CMADiff

Ask a generative protein model for "a secreted, disulfide-rich toxin that blocks a voltage-gated sodium channel" and most of them have nowhere to put the sentence. Conditioning arrives as a taxonomy label, a backbone geometry, or a curated alignment — never as the functional description the designer holds in mind, though UniProt already stores exactly that description in free text. CMADiff takes the sentence as its control signal, but routes it through an intermediate the generator can act on: the physicochemical profile of the sequence.

That indirection is the design decision worth understanding. Rather than conditioning a decoder on a text embedding directly, CMADiff learns a shared space between natural-language annotations and per-residue physicochemical descriptors — hydrophobicity, polarity, isoelectric pH, solvent-accessible surface area, flexibility — and lets the aligned property vector drive the diffusion process. Charge and hydropathy are what decide whether a sequence folds, stays soluble, and forms the disulfide pattern its description implies, so steering through them gives the prompt a mechanistic handle rather than a correlational one.

CMADiff was developed at Northeast Agricultural University in Harbin with collaborators at Zhejiang University, Northeastern University in Qinhuangdao, and Utah State University, and posted to arXiv in March 2025. It belongs with BioM3 and ProteinDT among the text-conditioned protein sequence designers; against unconditional sequence diffusion such as EvoDiff it trades breadth for explicit control. The released code and checkpoints carry the name PhysChemDiff, which is what to search for.

#Key Features

• Text as a design specification: A free-text functional description — organism lineage, keywords, subcellular location — is the conditioning input, so an intent expressible in a sentence needs no structural template.
• Physicochemical conditioning: Fifteen per-residue properties — steric parameter, helix and sheet propensity, hydrophobicity, polarizability, isoelectric pH, side-chain net charge and flexibility among them — enter both as local features and as a sequence-averaged global vector.
• BioAligner contrastive module: Matching text–property pairs are pulled together and mismatched ones pushed apart, with hard negatives drawn from high-similarity non-matches, yielding the text-guided vector the diffusion process consumes.
• Latent-space diffusion: Denoising runs over the compact CVAE latent rather than over amino-acid tokens, which is what makes sampling cheap — 1,000 sequences in 13.59 minutes on one GPU.
• Property-only sampling: Randomly assigned physicochemical conditions with no text still yield confidently folded designs, some with no Foldseek match in natural databases.

#Technical Details

The conditional variational autoencoder passes one-hot sequence and local physicochemical features through separate residual convolutional stacks, fuses them by cross-attention, and feeds a transformer encoder; the latent dimension is 512, the hidden dimension 256, the KL weight 0.5, and the decoder is conditioned on the global property vector. BioAligner embeds text with a pretrained all-MiniLM-L6-v2 sentence transformer and properties with a transformer encoder, both at dimension 384. Generation is a DDPM over the latent whose 1D U-Net noise predictor injects the aligned conditioning by cross-attention, with latent features as queries. Training used ProtSemantic, a corpus pairing Swiss-Prot sequences of up to 128 residues with their organism-classification, comment and keyword annotations; Adam at 1e-4, batch size 64, 500 epochs for the CVAE and 100 for the DDPM, on an A100 80GB cluster.

Across 1,000 generated sequences of length 10–128, AlphaFold 3 scored CMADiff designs at a mean pLDDT of 70.57 ± 14.13, above TaxDiff at 68.66, ProtGPT2 at 51.78, EvoDiff at 51.57 and CARP at 45.94, with natural sequences at 79.45. Foldseek gave a TM-score of 52.04% against Swiss-Prot and 53.62% against the AlphaFold database, at an RMSD of 4.28 Å. Textual semantic fidelity reached 0.94, and the novelty ratio against the PDB was 0.78 for designs above pLDDT 70. Ablations put pLDDT at 37.69 for the CVAE alone and 49.07 with BioAligner removed.

#Applications

The intended user is a protein engineer who can state a functional target in words — a thermostable hydrolase, a secreted knottin, an enzyme active at a given pH — and wants candidates already carrying the biochemical signature such a protein needs. Because conditioning is physicochemical, the same interface supports property-first exploration: fix a target charge and hydropathy profile, sample, filter on predicted structure. All reported evaluation is computational and the authors name wet-lab synthesis as future work, so designs still need the usual expression and assay pipeline.

#Impact

CMADiff's contribution is less its leaderboard position than its intermediate representation: physicochemical properties are a vocabulary both a text encoder and a sequence generator can share, which makes text-driven design interpretable where control is applied. The constraints are real — sequences cap at 128 residues, control quality inherits the quality of Swiss-Prot annotation, and the model reasons about primary sequence, not complexes or conformational change. The preprint has been revised through six versions with no journal publication, and while inference code and checkpoints are released under the MIT License, the training code is withheld until acceptance.