PhaseFlow | Yanjie Huang 黄䶮杰

PhaseFlow model architecture diagram — Model architecture: peptide sequence-phase learning, Flow Matching, and full-length protein context for LLPS prediction and design.

PhaseFlow application map — Application map: protein-level LLPS prediction, DPR scanning, peptide phase-diagram generation, and mutation-scale optimization.

PhaseFlow is a multimodal generative model for liquid-liquid phase separation (LLPS). It connects full-length protein understanding, peptide-scale phase diagrams, and mutation-level optimization in one workflow.

The core peptide module jointly models amino acid sequences and 4x4 PSSI phase diagrams. A Transfusion-style Transformer combines autoregressive sequence modeling with Conditional Optimal Transport Flow Matching over continuous phase values, enabling both sequence-to-phase prediction and phase-to-sequence design. A full-length GNN-Transformer branch then brings in long-range sequence, structural, disorder, and peptide-derived local phase evidence for protein-scale LLPS prediction and droplet-promoting region scanning.

What It Enables

Predicting full 4x4 LLPS phase diagrams from peptide sequences.
Generating novel peptide sequences conditioned on target phase behavior.
Predicting full-length protein LLPS propensity.
Scanning proteins for droplet-promoting regions.
Scoring mutation effects and steering in-silico directed evolution toward desired phase profiles.

My Role

I built data and quality-control pipelines for a high-throughput phase-separation dataset and developed a unified model combining sequence generation with flow matching for phase-diagram prediction. The model outperformed DDPM and Transformer-only baselines. I also implemented inference-time scaling for controllable design, improving self-consistency by 30%.

Release

The public repository provides the PhaseFlow codebase for training, inference, sequence-to-phase prediction, phase-conditioned sequence generation, and directed-evolution style optimization.