoch took 90 hours on eight H100 GPUs in bfloat16. Stage two specialized the model on the BiG-FAM biosynthetic gene cluster database, added the domain vocabulary, widened the context to 16,384 tokens, and ran four epochs in 80 hours on the same hardware. Both stages used AdamW at a 4.0e-5 peak learning rate with a 5,000-step warmup. Inference decodes masked tokens with Path Planning (P2) diffusion sampling at temperature 0.7 and eta 0.1; candidates are ranked by model pseudo-likelihood and screened by 1-gram Shannon entropy to discard low-complexity outputs.
The model targets biosynthetic engineering, where the goal is an enzyme that carries a non-natural intermediate through a sequence of chemical steps. In the reported campaign, round one redesigned the ketosynthase-acyltransferase didomain of a FluPKS/CmiPKS chimera with active-site residues fixed from Boltz-2 co-folding; all 45 tested designs matched or exceeded the starting titer and the best gave roughly 5.7-fold more product. Round two redesigned the dehydratase-enoylreductase- ketoreductase reduction loop with no fixed residues at all, a regime where structural priors are usually unavailable. Beyond polyketides, the same conditioning interface applies to other domain-organized clusters, serving metabolic engineers and natural-product chemists rather than requiring a structure for every target.
The work shows that evolutionary sequence statistics learned across natural clusters can coordinate changes over thousands of residues at once, producing variants up to 9.4-fold higher in δ-valerolactam titer than the rationally designed starting enzyme. Analysis of the best design found no convergence toward any single natural PKS: reconstructing 99% of its redesigned windows required combining segments from as many as 78 BiG-FAM or 58 NCBI sequences, indicating interpolation across sequence space rather than retrieval. It joins gLM2-SAE as a derivative of the same backbone released by the group. The weights and the OG dataset are public and the inference code is on GitHub, but both the code and the weights carry the same non-commercial CC BY-NC 4.0 license, Tatta Bio has a patent pending on genomic language modeling for biosynthetic assembly-line generation, the preprint has not been peer reviewed, and the wet-lab evidence covers one target molecule in one host.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.