Loading…
Phase separation prediction from sequence alone, pairing a protein language model with MD-trained conformational features to score every residue.
No providers recorded yet. Browse all providers
Whether a protein forms a biomolecular condensate is usually decided by one short stretch of it — a low-complexity domain, a charge-blocky disordered region, an RNA-recognition motif that engages a partner. A predictor returning a single score for the whole chain answers the easier question and leaves the biologist without the coordinates to truncate or mutate. The predictors that do best on that easier question generally buy their accuracy with immunofluorescence imaging, post-translational modification records, interaction data and functional annotations — none of which exist for an uncharacterized sequence, a rare isoform, or a frameshifted variant.
PSTP (Phase Separation's Transfer-learning Prediction) takes sequence as its only input and recovers residue resolution from protein-level labels. Each residue is described by two frozen pretrained encoders: ESM-2, carrying the evolutionary and compositional grammar of the sequence, and ALBATROSS, recurrent networks trained on coarse-grained molecular dynamics ensembles that report how compact or expanded a disordered chain is. Chain compaction is the physical property most tightly linked to phase separation — compact disordered regions favour it, expanded ones disfavour it — and it is precisely what masked-residue pretraining does not represent directly. Concatenating the two gives a 650-dimensional vector per residue.
The head that consumes it, PSTP-Scan, is deliberately small. It averages the embedding over a local window at every position, passes the result through a shared two-layer perceptron for a per-residue score, and takes the maximum over the sequence as the protein's predicted propensity. Because that maximum is what backpropagation updates, the network has to localize the driving region in order to classify the protein at all, so a residue-level track falls out of protein-level supervision. PSTP was developed at the Bio-X Institutes of Shanghai Jiao Tong University.
Each PSTP-Scan block applies average pooling with symmetric padding, a shared trainable perceptron of 20 then 5 neurons with a sigmoid output, and a max-pool over positions; three parallel blocks with pooling widths of 257, 129 and 33 capture driving regions from tens to hundreds of residues long. Training used binary cross-entropy, Adam at a learning rate of 0.003, 50 epochs, and a LASSO-style penalty applied separately to the language-model and conformational halves of the input. Supervision came from the PhaSePred curation — 201 self-assembly and 327 partner-dependent phase-separating proteins against 60,220 background sequences — plus a pooled set of 892 proteins drawn from LLPSDB, PhaSepDB, DrLLPS and PhaSePro. Held-out AUCs run around 0.89 for self-assembly and 0.80 for partner-dependent proteins, and roughly 0.9 for PSTP-Scan against PhaSePred, DeePhase, PSPHunter, PLAAC, PSAP, ParSe, catGRANULE, FuzDrop and PScore. The conformational half carries real weight: using it alone gives 0.81 and 0.72, and dropping it degrades performance across every evaluation.
The residue track is the practical output — it names the segment to delete, truncate, or swap when testing a condensate hypothesis at the bench, and it scores artificial polypeptides and truncated isoforms that have no homologs to align against. At proteome scale, PSTP scored more than 42,000 human proteins and isoforms and over 570,000 UniProt-reviewed entries, finding potential phase-separating regions in roughly 35% and 27% respectively. Its highest-value use is variant interpretation in poorly conserved disordered regions, where evolutionary predictors have little signal: across 4,756 pathogenic and 37,291 benign ClinVar variants there, pathogenic ones sit at systematically higher predicted propensity.
PSTP reframes condensate prediction as a localization problem rather than a classification one, and shows the localization can be learned for free from coarse labels. Its second contribution is methodological: pairing an evolutionary sequence encoder with a simulation-trained conformational one gives disordered-region models a physical channel that language models alone lack. The caveats are real. Partner-dependent phase separation stays harder than self-assembly, because predicting an interaction site requires knowing the partner; the definition of a "phase-separating region" varies enough between databases to cap agreement; and the learned component is small and single-task. The reusable artifact is as much the 650-dimensional dual embedding, which the package exposes directly, as the classifier built on it.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.