Phaseek

Liquid-liquid phase separation lets cells build compartments without membranes. Proteins driven by weak multivalent contacts demix into dense condensates that concentrate machinery for RNA processing, DNA repair, and carbon fixation, then dissolve when the job is done. When that goes wrong it shows up in neurodegeneration, cancer, and viral infection, which makes condensates a therapeutic target and their prediction a practical problem.

Existing LLPS predictors mostly score a fixed set of hand-chosen physicochemical features, and they generalize poorly: a model tuned on canonical disordered condensate proteins tends to fail on synthetic designs, on folded proteins that phase separate anyway, and on disordered proteins that do not. Phaseek, from Amir M. MohammadHosseini, Hossein Teimouri, Amir Pandi, and colleagues at Inserm with Sorbonne University, the CNRS, and the University of Washington, targets that gap by learning sequence context rather than curating features. It runs a transformer encoder over the raw sequence in parallel with statistical protein graphs derived from amino acid physicochemistry, fusing both into residue-level and sequence-level LLPS scores.

Phaseek is also generative: the released implementation optimizes a sequence directly against the model's own score, turning the predictor into a designer of phase-separating peptides. Posted to bioRxiv in January 2025 and revised that August, it ships with trained weights and a Colab notebook.

#Key Features

• Two parallel encodings: A GPT-architecture encoder supplies contextual sequence features while a protein graph generator supplies statistical matrices, so the model is not restricted to a predefined feature list.
• Residue-level profiles: A length-adaptive sliding window with Gaussian-weighted aggregation locates the regions that drive phase separation rather than only classifying the whole protein.
• Mutation sensitivity: Scores respond to single substitutions, separating experimentally validated impactful mutations from random ones.
• Peptide generation: A trainable logit matrix is optimized by gradient descent with temperature-annealed Gumbel relaxation and entropy regularization, designing sequences that maximize predicted propensity.
• Proteome-scale inference: Batch FASTA prediction and overlapping-window analysis let a fixed checkpoint score entire proteomes without per-species retraining.

#Technical Details

The core model has three modules. The encoder follows a GPT architecture with 6 layers, 6 attention heads per layer, a 192-dimensional token embedding, and a block size of 512. In parallel, FEGS encodes each sequence into 158 graphical representations; a random forest interpreted with SHAP values selected the 10 most informative matrices, merged by weighted summation. Encoder and graph outputs are pooled, passed through an MLP, and reduced to a probability vector. An XGBoost bottleneck then boosts high-scoring regions in the final score — worth a jump in evaluation AUC from 0.84 for the decoder-only version to 0.93 for the full model. Training used AdamW with cross-entropy loss over roughly 1,500 positive and 5,000 negative curated sequences, positives augmented to about 5,000 with a hidden Markov model, split 9:1 and converging after 35 epochs.

On an evaluation set of 300 phase-separating and 337 non-phase-separating sequences, Phaseek reached AUC 0.93, against 0.90 for PSPredictor, 0.88 for PSPHunter, and 0.67 to 0.87 for PSTP, PScore, and catGRANULE 1.0 and 2.0. Four external datasets probed generalization, and Phaseek led on each: 0.89 accuracy on synthetic designed proteins, 0.85 on biotinylated-isoxazole-precipitated Arabidopsis proteins, 0.72 on fusion oncoproteins, and 0.62 on the hardest set, folded proteins that do phase separate paired with disordered ones that do not. Removing either known LLPS-driving segment of the human LINE-1 retrotransposable element dropped its score from 0.82 to 0.43 and 0.37, below the 0.7 classification threshold, matching the experiment showing that either deletion abolishes the phenotype. Across 2,740 mutations on 137 proteins, sequence-level and cumulative regional metrics separated impactful from random substitutions (p = 1.1e-19 and 4.6e-12) while the mutated-residue-only metric did not (p = 0.83).

#Applications

Phaseek serves both fundamental and translational condensate research: shortlisting candidate phase-separating proteins from a newly sequenced proteome, pinpointing which regions drive condensate formation before designing truncations, and predicting whether a disease-associated variant strengthens or weakens phase separation. Its generative mode addresses synthetic biology, where designed phase-separating tags build synthetic condensates and concentrate enzymes. The 18-species predictions were used to examine conservation of LLPS among orthologs and its association with Gene Ontology terms.

#Impact

By benchmarking on four deliberately awkward external sets — synthetic proteins, precipitated plant proteins, fusion oncoproteins, and the folded-versus-disordered confound — Phaseek makes generalization the evaluation target rather than in-distribution accuracy, the axis on which earlier LLPS predictors were weakest. Accuracy of 0.62 on that last set shows how far the problem remains from solved. The preprint has not been peer reviewed, and all evidence is computational: predictions are compared against previously published experiments rather than new assays, and the designed peptides have not been tested at the bench. Trained weights, code, and a Colab notebook are on GitHub and Hugging Face under a research-purposes-restricted license administered by Inserm Transfert, limiting use to academic and non-commercial research.