ProtFlow

ProtFlow is a generative model for protein and peptide sequence design that applies rectified flow matching to a compressed latent space derived from a protein language model. It was developed by researchers in the College of Computer Science and Technology at Zhejiang University, first posted to arXiv in April 2025 and later expanded in a February 2026 bioRxiv manuscript focused on antimicrobial peptides. Where most sequence-design methods rely on autoregressive language models or discrete diffusion, ProtFlow uses flow matching — a continuous-time generative paradigm that transports noise to data along straight (rectified) paths — and adds a reflow stage that makes high-quality single-step generation possible.

The central design choice is where generation happens. Rather than modeling amino-acid tokens directly, ProtFlow encodes sequences with ESM-2, then compresses and smooths that embedding space so a continuous flow can operate on it stably. The bioRxiv version frames this as a semantic integration network that reorganizes the language model's representation space to capture the global semantic organization of protein space rather than local compositional statistics — the failure mode the authors argue causes existing generators to miss rare but functionally important regions of the distribution.

The project spans a wider task set than most single-purpose sequence generators: general peptides, long-chain proteins, antimicrobial peptides, and antibodies, the last handled through a joint multichain design pipeline that generates paired heavy and light chains together.

#Key Features

• Rectified flow matching with reflow: Straight probability paths let the model match diffusion-quality output in 25 ODE steps, and a reflow stage distills this to a single Euler step.
• Compressed language-model latent space: A learned compressor-decompressor pair reduces ESM-2 embeddings 16-fold while preserving over 99.9% reconstruction accuracy, improving distribution learning rather than degrading it.
• Multichain joint design: Separate encoders and decoders for antibody heavy and light chains feed concatenated embeddings through one flow model, so paired chains are designed jointly instead of independently.
• Antimicrobial-peptide fine-tuning: Adapting the pretrained model to a 195,000- sequence AMP corpus yields peptides with desired activity profiles across multiple pathogens, including underrepresented bacterial species.
• Released training code: The official implementation is public under an MIT license, covering decoder, compressor, and flow-matching training plus generation.

#Technical Details

ProtFlow freezes an ESM-2 encoder (8M and 35M parameter variants were both tested), fine-tunes a matching decoder, and inserts a compressor built from transformer and pooling layers that shrinks the embedding by a factor of 16 after z-score normalization, saturation truncation, and min-max smoothing. The flow-matching network is a 12-layer transformer with U-ViT-style long skip connections. Pretraining used 2.6 million UniProt peptides of length 2-50 and 470,000 SwissProt proteins; the AMP model was fine-tuned on 195,000 sequences from dbAMP, AMP Scanner, and DRAMP, and the antibody model on paired OAS sequences. On UniProt peptides the reflow variant reaches 72.86 OmegaFold pLDDT and 12.16 ESM-2 pseudo-perplexity in one function evaluation, against 100 steps for the continuous-diffusion baseline DiMA and 1,000 for EvoDiff. On AMP design it reports 12.24 perplexity and 0.84 and 0.63 classifier-predicted antimicrobial and anti-E. coli rates, ahead of AMP-Diffusion, HydrAMP, and AMPGAN; on antibodies it achieves the lowest physicochemical Wasserstein distance to the reference distribution. The code repository is MIT-licensed but ships training scripts only, without pretrained checkpoints; the bioRxiv manuscript is posted under CC BY-NC-ND.

#Applications

ProtFlow is intended for researchers designing functional peptides and small proteins, with antimicrobial peptides as the most developed use case and antibody variable regions as the most complex. Such models help triage and propose candidate sequences computationally — AMPs targeting drug-resistant pathogens, or paired antibody chains matching a target's biophysical property distribution — before synthesis and experimental assays, narrowing large design spaces to promising leads. The one-step generation path is the practical draw: sampling large candidate libraries costs a single network evaluation per sequence.

#Impact

ProtFlow was among the first flow-matching generative models built specifically for protein sequence design, and its main transferable result is that compressing a protein language model's latent space improves rather than harms distribution learning, up to a 16-fold ratio. Its evaluation is entirely in silico — perplexity, predicted structure confidence, and classifier-scored activity — so experimental validation of the designed peptides remains open. Without released weights, adopting it means retraining from the published code.