Understand
what supports the result.
Data sources, browser and server storage, and conclusions the current research beta does not establish.
Sources and calculations
UniProt — sequences and functional annotations; AlphaFold DB supplies existing predicted structures; FireProtDB provides measured variant records. A linear Foldiff model performs the main selection, with a separate ThermoMPNN neural prediction when available.
The service does not run AlphaFold or use every database entry for each calculation. A new source release does not automatically update the training dataset or local lookup.
What stays in the browser?
Laboratory measurements and pH series are processed by JavaScript in the browser. Storage uses localStorage, exports use CSV and JSON, and transfer uses manual JSON import. These data are not sent to the server queue or used for retraining.
Local-series access depends on the device, browser profile and retained site data. This supports a quick service trial; collaboration needs a separate account with access controls.
What does the server receive?
For a new calculation, the server receives job parameters and your CDS if supplied. Results are accessible through a hard-to-guess identifier without individual authentication. Storage is limited to the latest 100 completed jobs.
The shared beta is intended for public research data. Private accounts, agreed retention periods and a channel for confidential materials are not connected yet. This page explains technical behavior; it does not replace a contract or a formal personal-data policy.
Accuracy and validation
No validated universal accuracy percentage is claimed. Predicted ΔΔG is not a guaranteed Tm or activity gain. The original structure is visualized, not a mutant model. Combinations and primers require review.
Means and SD describe entered measurements. They do not determine significance or distinguish technical from biological replicates. Independent evaluation needs laboratory pilots and agreed criteria.
Current model metrics and limits →