Scinetics-ai/chain-of-structures

Chain of Structures

Weights for Chain of Structures: Connecting Reasoning and Generation for Protein Design. A language reasoning expert writes a design record over four rounds, catalytic core, active site, global backbone, and complete sequence and structure, and each written round re-conditions an all-atom generator whose output becomes the next round's context.

Code: github.com/Scinetics-ai/chain-of-structures

Files

Path	What it is
chain_of_structures.pt	The model. Text is injected at each of the last eight generator layers rather than at a single entry point.
base/qwen3.5-9b/	Reasoning expert base weights, unmodified.
base/dplm2-bit-650m/	All-atom generator base weights, unmodified.
base/struct_tokenizer/	Structure tokenizer for the 13-bit structure tokens.

chain_of_structures.pt is a delta checkpoint: it holds the trained parameters only, and is loaded on top of the three base components above.

Component	Parameters
model_trainable	160.8M, the unfrozen generator layers
qwen_lora	58.2M, rank 32 on the reasoning expert
nce_heads	7.4M
ground	3.3M
text_pos, stage_emb	0.7M

Recorded with the weights: step 9000, best validation 0.780707.

Third-party weights and their licences

Both base models are Apache-2.0 and are redistributed here unmodified, with their own licence files retained alongside them:

• base/qwen3.5-9b/ is Qwen3.5-9B. Its LICENSE is included in that directory.
• base/dplm2-bit-650m/ is DPLM-2-bit 650M, Apache-2.0 as stated in its own README, which is included in that directory.

We did not modify either set of base weights. Our contribution is the delta checkpoint above, applied on top at load time.