Method
How the atlas is made
A short, plain explanation of where the proteins come from and how each prediction is produced.
1. Which proteins
We start from the Tdark list of the Pharos database (NIH Illuminating the Druggable Genome). Tdark means a human protein with almost no published research, no known drugs and little functional annotation. Today that list has 5,501 proteins. The atlas currently shows 705 proteins: 700 dark proteins and 5 well-known controls used to check the model, chosen from different areas of biology (enzymes, receptors, transporters, structural, immune, neuronal and more).
2. Inputs per protein
For each protein we collect its amino-acid sequence from UniProt, the known building blocks (domains) from InterPro, and the predicted 3D shape from the AlphaFold database.
3. The model
BioReason-Pro (Wang Lab) reads the sequence and domains, uses the ESM3 protein language model (esm3-sm-open-v1) to understand the structure, and then writes out its reasoning, a function summary and Gene Ontology terms (what it does, which process, where in the cell).
4. Checking it
We include proteins whose function is already known as controls. If the model gets those right, we can trust its guesses on dark proteins a bit more. It does not give a confidence score, so every page is labelled as an AI prediction.
Limits
These are hypotheses, not discoveries. For dark proteins the answer often follows the protein's family, and it can include speculative links. Nothing here has been verified in a lab.