Open Research
Science Compounds in the Open
Every Astra model that powers Orbion is documented in an open preprint — methods, benchmarks, and head-to-head evaluation against public baselines. We put the science in front of the field to reproduce, challenge, and build on.
Prospective Validation
Tested Blind, at the Bench
Benchmarks show the models work across families. This shows one working prospectively — a wet-lab case study where the platform designed a construct, scored it before the experiment, and a partner lab put it to the test.
Model Performance Series
Benchmarked Across 18,367 Proteins
Five studies putting Astra head-to-head with public baselines — family by family, receptor by receptor. Each reports the methods, the metrics, and where the models win and where they don't.
Model Preprints
The Paper Behind Each Model
Peer-facing preprints for the models inside the platform — architecture, training signal, and evaluation, released openly on bioRxiv.
Beyond Structure and Affinity: Context-Dependent Signals for de novo Binder Success
Çağlar Bozkurt
A reanalysis of public benchmarks separating transferable from context-specific sequence signals for binder-design success.
AstraBIND: Graph Attention Network for Predicting Ligand Binding Sites
Aniruddh Goteti, Alexandra Vasilyeva, Çağlar Bozkurt
A lightweight graph neural network combining sequence, structure, and homology to localize ligand-binding sites.
AstraPTM2: A Context-Aware Transformer for Broad-Spectrum PTM Prediction
Çağlar Bozkurt, Alexandra Vasilyeva, Aniruddh Goteti
Predicts 39 post-translational modification types from sequence embeddings, structural features, and protein-level context.
AstraROLE2 & AstraSUIT2: Multi-Task Annotation Models for Functional Profiling of Proteins
Çağlar Bozkurt, Alexandra Vasilyeva, Aniruddh Goteti
Multi-task models that profile function, pathways, cofactors, domains, and localization directly from sequence.
Explorations
Questions We're Chasing
Not every idea is a benchmark yet. Notes and honest write-ups from the edge of what we're working on.
Can Quantum Computing Actually Improve Protein-Ligand Binding Prediction?
Aniruddh Goteti
An honest look at quantum-derived descriptors for binding prediction — where they help, and where they don't.
Calibration, Not Compounding: What Experimental Feedback Did for a Protein-Stability Model
Çağlar Bozkurt
Whether experimental feedback compounds across proteins — it calibrates per target, and does not transfer.
The Same Models Run on Your Targets
Everything documented here is what scores your constructs inside Orbion. Bring your hardest target and see how the predictions land against your own data.