In in-silico prediction for molecular binding of human genomes, promising results have been demonstrated by deep neural multi-task learning due to its strength in training tasks with imbalanced data and its ability to avoid over-fitting. Although the interrelation between tasks is known to be important for successful multi-task learning, its adverse effect has been underestimated. In this study, we used molecular interaction data of human targets from ChEMBL to train and test various multi-task and single-task networks and examined the effectiveness of multi-task learning for different compositions of targets. Targets were clustered based on sequence similarity in their binding domains and various target sets from clusters were chosen. By comparing the performance of deep neural architectures for each target set, we found that similarity within a target set is highly important for reliable multi-task learning. For a diverse target set or overall human targets, the performance of multi-task learning was lower than single-task learning, but outperformed single-task for the target set containing similar targets. From this insight, we developed Multiple Partial Multi-Task learning, which is suitable for binding prediction for human drug targets.

The identification of target molecules is important for understanding the mechanism of “target deconvolution” in phenotypic screening and “polypharmacology” of drugs. Because conventional methods of identifying targets require time and cost, in-silico target identification has been considered an alternative solution. One of the well-known in-silico methods of identifying targets involves structure activity relationships (SARs). SARs have advantages such as low computational cost and high feasibility; however, the data dependency in the SAR approach causes imbalance of active data and ambiguity of inactive data throughout targets. We developed a ligand-based virtual screening model comprising 1121 target SAR models built using a random forest algorithm. The performance of each target model was tested by employing the ROC curve and the mean score using an internal five-fold cross validation. Moreover, recall rates for top-k targets were calculated to assess the performance of target ranking. A benchmark model using an optimized sampling method and parameters was examined via external validation set. The result shows recall rates of 67.6% and 73.9% for top-11 (1% of the total targets) and top-33, respectively. We provide a website for users to search the top-k targets for query ligands available publicly at http://rfqsar.kaist.ac.kr . The target models that we built can be used for both predicting the activity of ligands toward each target and ranking candidate targets for a query ligand using a unified scoring scheme. The scores are additionally fitted to the probability so that users can estimate how likely a ligand–target interaction is active. The user interface of our web site is user friendly and intuitive, offering useful information and cross references.