OriginFlow

OriginFlow is a de novo protein backbone generator built on optimal-transport flow matching, developed at Beihang University with collaborators at dProtein Biotechnologies and Shandong Second Provincial General Hospital and posted to bioRxiv in April 2025. It represents backbones as rigid-body frames — rotations in SO(3) and translations in R³ — and learns to reverse an optimal-transport path from a noise prior to native structure. It also maintains a covariance-matrix view of the same structure in R³, which lets SDE and probability-flow ODE samplers be used without deriving the corresponding dynamics on SO(3).

The claim the work rests on is experimental yield rather than benchmark scores. Backbone generators such as RFdiffusion and Chroma established that generated structures can be highly designable in silico, but laboratory hit rates for de novo binders remain the field's bottleneck. OriginFlow pairs a binder-tuned checkpoint with an explicit iterative protocol: generate a backbone against a fixed target chain, fill the binder sequence with ProteinMPNN, screen with ESMFold, and accept only designs whose AlphaFold 3 complex predictions exceed 0.8 for both pTM and interface pTM, refilling sequences for one or two further rounds when they do not.

Separate checkpoints for monomer, motif, symmetric, and binder generation are distributed through a Google Drive link in the repository; the authors note these were trained at different times against slightly different network variants, so code and weights must be matched. The repository contains no LICENSE file, and the preprint is released under CC-BY-NC.

#Key Features

• Interchangeable ODE and SDE sampling: The dual frame and covariance-matrix representation makes constant-velocity ODE, SDE, and probability-flow ODE sampling equivalent routes to the same model, so samplers can be chosen per task.
• Length extrapolation: Trained on 384-residue crops, the model generates coherent monomers and complexes past 600 residues, with weaker metric decay against length than the diffusion baselines it is compared to.
• Task-specific checkpoints: Fine-tuned weights cover unconditional monomers and complexes, motif scaffolding, secondary-structure-constrained generation, symmetric assemblies, and target-conditioned binders.
• Symmetry by sampling-time guidance: Cyclic, dihedral, and tetrahedral assemblies of up to 720 residues come from symmetrizing the predicted clean structure at each reverse step and following the resulting gradient, with no symmetry-specific training set.

#Technical Details

The network stacks six transformer blocks with invariant point attention for global frame updates and three GNN-transformer layers that refine each residue against its K nearest Cα neighbours; rotary position encoding and length-based attention scaling support extrapolation beyond training length. Training used RCSB PDB structures better than 2.6 Å resolution and under 1,000 residues — roughly 17,000 chains, clustered at 30% sequence identity — cropped to 384 residues, with a steric-violation loss and an ELBO term over the covariance representation. Binder checkpoints were fine-tuned on complexes from PDBbind, and sampling uses 500 steps per design, or 1,000 for binders.

Across 40–600 residues, unconditional generation reached median pLDDT 80.8, self-consistency TM-score 0.90, and scRMSD 1.9 Å, against 86.8 / 0.97 / 1.0 Å for RFdiffusion and 57.35 / 0.65 / 6.9 Å for Chroma. Mean pairwise TM-score among samples was 0.45 versus 0.48 for RFdiffusion, and Foldseek's best match in CATH-S40 averaged 0.61, between Chroma's 0.59 and RFdiffusion's 0.66. On a 17-problem motif-scaffolding benchmark — solved when motif RMSD falls below 1 Å and ESMFold pLDDT exceeds 70 — OriginFlow solved 16, ahead of RFdiffusion, EvoDiff, and PVQD; across 25 length-discretized variants it solved 24.

#Applications

The intended use is early-stage binder discovery against a structurally characterized target, where screening cost dominates. Designs were generated against PD-L1, SARS-CoV-2 spike RBD, MDM2, GLP-1R, GP120, KRAS G12D, and VEGF, and seventeen designs against PD-L1, the RBD, and VEGF were expressed in E. coli. All appeared in the soluble fraction at 0.8–3.2 mg per litre of culture, circular dichroism confirmed the intended secondary structure, and melting temperatures ran from 60 °C to above the instrument's 90 °C ceiling. Surface plasmon resonance measured dissociation constants from 2.0 to 200 µM, the tightest binder engaging the RBD at 2.0 µM. The motif and symmetric modes serve active-site grafting, epitope presentation, and nanocage design.

#Impact

OriginFlow's contribution is a reproducible route from backbone generation to a laboratory-testable binder, and micromolar binding across three unrelated targets without high-throughput screening is a useful data point for how far in-silico filtering can carry binder design. The affinities are modest: the authors attribute this to separating backbone generation from sequence design and note that micromolar binders need affinity maturation before use. The headline success rate is also reported as both 89% and 90% at different points in the manuscript, reflecting the seventeen expressed and twenty SPR-measured designs. The work remains a preprint, the weights sit on Google Drive rather than a versioned registry, and the absence of a repository license leaves reuse terms unsettled.