Multiflow

The usual way to design a protein is a relay: a generative model emits a backbone, then an inverse-folding network such as ProteinMPNN writes a sequence onto it. The sequence never gets a vote on the fold. That split exists for a technical reason rather than a biological one — coordinates are continuous and amino acid identities are discrete, and flow matching, the framework that had displaced diffusion for continuous structure generation, had no discrete counterpart to pair with.

Multiflow supplies the missing half. Its authors show that the discrete analogue of continuous flow matching is a continuous-time Markov chain: define a probability flow that interpolates from a masked noise state to data, train a denoising network with plain cross-entropy to predict the clean token, and then pick the rate matrix that simulates that flow after training. They call the resulting class Discrete Flow Models (DFM), the name under which the method is often cited. Because the rate matrix is a post-hoc choice, a single scalar — the CTMC stochasticity η — becomes an inference-time knob rather than a training decision. Multiflow is the protein model built from it, coupling a DFM over the 20 amino acids plus a mask state to an SE(3) flow over residue frames in the style of FrameFlow. Structure and sequence carry independent noise levels, so fixing one of them at the clean end turns the same trained network into forward folding or inverse folding without retraining.

The model comes from Andrew Campbell, Jason Yim, Regina Barzilay, Tom Rainforth and Tommi Jaakkola at MIT CSAIL and the University of Oxford, published at ICML 2024.

#Key Features

• Discrete flows on state spaces: a CTMC realizes the discrete equivalent of a probability flow, subsuming discrete diffusion as one point in a family rather than replacing it.
• Four sampling modes from one checkpoint: co-design, structure-only generation, sequence-only generation, and forward or inverse folding, selected by which modality's noise level is held at 1.
• Stochasticity decided at sampling time: η is not baked into training, and varying it shifts the secondary-structure composition of the samples — helix content moves across roughly 0.67 to 0.70 as η is swept.
• Self-distillation for designability: replacing each training structure's native sequence with the best ProteinMPNN redesign and adding 4,179 self-generated structures to the 18,684 real ones lifts co-design designability from 0.42 to 0.86.

#Technical Details

The network is a modification of the FrameDiff trunk — invariant point attention interleaved with transformer blocks — widened to 8 blocks of 4 transformer layers, with IPA hidden channels cut to 16, skip connections and psi-angle prediction removed, and a three-layer MLP head added to emit amino acid logits, giving 21.8M parameters. Training used 18,684 PDB monomers of 60–384 residues for 200 epochs over three days on four A6000 GPUs with AdamW at a learning rate of 1e-4. Sampling runs 500 timesteps at η = 20. On the RFdiffusion benchmark — 100 samples each at lengths 70, 100, 200 and 300 — co-designed structure-sequence pairs are 0.86 designable under ESMFold refolding, against 0.63 for Protpardelle and 0.37 for ProteinGenerator, with 143 unique structural clusters. Judged on structure alone with eight ProteinMPNN sequences per backbone, designability reaches 0.99 versus 0.87 for RFdiffusion. On 449 held-out monomers the same weights reach 2.2 Å inverse-folding scRMSD against ProteinMPNN's 1.9 Å, but 15.3 Å forward-folding RMSD against ESMFold's 2.7 Å — competitive at writing sequences, far behind at predicting structures.

#Applications

Multiflow suits de novo design campaigns that want a backbone and a sequence that were generated to agree, rather than a backbone plus a sequence bolted on afterwards, and its diversity numbers matter when each candidate costs a wet-lab test. The η knob gives a designer a coarse handle on fold class without retraining, useful when a target calls for helical bundles or β-rich topologies. Beyond proteins, DFM is a general recipe: any problem mixing discrete tokens with continuous coordinates can use the same construction.

#Impact

Multiflow made joint sequence-structure generation a standard baseline rather than an aspiration, and later co-design work — including Pallatom, TriFlow, PhysFlow and DPLM-2 — measures itself against it. Its own limits are stated plainly by the authors: it models backbone frames only, with side chains left to future work; forward folding is well short of a dedicated structure predictor, so it is not a general-purpose protein model; and the reported results are computational designability metrics with no wet-lab validation. Code is MIT-licensed and the trained weights and processed datasets are archived on Zenodo under CC BY 4.0.