Machine learning scores short protein stretches that mark proteins for destruction

The work

Technical · from the record

The authors built MetaDegron, a pair of models that score short protein sequence segments as E3 ligase targeted degrons or as random peptides, trained on more than 300 curated degron instances and a length-matched random peptide background. MetaDegron-X is an ensemble of ten XGBoost classifiers over ten computed structural, evolutionary and physicochemical features, and MetaDegron-D is a deep network that works from sequence alone using embeddings from the SeqVec protein language model together with a jointly trained amino acid embedding branch. In five-fold cross-validation the average AUC was 0.87 for MetaDegron-X and 0.90 for MetaDegron-D, with independent-test AUCs of 0.86 and 0.90; on an additional dataset of experimentally validated β-TrCP2 substrates the reported AUCs were 0.9705 for MetaDegron-D, 0.9670 for MetaDegron-X and 0.9540 for Degpred. The models are deployed as a web server that predicts degrons for 21 E3 ligases in batch and annotates degron-related features.

How AI was used

Degron motifs and instances were collected from the ELM database and prior studies, and a background set of length-matched random peptides was assembled. Ten features per instance — flexibility, solvent accessibility, secondary structure, disorder, anchoring score, conservation, domain and PTM context — were computed with existing structural bioinformatics tools, and the dataset was split 90/10 into training and independent sets. On the feature representation, ten XGBoost classifiers were trained under a bootstrapping strategy with parameters selected by cross-validation, and their scores averaged to give a degron probability (MetaDegron-X). A second, sequence-only model (MetaDegron-D) combined 1024-dimensional per-residue embeddings from the pre-trained SeqVec protein language model, passed through BLSTM and convolution-pooling layers, with a separately trained context-insensitive amino acid embedding feeding its own BLSTM branch; the two branches were joined by dense layers ending in a two-node output for the degron and random peptide classes. Evaluation used five-fold cross-validation, the independent split, a feature-elimination study, and a separate dataset of β-TrCP2 substrates with repeated random background sampling against the published Degpred predictor. Both trained models are served from a web application that accepts FASTA input and selected E3 ligases and returns scored degron instances with feature and structure annotations.