Insilico reports that 83% of TargetPro+’s top predictions matched clinical-stage targets

Insilico reports that 83% of TargetPro+’s top predictions matched clinical-stage targets

On September 4, Insilico Medicine introduced TargetPro+, a model that ranks candidate genes as potential drug targets for a disease. The company reported that, on TargetBench 1.0, 83% of its top predictions matched targets that had already reached clinical development, compared with 29–34% for the general-purpose language models named in the announcement.

Drug development begins before the search for a molecule. Researchers first select a target, usually a protein whose activity the treatment is intended to affect. They then look for a molecule that acts on this target and test it experimentally.

TargetBench is one of the tests in DDD Benchmark, which divides drug development into separate tasks. TargetPro+ produces a ranked list of candidate genes for each disease. The test compares the top portion of this list with targets whose drug development programs have already reached the clinical stage. The reported 83% therefore measures how often the system reproduces those historical target selections. According to Insilico’s announcement, the comparison covered 100 diseases across eight groups.

The April methods paper describes its predecessor, TargetPro. It was trained separately for 38 diseases using biological data and scientific texts, with genes associated with drug development programs that had already reached the clinical stage serving as positive examples. In that study, general-purpose language models received the same prompt requesting a list of genes, repeated five times for each disease. The authors attributed the earlier TargetPro’s advantage partly to this specialized training. The September announcement reports a separate result for the new TargetPro+.

A candidate list saved in advance can later be compared with targets that entered clinical development only after the list was recorded. The authors propose this design for future evaluations: it tests the ranking against decisions that had yet to be made when the predictions were recorded.