Single cell transcriptomics is becoming more and more popular [1] because analysis of gene expression per cell basis gives us great insights. However most methods using single cell analysis deal with this complex, noisy, high dimensional and high volume data by aggressive feature selection, filtering and dimension reduction techniques [2]. These operations inevitably result in loss of some critical information within the data. We have been developing techniques help analyse such complex data without much filtering and feature engineering using advanced Deep Learning techniques. We assert that modern Deep Learning systems such as Convolutional Neural Networks can factor in perturbations in data such as inherent noise, batch effects, experiment preparation and execution differences and free-floating RNA be discovered and factored in, on par or better than existing systems eliminating over-simplified filtering and pre-processing that may strip data of valuable insights.
In this talk we will discuss approaching single cell data from a wholistic point and demonstrate results by a technique and a system we call 'DeepMapping', which allows feeding non-image data to Convolutional Neural Networks to analyse data with thousands to millions of parameters in supervised, semi-supervised and self-supervised settings and apply these techniques to single cell RNA expression or multi-omics data with examples.