enzymecot

EnzymeCoT - Stage III joint checkpoint

Joint adapters for reaction-conditioned, three-round de novo enzyme design: Qwen3-4B as the understanding model and RFdiffusion3 as the generator, trained together.

What this file is

s3_best.pt - a 553 MB PyTorch state_dict (755 tensors, 138.9 M parameters). Adapters only: no optimizer state, no base model weights.

Important caveat about the reported numbers

Evaluation figures reported for this line of work - catalytic-site RMSD 0.94 A, LigandMPNN-8 self-consistency 5/8 and 8/8 designable, pLDDT 77.9 / 94.5 - were produced with the step-2500 checkpoint (val 0.3526).

s3_best.pt is step-3500 (mean val L_gen 0.3611 - the lowest over the run, which is why it was saved as best). An earlier version of this card said step-4000 with val 0.3497; both were wrong - see the correction note below. The step-3500 weights have not been re-evaluated end to end, so please do not attribute the step-2500 numbers to them without re-running the evaluation.

s3_step2500.pt (the checkpoint those numbers DO come from) and s3_step4000.pt are now included in this repository so the reported figures can be reproduced against the weights they were actually produced with.

Files

This repository is self-contained: everything needed to run the model is here.

How the Qwen merge was verified

The Stage III Qwen LoRA is r=64, alpha=128, targets q/k/v/o/gate/up/down_proj (RFdiffusion3's is r=16, alpha=32, targets linear_1/linear_2/linear_3). The Qwen LoRA was merged in float32 and then cast to bfloat16, and compared against the unmerged model on the same prompts:

The merge itself is exact (row 1); the error is dominated by bfloat16, which the adapter path already incurs. The published merged model matches the adapter path on argmax and is slightly better on top-5.

RFdiffusion3 is shipped unmerged on purpose: its checkpoint carries EMA copies (1616 tensors / 336.1 M parameters) and merging into it is easy to get silently wrong, so the upstream file is kept byte-identical and the adapter is handed to you instead.

Provenance of s3_best.pt (corrected 2026-09-17)

Earlier versions of this card made two claims about s3_best.pt that do not hold. What is actually verifiable:

• s3_best.pt is value-identical to s3_step2500.pt of the training run s3prod_run1. Verified by loading all 755 tensors of the published file and of every checkpoint of every s3prod_* run on our cluster and comparing them element-by-element: zero tensors differ from that one file, while 754 of 755 differ from every checkpoint of the later s3prod_pp4 run. So the published weights are a step-2500 checkpoint - which means the evaluation figures reported for this line of work (catalytic-site RMSD 0.94 A, LigandMPNN-8 self-consistency 5/8 and 8/8) were produced with this checkpoint, not a different one. The earlier caveat to the contrary was based on a misreading.
• The card previously said "step-4000". That was wrong. The likely cause: in s3prod_run1 the file s3_best.pt carries the modification time of the step-4000 save while containing the step-2500 weights, so a timestamp-based reading gives the wrong step. Step numbers here were established by comparing tensors, not timestamps or file names.
• The card previously quoted validation 0.386 -> 0.3497, and a later revision of this card quoted 0.3611. Neither describes this file. The training log of s3prod_run1 is not retained on our cluster, so no validation number for s3_best.pt can be verified, and none is claimed here. The strings 0.3497 and 0.386 do not appear in any surviving log.
• For the s3prod_pp4 run, whose log is retained, the verified curve (mean L_gen over the three rounds) is 0.3804 (step 250) -> 0.3611 (step 3500, the minimum) -> 0.3623 (step 4000). Those numbers describe the files under s3prod_pp4/ only, and s3prod_pp4 has not been evaluated end to end.

No weights were changed by any of these corrections; only the description was wrong.

Requirements

This is an adapter checkpoint, not a standalone model. You need:

• Qwen3-4B base weights
• RFdiffusion3

Task definition

Given a reaction (SMILES), the ligand's chemistry and 3D coordinates, and a target protein length, the model designs an enzyme in three rounds:

1. R1 - an unindexed catalytic-residue set (identity + catalytic/support role) together with local geometry relative to the ligand. Object IDs carry no sequence position.
2. R2 - a full backbone, then an assignment of the round-1 objects onto positions in that generated backbone.
3. R3 - the remaining sequence and side chains, inheriting the committed core identities.

Data

Trained on the EnzymeCoT corpus (19,888 train / 1,123 validation records, lengths 100-997) with a reaction- and homology-disjoint split. Ligand binding conformations are taken from reference complexes, which are Boltz predictions rather than experimental structures.

Limitations

• No experimental validation. No catalytic activity has been measured.
• The metrics above are computational self-consistency and RMSD-to-reference; neither establishes catalysis.
• Autonomous three-round design quality remains weak; treat this checkpoint as a research artifact, not a working enzyme designer.