SparseGO’s architecture

SparseGO was designed based on DrugCell’s two-branch structure, which includes a VNN that captures the hierarchical relationships of the GO graph, and an Artificial Neural Network (ANN) that integrates the Morgan fingerprint⁸ of the compounds (Fig. 1). It is important to note that while the ANN architecture in SparseGO is comparable to DrugCell, the layers of our VNN differ significantly.

SparseGO architecture. The network has two branches: an artificial neural network that takes the Morgan fingerprint of a drug as input, and a sparse VNN that takes the gene expression or mutations of a cell line as input. In the VNN, each GO term is represented by “k” nodes, where the hyperparameter “k” is set to 6 (in all cases). The connections between the layers of the hierarchy are represented using sparse matrices. The output of the network corresponds to the cell line’s viability, measured by the area under the dose-response curve. To assess the significance of the VNN nodes, we employed the DeepLIFT method.

To generate a response for a specific drug, the output of both branches is combined and integrated into another fully connected network. The predicted continuous value represents the area under the dose-response curve (AUDRC) normalized such that AUDRC = 0 represents complete cell death, AUDRC = 1 represents no effect, and AUDRC > 1 represents that the treatment favours cell growth.⁷

SparseGO uses a sparse matrix representation to depict the connections of the GO hierarchy. A matrix is sparse if most of its entries are null. There are different methods to store sparse matrices. All of them (if the proportion of null entries is large) require less memory to store and are more efficient when performing computations (since zeros can be skipped in many operations). This matrix representation was used to create a neural network of sparse linear layers.

Using this architecture, we developed three models. The first model closely resembles DrugCell, utilizing mutations as the input. The other two models incorporate either mutations or gene expression as inputs, with some modifications implemented to enhance the prediction of drug response.