On tower and checkerboard neural network architectures for gene expression inference

BACKGROUND: One possible approach how to economically facilitate gene expression profiling is to use the L1000 platform which measures the expression of ∼1,000 landmark genes and uses a computational method to infer the expression of another ∼10,000 genes. One such method for the gene expression inference is a D–GEX which employs neural networks. RESULTS: We propose two novel D–GEX architectures that significantly improve the quality of the inference by increasing the capacity of a network without any increase in the number of trained parameters. The architectures partition the network into individual towers. Our best proposed architecture — a checkerboard architecture with a skip connection and five towers — together with minor changes in the training protocol improves the average mean absolute error of the inference from 0.134 to 0.128. CONCLUSIONS: Our proposed approach increases the gene expression inference accuracy without increasing the number of weights of the model and thus without increasing the memory footprint of the model that is limiting its usage.