2026-09-19
gLM2 plus discrete diffusion redesigns two multi-domain regions of a chimeric type I PKS, raising δ-valerolactam titer up to 9.4-fold over the starting enzyme.
Type I polyketide synthases (PKSs) are microbial assembly lines. Catalytic domains sit in order and stitch small building blocks into complex natural products: antibiotics, anticancer agents, and more. Over the past four decades, more than half of FDA-approved small-molecule drugs have been natural products, derivatives, or synthetic mimics. Engineers have long wanted to swap modules from different natural lines and program molecules that nature never made.
Swapping modules is slow and low-yielding. Function depends on substrate channeling and interdomain contacts that nobody can write down as a design rule. Even chimeras cut at evolutionary recombination boundaries often limp. Keasling's group previously fused pieces of the Fluvirucin and Cremimycin PKSs into VL-Start, a chimera that makes δ-valerolactam, a nylon-5 monomer that PKSs do not produce in nature. VL-Start was already the best of 140 module-swap constructs. This paper asks whether a genomic language model can rewrite large multi-domain stretches in the context of the full enzyme and push that non-natural line further.
The base model is gLM2, a 650-million-parameter mixed-modality genomic language model trained on 3.1 trillion base pairs from OpenMetaGenomic. Coding regions are amino acids; intergenic DNA is nucleotides. gLM2 was a masked encoder, so it could not sample. The authors attach a discrete diffusion head so the same model can in-paint, optimize a template, or generate from scratch.
Fine-tuning comes in two stages. First, 6.2 million microbial contigs from OpenGenome (about 400 million proteins), intergenic sequence stripped, masking rate drawn uniformly from 0% to 100% per coding sequence. That covers both unconditional generation and conditioned redesign around fixed neighbors or domains. Second, BiG-FAM: 1,225,071 biosynthetic gene clusters, including 211,243 PKSs, 337,201 nonribosomal peptide synthetases, and 79,506 hybrids. Context grows to 16,384 amino acids. The vocabulary gains 1,999 antiSMASH domain-family tokens, prepended as unmasked prompts, so domain order becomes a control knob at inference. Sampling uses Path Planning (P2) diffusion. Stage 1 took about 90 hours on 8×H100; stage 2 about 80 hours.
The wet-lab target is VL-Start, chromosomally integrated into Pseudomonas putida to make δ-valerolactam. Redesign is split into two rounds, in part to avoid synthesizing one 7.5 kb, 60% GC fragment.
All 45 Round 1 designs matched or beat VL-Start; 84% (38/45) cleared one standard deviation. The top variant, R2-Start, made about 5.7-fold more valerolactam. Round 2 roughly doubled that, reaching 9.4-fold over VL-Start. The introduction quotes sequential gains of 5.7-fold and 1.67-fold, which multiply to about 9.5.
| Round | Window | Tested | vs VL-Start |
| 1 | KS-AT, 79 residues pinned | 45 | up to 5.7× |
| 2 | DH-ER-KR, nothing pinned | 28 | up to 9.4× |
With no residues pinned, Round 2 still left annotated catalytic positions untouched in every tested sequence. Median substitution rate at predicted active sites was 38-fold lower than the rest of the window. Changes piled up on solvent-exposed surfaces; the core and pockets stayed quiet. Against ClusterCAD natural PKS alignments, about 45% of substitutions hit the consensus residue, another 45% a common residue (≥5%), about 9% rare, about 1% unseen at that site. The amino-acid chemistry is conservative, mostly within-class swaps.
R2-Top differs from VL-Start at 256 positions across the two windows (86% identity inside them). It does not collapse onto one natural PKS. Covering 99% of the designed residues takes the union of 78 BiG-FAM sequences or 58 NCBI nr hits. AlphaFold3 pLDDT and gLM2 pseudo-likelihood do not track titer; Round 1 even shows a weak negative link between pLDDT and production. Proteomics says expression explains some of the spread, not all of it.
The model is being asked to expand function: the product is a nylon monomer outside the natural PKS repertoire. For a synthetic-biology lab, the recipe is concrete. Pin the pocket, rewrite a didomain, then free the reductive loop, and screen tens of variants rather than thousands. Weights and code are public.
The mutations that ride with higher titer are scattered across domains, linkers, and sites far from the original chimera junction. That undercuts the old Lego story of plug-and-play modules. For genome-mined clusters with no ligand-bound structure, unconstrained generation of the kind used in Round 2 is the more realistic mode.
Still incremental. The template already worked, and was the best of 140 hand-built chimeras. The model is optimizing a functional enzyme, not inventing an assembly line from a blank page.
The authors cannot say which of the 256 substitutions did the work, or whether a handful of local changes carry most of the gain. Titer does not line up with substitution location, tree clades, or global sequence and structure scores. Round 2 used shifted genomic integration loci, so the two rounds are not fully matched. Expression is one driver and not a full account.
Only one chimeric PKS, one product, one host. Pseudo-likelihood was used to pick candidates, then failed to rank the ones that actually produced. Wet lab remains the bottleneck. Boltz-2 ligand placement on PKS structures is shaky: 38% success for KS, 18.2% for DH/ER, 0% on multi-domain docking. The 79 pinned residues in Round 1 rest on that pocket call. Tatta Bio has a patent pending on genomic language modeling for biosynthetic assembly-line generation. The manuscript is a bioRxiv preprint, not yet peer reviewed.