GPFlow

GPFlow is a generative framework for protein design that treats the length of a protein as part of what the model generates, rather than a fixed quantity chosen in advance. In most generative protein models, the number of residues must be specified before sampling, or supplied by an external length oracle. This is a poor fit for real design problems, where the optimal length of a scaffold, binder, or peptide is itself unknown and part of what a designer wants to discover. GPFlow addresses this by learning the rate function of an inhomogeneous generalized Poisson process, yielding a single generative model that samples proteins of varying length without access to a native-length oracle.

The method's central idea is a variable-dimensional probability flow that couples a process over sequence length with within-length dynamics, allowing the same formulation to operate over continuous Euclidean coordinates, Riemannian quantities such as backbone rotations and torsion angles, and discrete amino-acid tokens. Because these modalities are unified under one framework, a single trained model spans unconditional structure generation, unconditional sequence generation, conditional motif scaffolding, and peptide co-design, rather than requiring a separate architecture per task.

GPFlow was developed by Chaoran Cheng, Zhanghan Ni, Yanru Qu, Yuxin Chen, Ruihan Guo, Jiajun Fan, and Ge Liu at the Siebel School of Computing and Data Science, University of Illinois Urbana-Champaign, and released as a preprint in July 2026. It sits alongside other flow- and diffusion-based protein design models but is distinguished by its native handling of variable length across multiple data modalities.

#Key Features

• Variable-length generation: Models sequence length as part of the generative process through a generalized Poisson formulation, sampling proteins of differing sizes without a native-length oracle.
• Multi-modal probability flow: A single variable-dimensional flow handles continuous Euclidean coordinates, Riemannian rotations and torsion angles, and discrete categorical tokens.
• Unified task coverage: One framework performs unconditional structure and sequence design, motif scaffolding, and receptor-conditioned peptide co-design.
• Theoretical grounding: The formulation provides guarantees on distribution recovery and bounds on KL divergence to the target distribution.
• Strong in-silico benchmarks: Leads on 10 of 16 structure-based motif scaffolding tasks and improves designability over strong flow and diffusion baselines.

#Technical Details

GPFlow builds a unified probability flow from inhomogeneous generalized Poisson processes combined with within-length dynamics, with distinct parameterizations for Euclidean, Riemannian, discrete, and mixed length-dependent modalities. The structure model has roughly 65M parameters, while the sequence model is a 642M-parameter diffusion-transformer (DiT); the peptide co-design setup adapts the PepFlow architecture. Structure models are trained on 27k single-chain PDB entries (lengths 50–256) and 713k AlphaFold Database samples, sequence models on 41M UniRef50 sequences (lengths 10–1024), motif scaffolding on AFDB with motif augmentation, and peptide co-design on a receptor-pocket benchmark from Li et al. (2024).

On unconditional structure generation, GPFlow reports 96.1% designability when trained on the PDB and 94.5% on AFDB, versus 77.6% and 90.6% respectively for Proteina, alongside novelty (TM-score) near 0.75. On motif scaffolding it ranks first on 10 of 16 tasks and averages 117.4 unique successes across 1000 generations, compared with 94.9 for Proteina, 64.8 for Genie2, and 41.5 for RFDiffusion, with up to a tenfold gain on challenging targets. For sequence generation, its pLDDT distribution aligns more closely to UniRef50 than DPLM, and in peptide co-design it achieves 53.71% amino-acid recovery, 1.39 Å RMSD, and 86.04% binding-site recovery.

#Applications

GPFlow is aimed at protein and peptide design workflows where length is a free parameter. Designers can generate de novo backbones and sequences of varying size, scaffold functional motifs into proteins whose overall length is not predetermined, and co-design peptide binders against a target receptor pocket. Because the same model covers structure, sequence, and conditional generation, it can serve as a shared backbone across an early-stage design campaign — from unconditional exploration of the structural landscape to motif- and receptor-conditioned generation of candidate therapeutics or functional proteins for downstream computational screening.

#Impact

GPFlow demonstrates that variable-length generation across continuous, Riemannian, and discrete modalities can be handled by a single principled flow, unifying tasks that are typically split across bespoke models. Its reported gains on motif scaffolding and designability position it as a competitive entry among flow- and diffusion-based protein design methods. As a preprint awaiting peer review, its results are established through in-silico benchmarks rather than experimental validation, and no public code or model weights have been released, which currently limits independent reproduction and direct adoption.