Drug discovery is underlined by high costs and long lag times due to our lack of knowledge on which ligand could possibly bind well to a disease protein. However, with the advent of machine learning, convolutional neural networks have shown a potential to uncover the underlying patterns on how the structures of ligands and proteins give rise to the interactions between them and hence determine their affinity for each other. DeepDTA is a popular convolutional neural network for such drug-target affinity prediction as its predictions are highly accurate despite its simple model architecture. In this paper, the effectiveness of DeepDTA as an affinity prediction model is investigated using traditional evaluation metrics and some potential updates to the source code are evaluated as well. Some such updates include changing the convolution functions found within the source code as well as changing the optimiser used in the model. Although the latter of the two changes is more promising than the other, further studies are necessitated to better optimise the model and find the suitable hyperparameters to investigate if the model’s performance can be further improved.
License type:
Publisher Copyright
Funding Info:
There was no specific funding for the research done