Proteina

Generative models of protein backbones had, by 2025, converged on a recipe: an equivariant diffusion or flow network, a few tens of millions of parameters, and a training set of at most half a million structures. Image and language generation had meanwhile moved in the opposite direction — larger non-equivariant transformers, far more data, and semantic conditioning that lets a user say what they want rather than sample and filter. Proteina, from NVIDIA, asks what happens when that second recipe is applied to protein structure.

The answer has three parts. The network is a non-equivariant transformer scaled to as much as five times the parameter count of prior backbone generators. The data is scaled to roughly 21 million synthetic structures mined from the AlphaFold Database, against the ~0.6M used by the largest previous effort. And conditioning is hierarchical: rather than the local motif constraints that earlier models accepted, Proteina conditions on CATH fold-class labels, so a user can request a mainly-β protein, a specific architecture, or a particular topology, in the way an image model conditions on a class label.

Proteina is the root of a family. NVIDIA's later La-Proteina carries the approach into joint all-atom sequence-and-structure generation, and Proteina-Complexa extends it to target-conditioned binder design. The paper was an oral presentation at ICLR 2025.

#Key Features

• Hierarchical fold-class conditioning: Labels drawn from the CATH class, architecture, and topology levels are supplied via TED annotations of the AlphaFold Database, giving coarse secondary-structure control and fine fold-specific control from the same model.
• Long-chain generation: Dropping triangular multiplicative pair updates makes training on large proteins tractable, and the resulting model produces designable, diverse backbones up to 800 residues.
• Guidance at sampling time: Classifier-free guidance sharpens conditional samples, and autoguidance — steering with an undertrained copy of the same model — improves unconditional samples where no label exists to guide with.
• LoRA fine-tuning for structures: Low-rank adaptation transfers a model pretrained on synthetic AlphaFold structures onto experimental PDB data, roughly doubling β-sheet content in its samples.
• Distributional evaluation metrics: FPSD, fS, and fJSD score how closely a generator's output distribution matches a reference set, measuring mode collapse and coverage that designability and diversity together miss.

#Technical Details

Proteina models Cα coordinates with flow matching, using a non-equivariant transformer with conditioned multi-head attention, pair bias, QK normalization, register tokens, and optionally triangular multiplicative pair updates. Released checkpoints span ~60M, ~200M, and ~400M transformer parameters, several with an additional ~15M parameters of triangle layers. Two training corpora are used: D_FS, the 588,318 Foldseek/MMseqs2 AlphaFold Database cluster representatives of 32–256 residues also used by Genie2, and D_21M, a ~20.9M-structure AlphaFold Database subset filtered on pLDDT ≥ 85, coil content, and radius of gyration.

Designability is measured by generating eight sequences per backbone with ProteinMPNN and refolding them with ESMFold, counting a backbone designable at scRMSD ≤ 2 Å. Under that protocol the D_FS model reaches 98.2% designability with 0.49 diversity and the D_21M model 99.0%, against 95.2% for Genie2, 94.4% for RFdiffusion, 94.2% for Proteus, 88.6% for FrameFlow, 74.8% for Chroma, and 22.0% for ESM3. Fold-class-conditional sampling with guidance reaches 89.2% designability while raising β-sheet content from 5.8% to 14.6%. On the standard motif-scaffolding benchmark Proteina records 2,094 unique successes, against 1,445 for the next-best method.

#Applications

The practical draw is control. A designer who needs a TIM barrel, an immunoglobulin-like fold, or simply a β-rich scaffold can condition on that fold class instead of sampling unconditionally and screening for it, which changes the economics of scaffold search. The 800-residue regime opens up multi-domain scaffolds and larger binding surfaces than earlier generators could reach, and motif scaffolding supports the common workflow of building a protein around a known functional site. The LoRA recipe lets groups specialize a released checkpoint on a structural subset of their own without retraining from scratch.

#Impact

Proteina established that data and parameter scaling transfer to protein structure generation, and that non-equivariant architectures — already standard in AlphaFold 3's diffusion module — are sufficient for unconditional generation, where equivariant methods had been assumed necessary. Its distributional metrics have given the field a way to detect generators that score well on designability while collapsing onto a narrow set of folds. Two caveats bound the result. All evaluation is in silico: designability is self-consistency against ProteinMPNN and ESMFold, not experimental expression, and the paper reports no wet-lab validation. And the release — code, weights, dataset indices, and auxiliary files alike — is under an NVIDIA license restricted to non-commercial research use, so commercial protein design programs cannot build on it directly.