Hello @DREAMxCACHETarget2035ChallengeParticipants, despite the large dataset (>6000 labelled molecules) released for the Active Learning step, the objective of the challenge is still to use DEL data to train machine learning models for hit prediction. For this reason, qualifying workflows must include a significant DEL-ML contribution. Here we propose a few examples of methods that would not qualify for the challenge: * Using a model trained exclusively on Active Learning data. Pairing such model with a DEL-ML model can be admissible (see below) * Selecting a small number of close analogs of Active Learning hits (e.g. 500 closest analogs) and then scoring the remaining molecules by DEL-ML. On the other hand, selecting a larger pool of analogs (e.g. 10,000) is acceptable. * Using the DEL-ML predictions to remove a small fraction of the molecules from the test set (e.g. removing the worst 20% of molecules), and then use non-DEL-ML methods to rescore the remaining molecules. * Using a weighted blended score to rank the compounds in the test set. This is unacceptable if the DEL-ML contribution is small. A way to verify the role of the DEL-ML contribution is to perform an ablation study, showing a significant difference in the 50 molecules selected for evaluation when the DEL-ML score is not used. Workflow that we deem to not include a significant DEL-ML contribution will not be considered. If you have doubts on the validity of the method, please reach out to a member of the organizing team, we will be more than happy to assist you. For the organizing team.