AstraPTM

Post-translational modifications are the layer of protein regulation that sequence alone does not reveal. Phosphorylation, glycosylation, ubiquitination, acetylation and dozens of rarer chemistries change a protein's stability, localization and interaction partners, and mapping them experimentally by mass spectrometry remains slow, incomplete and biased toward abundant proteins. Computational PTM predictors have existed for two decades, but most share two structural limitations: they train a separate model per modification type, and they operate on short fixed-width windows around a candidate residue, which discards the long-range context that governs many modification events and makes very large proteins awkward to handle.

AstraPTM, from Orbion in Berlin, is a transformer-based framework built to remove both constraints at once. A single model scores every residue of an input protein for 25 different PTM types in one forward pass, with no ceiling on sequence length. It was described in a bioRxiv preprint in February 2025 by Aniruddh Goteti and Çağlar Bozkurt, and is served through the company's hosted Astra prediction platform rather than as downloadable code.

Its place in the landscape is on the annotation side of PTM modeling — the complement to generative work such as PTM-dCN, which designs sequences around chosen modification sites rather than reading modifications off existing ones. Where earlier tools ask "is this specific serine phosphorylated?", AstraPTM asks the whole-protein question in a single call.

#Key Features

• Twenty-five PTM types in one pass: A single unified model covers 25 modification chemistries instead of maintaining a separate per-modification classifier, so a protein is annotated comprehensively in one inference.
• No sequence length limitation: Full-length proteins are processed whole rather than chopped into fixed windows, preserving the long-range context that short-window predictors discard.
• Two-stage prediction head: A binary module first decides whether a residue is modified at all, and a multi-label module then assigns which modification types apply — a factoring that keeps sensitivity for rare chemistries while the binary gate handles the heavy class imbalance.
• Protein language model embeddings: Residue representations come from ESM-2, so the model inherits evolutionary sequence context learned from large protein corpora rather than learning motif grammar from PTM labels alone.
• Hosted inference: Predictions run through Orbion's Astra platform, which removes the local GPU and environment setup that self-hosted PTM predictors require.

#Technical Details

AstraPTM builds on embeddings from ESM-2, the protein language model family from Meta, and feeds them to a transformer that captures both short-range sequence motifs — the canonical kinase and glycosyltransferase recognition patterns — and long-range interactions across the full chain. Prediction is factored into two coupled outputs: a binary classification module indicating whether a given residue carries any modification, and a multi-label module that resolves which of the 25 PTM types are present. Training used dbPTM, a high-coverage aggregation of experimentally reported modification sites drawn from the published literature and from public proteomics repositories.

On benchmark comparisons against MusiteDeep and MIND-S, two established multi-PTM predictors, AstraPTM reported competitive or superior performance, with AUC-ROC above 99% for well-represented modifications such as phosphorylation and glycosylation. Those headline figures should be read with the class imbalance in mind: AUC-ROC is generous on tasks where negatives vastly outnumber positives, and the paper is explicit that the well-represented chemistries carry the strongest numbers while rarer modifications are harder. No trained weights or source code accompany the preprint, so independent reproduction of these benchmarks is not currently possible.

#Applications

The natural use is proteome-wide annotation: taking a set of proteins of interest — a pathway, a target family, an entire organism's predicted proteome — and generating a residue-level map of likely modification sites to prioritize which ones deserve targeted mass spectrometry. Because the model returns all 25 modification types on the same residues, it also supports crosstalk questions, where one modification blocks or licenses another on a neighboring site, which per-modification models cannot address without stitching independent predictions together. For protein engineers, the same output flags modification-prone residues that may complicate expression in a heterologous host or affect the developability of a biologic candidate.

#Impact

AstraPTM is a commercial model, and that shapes both its reach and its evaluation. There is no public repository, no released checkpoint, and no HuggingFace model card; access is through Orbion's hosted platform, which means the results in the preprint cannot be independently reproduced and the model cannot be fine-tuned or audited by users. Set against that, the architectural argument — one multi-task model over full-length sequences, rather than an ensemble of window-based per-modification classifiers — is a real contribution to how PTM prediction is framed, and the design has continued to develop: a successor, AstraPTM2, extends coverage to 39 modification types and adds predicted structural features alongside the language model embeddings. The preprint has not completed peer review.