Objective

Vast amounts of biological data accumulated over the decades of biomedical research contain numerous potentially unexplored phenotypes. Experimental groups often lack the necessary resources to address phenotypes that are less obvious and require hefty computational resources to explore. Reanalysis of such datasets may shed light on buried biological complexity and, in certain cases, retrospective analysis of published biomedical work may explain the lack of experimental reproducibility. Recent advances in machine learning and artificial intelligence empowered by the progress in parallel computing allows for rapid exploration of such datasets. Here we propose an implementation of convolutional neural network (CNN) to context folded table dataset. Using several genome wide image-based RNAi screens analyzed using conventional computer vision methods we show that our approach is capable of discovering novel complex phenotypes in the absence of raw pixel-based data. For this we devised a CNN architecture inspired by natural language processing - OmicNet. OmicNet is capable of learning complex data patterns of independent measurements folded in a fashion relevant to biological context. Additionally, by performing pattern recognition on significantly leaner datasets we were able to relax the computational encumbrance of large image-based dataset neural network training. Thus we present a flexible, adaptable, and scalable methodology for re-exploring high-content biological datasets.