ProtScape preprint places all 18 withheld Parkinson's drug targets within the top 511

ProtScape preprint places all 18 withheld Parkinson's drug targets within the top 511

On September 14, a preprint describing ProtScape was posted on bioRxiv. The model builds separate protein interaction maps for 207 cell types and states. To test drug target discovery, the authors withheld 18 proteins already targeted by drugs with clinical evidence in Parkinson's disease. ProtScape placed all of them within the top 511 of its ranking; PINNACLE, the previous model for the same task, required 8,186 positions.

A conventional protein interaction map pools edges observed across different tissues and conditions. It poorly resolves which neighbors a protein has in any one cell type. For each of the 207 contexts, the authors selected proteins whose genes are active in that context. ProtScape matches a protein's amino acid sequence against its neighbors in the context-specific network and proposes additional edges for that cell type.

In the preprint, validation works as follows: a fraction of edges were removed from the original maps, and the model recovered them from the remaining neighbors and the protein sequence. Each protein pair was assigned to a single data partition across all cell contexts, either training or validation. This ensured that every test pair remained unseen by the model in every cell context.

For 15 diseases, the authors trained a target protein ranking. Positive examples were proteins targeted by drugs that have completed phase II human trials or have stronger clinical evidence. Ranking depth indicates how many proteins one must inspect to reach all 18 withheld Parkinson's targets: 511 for ProtScape and 8,186 for PINNACLE. The roughly 16-fold difference sets the size of the queue for further experimental validation.

The authors then applied a pre-selected threshold to proteins without training labels and obtained 102 candidates. This protein list is intended for follow-up experiments.