Efficient online-group-screening designs for agent identification

Identifying significant causal agents among a large number of candidates is challenging. When experimental resources are limited, exhaustively screening a large number of agents for the desired effect could incur a large cost and take a substantial amount of time. However, in many large scale experiments, such as high-throughput screening (HTS), the ratio of causal to non-causal agents is usually very low. In this paper, we introduce a group-screening strategy to efficiently screen causal agents by grouping them into treatments. Our analysis shows that when a large number of candidates factors are screened and true agent percentage is very low (less than 1%), even in the worst case we could save up to 80% of the experiment runs. In the case where experiments span many rounds, we provide an online version of the group-screening that can determine the best strategy automatically based on the existing results. We applied this method to a real HTS experiment with 50,000 candidates that would require 9 rounds to finish in an exhaustive case. Our analysis showed that by applying the online-group-screening method, in the worst case, we can use 3 rounds and 19.7% (9828/50000) total tests to identify all the agents. Finally, we show that with minor modifications, this framework extends to more complex agent discovery problems.