Fig. 1. Overview of the LECIF method.
a Supervised learning procedure of LECIF. For every pair of human and mouse genomic regions, two feature vectors are generated from their functional genomic annotations, one vector for the human region (beige) and the other vector for the mouse region (gray). Each feature vector consists of thousands of functional genomic annotations, as listed in Supplementary Data 1. Only a subset of the features is shown here. These two species-specific feature vectors are given to an ensemble of neural networks (ENN). The ENN is trained to distinguish positive pairs (green), which are aligning human and mouse regions, from negative pairs (red), which are randomly mismatched human and mouse regions that do not align to each other, but somewhere else in the other species. Here, we provide about 2 million positive and 2 million negative training examples. Feature labels (e.g., DNase in liver) and matching of features across species are not provided to LECIF. b Genome-wide prediction procedure of LECIF. Once trained as illustrated in a, the ENN can estimate the probability of any given pair of human and mouse regions being classified as a positive pair. We consider this probability, the LECIF score, to represent the evidence of conservation observed in the functional genomics data annotating the given pair. Here, we generate the LECIF score for all pairs of aligning human and mouse regions. Although not shown here, for model evaluation we also generate predictions for randomly mismatched negative pairs held out from training. When generating a prediction for a pair, LECIF uses an ENN trained on data excluding the pair as described in “Methods” and Supplementary Data 2.