Predicts 39 post-translational modification types on every residue of a full-length protein from ESM-2 embeddings and AlphaFold2 structural features.
No providers recorded yet. Browse all providers
A multi-PTM predictor trained on raw literature data learns mostly phosphorylation, acetylation and ubiquitination, because those chemistries dominate every curated database. Rare modifications such as neddylation, sumoylation or S-nitrosylation contribute a handful of sites each, and a single loss over all labels simply learns to ignore them. Most existing tools respond by narrowing scope to a few well-studied modifications, or by scoring short peptide windows that cut away the protein-level context a modification may depend on.
AstraPTM2, from Orbion in Berlin, keeps both the breadth and the full-length context and addresses the imbalance directly in how it trains. A single residue-level model scores 39 PTM types across entire protein chains, and training introduces labels in stages from common to rare, with a focal loss that re-weights hard and rare positives and per-label calibration applied afterwards. It is the successor to AstraPTM, which covered 25 modification types from ESM-2 embeddings alone. Version 2 adds structural features taken from AlphaFold2 predictions and protein-level physicochemical descriptors. Çağlar Bozkurt, Alexandra Vasilyeva and Aniruddh Goteti described it in a bioRxiv preprint in October 2025.
Each residue is represented by a 1,377-dimensional feature vector: a relative position channel, a one-hot amino acid, Orbion's protein- and residue-level enrichment descriptors, a six-channel AlphaFold2 block drawn from the AlphaFold Protein Structure Database, and the ESM-2 esm2_t33_650M_UR50D embedding. Three depth-wise dilated convolutions (kernel sizes 7, 5 and 3) add local motif context. A global gate built from the mean-pooled sequence vector modulates each token, and eight transformer blocks with 24 attention heads and rotary position embeddings follow. Outputs go to grouped linear heads for common, medium and rare labels, plus a dedicated phosphorylation head for STY, HK and any-residue signals. Auxiliary site-window and kinase-group classification tasks add supervision, and an optional graph-convolution branch reads AlphaFold2 contact graphs. The training set is a label-balanced subset of 100,000 proteins drawn from a curated corpus of 237,813, split 70/15/15, with rare labels oversampled in the training split only.
On the held-out test split, AstraPTM2 reaches AUROC 0.99 overall (median 0.95 per label) and macro-F1 of 59% across all 39 types. Performance is uneven in an instructive way. Motif-driven chemistries score highest, including sumoylation (98.8% F1), O-linked glycosylation (97.6%) and methylation (91.5%). Broad, heterogeneous modifications score lowest: phosphorylation at 41.8%, acetylation at 32.1% and ubiquitination at 19.6%. The authors report a 9–14 point F1 gain over AstraPTM. They also note that label sets and metrics differ too much across published tools, among them MeToken, for head-to-head comparisons to be clean.
AstraPTM2 fits the planning step before an experiment. Examples include flagging glycosylation sites to remove before crystallization, choosing an expression host that can install the modifications a protein needs, and picking which residues to mutate or target by mass spectrometry. Its breadth matters most for less-studied chemistries, which dedicated single-PTM tools rarely cover. The authors recommend calibrated mode for mutagenesis decisions and exploratory mode for triaging poorly annotated regions.
AstraPTM2 is a commercial model with no released code, weights or training dataset. Predictions come only through Orbion's hosted platform, which renders them in 2D and 3D views and exports annotated PDB files stamped with the model version and calibration set. The benchmark figures therefore cannot be reproduced independently. All of them come from a test split drawn from the same curated corpus as training. The preprint's lasting contribution is methodological: its ablations report that a balanced 100,000-protein subset outperformed the full 230,000-protein corpus, that staged curriculum added more than 10 macro-F1 points, and that depth-wise convolutions gave the largest architectural gains by capturing short kinase motifs that attention alone under-represented. The preprint has not completed peer review.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.