Both methods make use of a machine learning technique, the support vector machine (SVM), to assign each SNP as deleterious or non-deleterious to protein function. The SVM is trained on monogenic disease data, so that the definition of deleterious is 'sufficiently damaging to protein function in vivo as to be consistent with a monogenic disease outcome'. Benchmarking has yielded false positive and false negative rates of 15% and 26% for the stability method and 10% and 20% for the sequence profile method. The higher false negative rate for the stability method reflects the fact that only stability effects on in vivo function are included. Approximately 30% of the non-synonymous SNPs in dbSNP are assigned as deleterious. Very few of the dbSNP cases are known to be associated with monogenic disease, and so most the deleterious ones are candidates for contributing to complex disease traits. As illustrated later, in many cases, low impact on the phenotype is likely the result of network level buffering against loss of function for individual proteins.
