Ubicon

Ubicon is a deep learning framework for predicting which substrates a given E3 ubiquitin ligase targets for degradation, developed by the Shimizu lab in the Department of AI Systems Medicine at the Institute of Science Tokyo and released as a preprint in April 2025. E3 ligases are the specificity component of the ubiquitin-proteasome system: several hundred human E3s decide which proteins are marked for destruction, wiring that runs through cell cycle control, DNA damage response, and cancer. Experimental methods for finding E3-substrate interactions (ESIs) are low-throughput and context-dependent, so most of the interactome has never been measured.

The model is a three-way fusion. Sequence context comes from ESM Cambrian, but not from the frozen general-purpose model — the authors adapt a 300M-parameter ESM-C checkpoint with Low-Rank Adaptation and a contrastive objective over known ESI pairs, producing representations that beat the generic 6B-parameter ESM-C on every metric. Structural context enters as 3Di structural-alphabet token frequencies computed with Foldseek over AlphaFold 2 predicted structures. Cellular context enters as predicted localization probabilities across ten compartments, on the reasoning that two proteins that never share a compartment cannot interact however compatible their surfaces look.

Once trained, the same fixed model is applied across the proteome to score candidate ligase-substrate pairs, producing a calibrated interactome map the authors call the ESIome and using it to identify cancer-associated rewiring.

#Key Features

• Task-adapted protein language model: LoRA modules (rank 4, alpha 32, dropout 0.1) trained contrastively on known ESI pairs turn a 300M-parameter backbone into a specialist that beats a generalist twenty times its size — adaptation outperforming scale.
• Three-modality feature fusion: Adapted sequence embeddings, Foldseek 3Di structural tokens, and DeepLoc 2 subcellular localization probabilities are concatenated into one feature vector, each modality contributing measurable gains.
• Calibrated probability outputs: Raw scores pass through an isotonic regression calibrator fitted on out-of-fold predictions, so a score of 0.8 means an empirical 80% chance of a true interaction — a prerequisite for prioritizing wet-lab experiments.
• Proteome-scale interactome map: The trained model scores candidate pairs between 486 human E3 ligases and roughly 20,000 reviewed human proteins, producing a full ESIome atlas rather than predictions for a handful of curated targets.
• Cancer network analysis: Intersecting the predicted interactome with TCGA expression data links the overexpressed ligases UHRF1, AURKA, and CDC20 to distinct substrate sets and to poor patient survival.

#Technical Details

The positive training set comprises 2,727 high-confidence human ESI pairs curated from UbiBrowser 2.0; negatives are 28,125 physically interacting protein pairs from BioGRID with all known ESIs removed, making the task discrimination between "interacts" and "interacts as ligase and substrate" rather than the far easier interacts-versus-random. Sequence embeddings, 3Di frequencies, and a ten-dimensional localization vector feed a gradient-boosted tree classifier; CatBoost beat LightGBM, XGBoost, Random Forest, TabNet, and TabPFN in five-fold cross-validation, and all of them beat an initial CNN.

The final pipeline reaches a mean AUROC of 0.9305 ± 0.0035 and AUPRC of 0.6812. The AUPRC is the meaningful figure given the roughly 1:10 class imbalance, and it improves substantially on DeepUSI, the previous leading sequence-only method, at 0.4445 on the same data. For the proteome-wide sweep, DeepTrio first filters the ligase-by-proteome search space down to pairs with a plausible physical-interaction prior, after which Ubicon scores nearly six million candidate pairs; the published network retains edges with a calibrated score of at least 0.8.

#Applications

The most direct use is target selection for induced protein degradation. Designing a PROTAC or molecular glue requires knowing which E3 ligase can be recruited to a protein of interest, and Ubicon supplies ranked, calibrated candidates rather than the narrow set of ligases current degraders reuse by convention. Ubiquitin biologists can mine the atlas for testable hypotheses about uncharacterized ligases, since the model scores pairs it never saw in training. Cancer researchers can intersect the network with tumor expression data to find degradation pathways rewired in a disease context, as the authors demonstrate for breast cancer.

#Impact

Ubicon's clearest contribution is methodological: a parameter-efficiently adapted mid-size protein language model beats a much larger frozen one on a specialized interaction task, and adding cheap non-sequence context — a predicted structural alphabet and a localization prior — yields further gains pure sequence models cannot reach. The proteome-wide ESIome is distributed as supplementary data with the preprint. Several caveats bound its use. The work is a single-lab preprint that has not been peer reviewed. Code and weights were released after the preprint, under an MIT licence: the repository ships the LoRA adapter parameters, the CatBoost classifier, the isotonic calibration model, and a command-line predictor that scores any ligase-substrate pair, alongside a Colab demo. What it does not include is the evaluation code or the curated interaction set behind the reported accuracy, so the headline metrics cannot be recomputed independently. Predictions are entirely computational, with no experimental validation of novel ESIs. And because the positive set is drawn from curated literature, the model inherits that literature's bias toward well-studied ligases — the same bias a proteome-wide map is meant to correct.