deepGPS

Ask a subcellular localization predictor where a protein goes and it hands back a word: "nucleoplasm", "cytoplasm". A cell biologist checking that answer reaches instead for a fluorescence micrograph — a shape, a texture, a signal diffuse or punctate relative to the nucleus in the same field. The word and the picture are not interchangeable kinds of evidence, and text-to-text predictors stop short of the second.

deepGPS makes the picture an output. Given a protein's primary sequence together with a nucleus fluorescence image that supplies the spatial frame, it emits both a textual localization label and a synthetic fluorescence image showing where that protein should appear in that particular cell. Sequence understanding comes from a frozen ESM-2 650M encoder; the image is assembled by a U-Net whose bottleneck is fused with the sequence vector through cross-attention, so the generated pattern is conditioned on the protein rather than copied from the input nucleus. A cross-assay makes that concrete: holding one cell's nucleus image fixed and swapping in the sequences of AGO1, FAM120A, HNRNPD and SMARCD2 produces cytoplasmic patterns for the first two and nuclear patterns for the second two.

The model was developed by Li Yang's group at Fudan University with Nanqing Dong and Wanli Ouyang at Shanghai AI Laboratory and the Shanghai Innovation Institute, and published in Briefings in Bioinformatics after a 2024 bioRxiv preprint.

#Key Features

• Text-to-image localization: One forward pass returns a classification label and a 208 × 208 pixel artificial fluorescence image, giving a visual prediction that can be compared directly against an experimental micrograph.
• Frozen sequence backbone: ESM-2 650M serves as a fixed feature extractor, mean-pooled to one 1280-dimensional vector per protein, so no language model training is repeated here.
• Organelle resolution in tiers: deepGPS-single-2 separates cytoplasm from nucleoplasm, deepGPS-single-4 adds vesicles and endoplasmic reticulum, and deepGPS-all generates across all 17 annotated organelles.
• Cell-line-specific checkpoints: Training on HEK293T images from OpenCell and on U2OS images from the Human Protein Atlas yields distinct models, so the same protein's predicted distribution can be compared across cell backgrounds.
• Hosted prediction and data return: The openGPS web platform serves the trained checkpoints for arbitrary query sequences and accepts user-contributed experimental images, both in batch.

#Technical Details

A protein sequence of length L becomes an L × 1280 ESM-2 embedding, mean-pooled to 1280 dimensions and projected by a linear layer to 1024. That vector drives a four-layer MLP classification head and, separately, acts as key and value in an attention step whose query is the U-Net encoder's 1024 × 169 image bottleneck; the decoder up-samples the fused feature back to the input resolution. Training data came from 6,239 paired OpenCell images of 1,301 endogenously tagged HEK293T proteins under 2,700 amino acids — normalized with CSBDeep, nucleus-segmented with StarDist, cropped to 62,108 tiles and split 4:1 with proteins held mutually exclusive between train and test. On the 354-protein cytoplasm/nucleoplasm task, deepGPS-single-2 reaches 0.744 accuracy, 0.761 AUROC and 0.838 AUPRC, with mean PSNR 14.5 and SSIM 0.20 on generation; deepGPS-single-4 reaches 0.81 accuracy and 0.83 AUROC across four classes. Retraining deepGPS-single-2 on 2,087 proteins and 42,204 crops from U2OS raises accuracy to 0.797 and F1 to 0.793, above MULocDeep (0.763 / 0.697) and DeepLoc v1.0 (0.731 / 0.667) on the same test set. Adding AlphaFold 2 structures as PyUUL point clouds encoded by PointNet improved neither task, which the authors read as evidence that the sequence embedding already carries the signal.

#Applications

The practical use is triage before imaging: a researcher with a protein of interest gets a predicted label and an expected fluorescence pattern, then decides whether an endogenous tagging or immunofluorescence experiment is worth running and what the result should look like. Because openGPS accepts batches, the workflow scales to a candidate list from a screen or an interactome, and its "Submit" module takes user-contributed localization images for side-by-side comparison with the generated ones. Predictions for CPSF6, CSTF2 and EEF2 — proteins absent from the training set and imaged in HeLa rather than HEK293T — matched their known compartments in both label and image.

#Impact

deepGPS is best understood as a working prototype for generative localization prediction rather than a finished production tool. Its classification numbers are competitive with, not dominant over, established text-to-text predictors, and its generated images reproduce overall shape more reliably than fine texture, especially for organelles with few training proteins — mitochondria and cytoskeleton have roughly ten single-localized proteins each in OpenCell. What it establishes is the pairing itself: a general protein representation supplying the biology and an image decoder supplying a readout in the format the wet lab already uses. The repository holds training code only, with no released checkpoints and no license file, so the openGPS website is the sole route to running the trained models.