ProDiT

ProDiT (Protein Diffusion Transformer) is a generative protein design model that treats amino-acid sequence and three-dimensional structure as two halves of one diffusion process rather than as separate modeling problems. It was released as a bioRxiv preprint in September 2025 by Bowen Jing, Bonnie Berger, and colleagues at MIT, with collaborators at Harvard Medical School, UT Austin, and the University of Washington.

The model targets a split that has shaped de novo protein design. Structure diffusion models such as RFdiffusion generate high-quality backbones, but function must be hand-specified as a binding partner or motif, and training is confined to the comparatively small Protein Data Bank. Sequence generative models such as EvoDiff scale to far more data but cannot natively reason about the structures their sequences fold into. Earlier multimodal attempts either tokenized structure into a discrete vocabulary, as in ESM3, or forced sequences into a continuous space; both routes cost generation quality.

ProDiT instead runs continuous diffusion over Cα coordinates and discrete masked diffusion over amino-acid tokens inside one transformer, with molecular function supplied as an explicit conditioning input. Because either modality can be held fixed while the other is denoised, unconditional generation, co-generation, inverse folding, and motif scaffolding all become sampling protocols over a single checkpoint. That flexibility enables the paper's most distinctive contribution: coupled structure diffusion, which ties two structure trajectories to one shared sequence trajectory to design proteins that adopt two conformational states.

#Key Features

• Joint sequence-structure diffusion: Continuous diffusion on backbone coordinates and discrete masked diffusion on amino acids train under one loss, so both modalities are generated consistently rather than chained through a separate inverse-folding step.
• Function conditioning at scale: A learned embedding for each of 8,220 molecular-function Gene Ontology terms steers designs by biological function directly, rather than by a hand-built geometric specification.
• Coupled structure diffusion: A multistate design protocol that needs no training on structural ensembles, used to scaffold enzymatic active sites intended to be allosterically switched off by a calcium effector.
• Beyond the PDB: Training on predicted structures for 214M UniProtKB proteins pushes structural coverage far past the experimental structures that constrain PDB-trained generative models.
• No geometric machinery: Coordinates enter through a plain linear layer with no equivariant modules or pairwise tracks, which is what makes this data scale and crop length practical.

#Technical Details

ProDiT is a 30-block diffusion transformer adapted from image-generation practice, using adaLN-Zero conditioning on the structural diffusion time, pre-norm, QK-norm, GeLU activations, and an unusually high 20% dropout rate the authors found necessary for sequence quality. Two checkpoints were trained: 321M parameters at width 768 without function conditioning, and 576M at width 1024 with it, both on a 512-residue crop with batches of 250k-500k tokens across 16-32 NVIDIA H200 GPUs for roughly 500k steps. Training drew on the 214M UniProtKB entries with AlphaFold Protein Structure Database models, filtered to pLDDT above 80 and sampled hierarchically from structural rather than sequence clusters — 1.24M FoldSeek clusters over 18.9M MMseqs clusters and 128M sequences.

On unconditional generation of 100 samples each at lengths 100 through 500 — success being pLDDT above 70 for sequences, and self-consistency RMSD under 2 Å for structures redesigned with ProteinMPNN and refolded with ESMFold — ProDiT matches RFdiffusion on structure generation and holds quality better beyond 300 residues, while generating substantially more successful sequences than ESM3 or EvoDiff. Co-generation stays self-consistent out to 500 residues. Of 915 Gene Ontology molecular-function terms screened with the DeepFRI predictor, 465 yielded at least one successful design, with success rates tracking term frequency in the training data.

#Applications

ProDiT suits protein engineers who want to specify a design by biological function rather than geometry: generating candidate enzymes for a target activity, scaffolding a catalytic motif into a novel fold, or producing starting points for directed evolution. Designs conditioned on aldehyde dehydrogenase and malonyltransferase activity recovered catalytic residues and cofactor-binding pockets to sub-ångström accuracy at only 41-45% sequence identity to reference enzymes. The multistate protocol extends this to regulated proteins — switchable enzymes, sensors, signaling components — where one sequence must support two conformations.

#Impact

ProDiT shows that a plain transformer trained on predicted structures at evolutionary scale can match specialized equivariant structure-diffusion architectures while adding sequence and function generation, and it makes multistate design reachable by sampling alone. The work remains a preprint awaiting peer review, and its results are entirely computational: the allosteric lysozyme and carbonic anhydrase designs were assessed by co-folding with Chai-1 with and without calcium, and the authors state that scaffolding the selected active-site residues does not guarantee catalytic activity and that experimental verification is still required. No code or weights have been released — the project's GitHub repository has held only a "coming soon" note since September 2025 — so the model is not yet reproducible outside the authors' group.