Func2Prot

A generative model fitted to one protein family learns what that family's sequences look like but nothing about what the family does, and a model conditioned on a bare function keyword treats "DNA binding" and "double-stranded DNA binding" as two unrelated labels rather than a term and its child. Func2Prot replaces the keyword with a learned Gene Ontology embedding: each Molecular Function term's definition and description are encoded with BioBERT, then propagated across the ontology's directed acyclic graph by a two-stage sequential graph convolution that aggregates along the hierarchy's direction rather than isotropically. Functions near each other in the ontology end up near each other in the conditioning space, so what the model learns about one annotated family carries over to families labeled with related terms.

Func2Prot is the joint sequence-structure member of that framework. Where its sequence-only sibling Func2Seq models P(seq | func), Func2Prot models P(seq, str | func): the function-conditioned autoregressive decoder emits amino acids one at a time, and a second head reads the same final token representations to produce backbone geometry, so a design arrives as a sequence and a three-dimensional backbone rather than a sequence to be folded afterwards. Both models come from Yang Shen's group at Texas A&M University, published together in Biophysical Journal in 2026.

#Key Features

• Function is the specification: Generation is conditioned on a Gene Ontology Molecular Function embedding rather than a fold or template, and any family in the Pfam-to-GO mapping can serve as the conditioning target.
• Backbone-angle decoder: For a length-L protein the structure head emits an L × 6 tensor — the dihedrals φ, ψ, ω plus three bond angles per residue, the parameterization used by FoldingDiff — converted to Cartesian coordinates with bond lengths fixed at standard values.
• One meta-pretrained checkpoint for both jobs: The released meta-pretrained checkpoint is loaded unchanged for conditional log-odds variant scoring and for GO-conditioned joint sampling alike; family-specific fine-tuning is layered on top, not a prerequisite.
• Coupled but asymmetric decoding: Residues are sampled autoregressively with top-k sampling, each choice fed back in, while the backbone angles come in a single pass from the final sequence embeddings — coupling the two through shared representations without autoregressive sampling over continuous variables.

#Technical Details

The sequence decoder is a function-conditioned transformer trained in both forward and reverse directions, each carrying its own structure head; the reference training configuration is four layers, 12 attention heads and 768-dimensional hidden states. Meta-pretraining uses the Pfam 34 RP15 corpus, filtered to 8,772,255 sequences shorter than 800 residues drawn from GO-annotated families and free of non-natural amino acids. Angles are supervised against experimentally determined structures under a smooth L1 loss, and the released checkpoint was trained at a learning rate of 5e-5 with the structure term weighted equally against the sequence term. Paired sequence-structure supervision is the scarce ingredient: for chorismate mutase, 798 of 13,479 fine-tuning sequences had a corresponding experimental structure, and coverage is thinner elsewhere.

Evaluation is deliberately narrow. Across two enzyme fitness case studies, joint modeling beats the sequence-only counterpart for chorismate mutase — a compact domain with tight structure-function coupling — while for malate dehydrogenase the benefit is mixed across solubility and activity endpoints. For controllable design the authors fine-tuned on the ARID DNA-binding domain (Pfam PF01388, 18,740 filtered sequences): against the sequence-only model, Func2Prot raises maximum sequence identity to natural ARIDs and lowers sequence coverage while producing backbones of high structural similarity and coverage. A representative design from the twilight zone of 20–30% global identity reached a TM-score of 0.867 against a reference ARID fold.

#Applications

Func2Prot suits design campaigns whose specification is a molecular function rather than a target fold: ask for a sequence that performs a Gene Ontology activity, and get a candidate backbone alongside it, so structural filters can run before anything is ordered. The ARID study is the pattern — preserve residue-level constraints at the DNA-binding L1 and L2H5 regions while pushing global sequence identity into novel territory. The same checkpoint also scores variants by conditional log-odds, most usefully in compact domains where structure and function are tightly coupled and alignment-based predictors are short of homologs.

#Impact

The contribution that generalizes is the conditioning signal: encoding Gene Ontology terms as text plus topology gives a generative protein model a functional specification with real structure in it, rather than a one-hot family index. Func2Prot then tests whether adding structure as an explicit modeling variable buys anything, and reports a candid answer — it helps for chorismate mutase, is mixed for malate dehydrogenase, and its clearest advantage is emitting a backbone at all. The limits are stated plainly: the design evidence rests on single-family case studies, paired sequence-structure training data is sparse, and every result is computational, with no experimental characterization of a designed protein.