Research map / Probabilistic and scientific modeling / Molecules and proteins

Variable-Length Generative Protein Design via Generalized Poisson Flow

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Generative protein structure design · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces Generalized Poisson Flow (GPFlow), a variable-length generative framework that learns the rate of an inhomogeneous generalized Poisson process via Poisson NLL, applied to protein structure, sequence, motif scaffolding, and peptide co-design.

Problem

Protein generative models based on diffusion and flow matching require the sequence/structure length to be fixed before sampling, which limits exploration of feasible designs when the optimal length is unknown and coupled to designability.

Approach

Length evolves as an inhomogeneous generalized Poisson process whose rate is learned by minimizing the exact point-process negative log-likelihood (with insertion times analytically marginalized, yielding a simulation-free loss), coupled with within-length continuous, discrete, or Riemannian generators; a marginalization theorem gives population-level recovery of the joint distribution and a gener

Key results

• Structure model (65M, AFDB) achieves designability 94.5 and PDB-trained variant 96.1, vs Proteina AFDB 90.6 and PDB 77.6 in unconditional structure design.
• On the RFDiffusion motif scaffolding benchmark, GPFlow ranks first on 10 of 16 tasks with 117.4 average unique successes vs 94.9 for Proteina.
• On peptide co-design GPFlow achieves AAR 53.71%, RMSD 1.39 Å and affinity 21.03%, best on 4 of 8 metrics, without requiring native length.

Surprising findings

• Proteina-family models including GPFlow show lower diversity than FrameFlow and RFDiffusion in unconditional structure generation.
• Proteina's CA-only design remains a strong baseline on several long-length motif tasks such as 5TRV and 6E6R where it wins raw counts.

Limitations

• Sensitivity to scheduler choices, especially in multimodal setups where modalities interact.
• Sampling techniques such as tau-leaping and localized probability paths introduce subtle trade-offs.
• Diversity of Proteina-family models including GPFlow remains lower than some baselines in unconditional structure generation.

Experimental setup

Datasets: PDB AFDB UniRef50 RFDiffusion motif scaffolding benchmark peptide co-design benchmark (PepFlow)

Benchmarks: Unconditional structure generation (Proteina protocol) UniRef50 sequence statistics RFDiffusion motif scaffolding benchmark (16 tasks) peptide co-design benchmark

Metrics: designability (scRMSD<2Å) diversity novelty (TM-score) pLDDT pdb-TM unique successes AAR RMSD SSR BSR

Baselines: FrameDiff FrameFlow RFDiffusion Proteina EvoDiff-OADM DPLM ESM3 SCISOR Genie2 ProteinGenerator PepFlow

Builds on: Proteina EditFlow PepFlow TDDM flow matching

Closest papers by meaning

• Generative Flows on Discrete State-Spaces: Enabling Multimodal Flows with Applications to Protein Co-Design (ICML 2024 · similarity 0.89)
• SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · similarity 0.88)
• AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.88)
• Sequence-Augmented SE(3)-Flow Matching For Conditional Protein Generation (NeurIPS 2024 · similarity 0.88)
• A Variational Perspective on Generative Protein Fitness Optimization (ICML 2025 · similarity 0.88)
• Proteina: Scaling Flow-based Protein Structure Generative Models (ICLR 2025 · similarity 0.88)