ProtGPS

A protein bound for the nucleus carries a nuclear localization sequence — a short, readable address that biology worked out decades ago. Proteins bound for a biomolecular condensate carry no such tag. Nucleoli, stress granules, nuclear speckles, P-bodies and PML bodies have no membrane and no import machinery, yet each concentrates a specific and reproducible set of proteins. Whatever specifies that membership is distributed across the whole sequence rather than packed into a signal peptide, which is exactly the kind of signal a sequence model can learn and a human reader cannot.

ProtGPS treats compartment membership as a multi-label prediction over twelve destinations and learns it directly from primary sequence. Built at the Whitehead Institute with MIT CSAIL, it fine-tunes an ESM-2 backbone under a small classifier head that emits an independent probability for each compartment, so a protein that resides in both the nucleolus and chromatin scores for both rather than being forced into one bucket.

The result that distinguishes it is generative rather than predictive. Holding the trained weights fixed, the authors used ProtGPS as the scoring oracle inside a sequence sampler, searching for novel sequences the model believed would land in a chosen compartment. An initial greedy autoregressive search produced eight nucleolus-targeted designs that entered the nucleus but did not concentrate in nucleoli. Replacing it with Markov chain Monte Carlo sampling constrained to sequences ESM-2 finds plausible and DR-BERT scores as disordered produced designs that worked: all ten 100-residue nucleolar candidates concentrated in nucleoli when expressed in cells, and six of ten nuclear-speckle candidates were enriched in speckles.

#Key Features

• Twelve condensate destinations: nucleolus, nuclear speckle, P-body, PML body, stress granule, Cajal body, RNA granule, nuclear pore complex, chromosome, transcriptional condensate, post-synaptic density and cell junction.
• Multi-label, not single-class: independent per-compartment probabilities reflect the biology, where many proteins participate in more than one assembly.
• A frozen oracle for design: the released checkpoint scores candidate sequences inside a sampling loop, so generation updates the sequence rather than the model.
• Variant screening at database scale: predictions across 205,182 ClinVar mutations — 83,211 truncations and 121,971 point substitutions — quantify how far a mutation shifts a protein's predicted compartment distribution.
• Small enough to run anywhere: the backbone is the smallest ESM-2 release, so proteome-scale scoring needs no unusual hardware.

#Technical Details

Training used 5,541 human protein sequences with compartment annotations curated from UniProt and CD-CODE, split 70/15/15. The encoder is ESM-2 esm2_t6_8M_UR50D — six layers, roughly 8 million parameters, 320-dimensional representations — read out at layer 6 and passed to a two-layer 512-unit MLP with batch normalization and 0.1 dropout, trained under binary cross-entropy for up to 30 epochs with Adam at a 1e-3 learning rate over sequences up to 1,800 residues. The backbone is fine-tuned end to end rather than frozen, so the released checkpoint is a compartment-specialized protein language model rather than ESM-2 plus a probe. Across the twelve compartments, area under the ROC curve on held-out proteins ranges from 0.83 to 0.95. Inference code is MIT-licensed and the checkpoint is archived on Zenodo under CC BY 4.0; the identical checkpoint, with the hyperparameter dump the prediction notebook needs, also ships as a release asset on the MIT-licensed repository, so there is a permissively licensed route to the weights. Reproducing the full pipeline additionally requires the public ESM-2 and DR-BERT checkpoints, neither of which is vendored.

#Applications

The immediate use is annotation: scoring an uncharacterized protein, or a proteome, for condensate membership without an imaging experiment. The variant analysis extends that to disease, ranking mutations by how much they perturb a protein's predicted distribution — twenty candidates selected this way were expressed in mouse embryonic stem cells, where confocal imaging confirmed altered localization for most truncations and several point mutants. The generative path is the most consequential for engineering: a designer who wants a synthetic protein, degrader or sensor to accumulate in a particular condensate can sample against ProtGPS instead of grafting motifs by hand.

#Impact

Published in Science in 2025, the work argues that amino acid sequence carries a compartmentalization code alongside the folding code, and backs the claim with designed proteins that go where the model says they will. That closes a loop most sequence-property models leave open, since the prediction was tested by synthesis and microscopy rather than a held-out split. The limits are real: twelve curated compartment classes over a few thousand human proteins, evaluation on a random split of that same set rather than an independent cohort, and a prediction path that runs through a notebook with hard-coded paths rather than a packaged tool.