Transfer learning for predicting gene regulatory effects of chemicals

Chemogenomics is the study of the genomic and/or proteomic reaction of biological targets to chemical compounds, with the goal of aiding drug design and screening. One of the many difficulties in chemogenomics comes from the dependency on real-life experiment data; the combination of different chemical compounds and drug targets creates an unrealistic number of possible experiments, which results in datasets that are biased towards certain chemicals and targets. The recent developments in machine learning resulted in an over-saturation of powerful models that easily pushed the limits of these datasets. To undo the effects of these biases, we employ transfer learning, the method of leveraging knowledge from similar problems. The most important bias of chemogenomics datasets is the bias towards drug targets. The availability and significance of certain cell lines greatly increase the chance of it being used as a drug target for these experiments, while others have barely enough data to train machine learning models. We experimentally demonstrate that transfer learning greatly increases the trainability of various drug targets, while using the work done on the Deep Compound Profiler (DeepCOP) as a basis. While focused on one of the simple methods of transfer learning, our experiments showed up to 22.81\% and an average of 9.00\% improvement on the area under ROC curve scores and showed great potential to be improved upon if accompanied by hyperparameter optimization and correct cell line as the transfer source.