Transfer Learning in Heterogeneous Drug-Target Interaction Predictions Using Federated Boosting

Abstract

In multitask federated learning, when small amounts of data are available, it can be harder to achieve proper predictive performance, especially if the clients’ tasks are different. However, task heterogeneity is common in modern Drug-Target interaction (DTI) prediction problems. As the data available for DTI tasks are sparse, it can be challenging for clients to synchronize the tasks used for training. In our method, we used boosting to enhance transfer in the multitask scenario and adapted it to a federated environment, allowing clients to train models without having to agree on the output dimensions. Boosting uses adaptive weighting of the data to train an ensemble of predictors. Weighting data boosting can induce the selection of important tasks when shaping a model’s latent representation. This way boosting contributes to the weighting of tasks on a client level and enhances transfer, while traditional federated algorithms can be used on a global level. We evaluate our results extensively on the tyrosine kinase assays of the KIBA data set to get a clear picture of connections between boosting federated learning and transfer learning.