The first main objective of the project was the production of RBP Binding Sites datasets that can be used for the training of Machine Learning models. We have used the widely used RBP-24 and RBP-31 datasets, as well as produced a large dataset from ENCODE CLIP-Seq data. In total we have produced datasets for over 100 RBPs, including millions of RBP binding sites. These datasets are disseminated freely and available to the community for training and testing their models.
We proceeded with the development of the deep neural network models of RBP binding. We have not only produced such models that outperform the state of the art for the RBP-24 and RBP-31 datasets, but have also produced and published ENNGENE – a Graphical User Interface equipped method that allows any researcher to easily train such a model on the dataset of their liking. Paired with our ever increasing dataset, and trained model, collection this will become an invaluable resource to RBP researchers.
The second objective of the project was the interpretation of the machine learning models in order to understand what combination of sequence, secondary structure, and evolutionary conservation patterns they learned. We have, for the first time, implemented the Integrated Gradients technique on multi-branch convolutional neural networks, and interpreted the importance of each nucleotide on all three trained modalities. We are finalizing a method that can extract binding motifs from such trained models of RBP binding.
Finally, we had the objective of dissemination of our methods via standalone programs and web-servers. The stand-alone program part was achieved using ENNGENE, not only with the publication of our trained models, but also with empowering researchers to train their own models using our GUI. We are in the process of publicizing our web-server that includes all our collected experimental datasets, as well as binding site predictions for all our trained models.
