OnCo
ideasIdea

Pre-registered, publicly scored AI ranking of repurposing candidates for cancer

AI systems claim to find new uses for old drugs, but their predictions are rarely tested fairly. Publish their cancer predictions in advance and score them against trial results.

Knowledge-graph and language-model approaches to repurposing (Every Cure, funded by ARPA-H; academic systems such as those built on Hetionet and Open Targets) generate ranked lists of drug-disease pairs, but their forward-looking accuracy in oncology is unknown because predictions are published selectively after the fact. The proposal is a public benchmark: each participating system deposits time-stamped ranked predictions for defined cancer indications; a neutral body scores them annually against subsequent randomised trial results and target trial emulations, and the repurposing fund preferentially trials candidates on which independent systems agree.

Hypothesis
Candidates ranked highly by at least two independent pre-registered systems have a positive-trial rate at least twice that of candidates selected by conventional literature review.
Rationale
Prospective, pre-registered evaluation is the only way to know whether these tools add value beyond the literature they were trained on; if they do, they should direct scarce trial funding, and if they do not, that should be known.
What would test it
Launch the benchmark with three systems and a five-year horizon; compare the fate of top-ranked candidates with a matched set chosen by expert panel.
Maturity
early clinical
Who has to act
data
Cost to try
Small (under $1M)
Years to first evidence
4
Bottlenecks it attacks

Connected

4top