Sequential data selection for predicting the pathogenic effects of sequence variation

Recent improvements in sequencing technologies provide unprecedented opportunities to investigate the role of genetic variation in human disease. In previous work we have proposed a machine learning approach to predicting whether single nucleotide variants (SNVs) are functional or neutral in human disease. Many data sources from the Encyclopaedia of DNA Elements (ENCODE) may be relevant to this problem. To integrate these data sources, we applied integrative multiple kernel learning (MKL) that weights each source according to its relevance. Using an MKL optimization that yields sparse weights, we were able to eliminate the least informative data sources from our model. However, when selecting from a wide assortment of data sources, we have found that MKL may not be an efficient method for eliminating uninformative sources. Many data sources related to the human genome are incomplete: this can reduce dramatically the data available for training and the proportion of novel predictions that exploit all relevant sources. Here we introduce a greedy sequential selection method that assesses data sources in a structured fashion prior to MKL weight optimization. This method allows us to eliminate a majority of uninformative data sources prior to assigning kernel weights. When we use this method with our coding-region predictor, we select just five kernels for our final model, yielding increased accuracy over our previous model. In addition, by reducing the amount of data required for novel predictions, we are able to increase by five fold our model's coverage for new predictions.