| A variety of treatment options is now available for patients infected with Human Immunodeficiency Virus (HIV). Often, antiviral treatments do not lead to complete suppression of the virus, due to the rapid development of drug-resistant mutations in the viral genome. For some patterns of mutations, the degree of resistance to some or all of the available antiviral drugs, including protease inhibitors, has been measured in vitro. These measurements can aid in the choice of antiviral treatment options against a viral subtype containing one of these patterns of mutations. However, experimental testing to determine resistance values for all possible variants is combinatorially prohibitive. The primary goal of this thesis is to produce a computational system that learns from a collection of genetic sequences with known drug resistance values, and estimates resistance values for sequences that have not been tested.; The neural network Analog ARTMAP, which learns nonlinear multi-dimensional continuous-valued maps, is introduced and applied to estimate protease inhibitor resistance from viral genotypes. A feature selection method is also introduced and applied to these trained networks, producing insights into the mutation locations that are most predictive of resistance. The nonlinearity of the maps learned by the networks allows the feature selection method to detect genetic positions that contribute to resistance both alone and through interactions with other positions. This method has identified positions 35, 37, 62 and 77, for which traditional linear feature selection methods have not detected a contribution to resistance.; At several positions in the protease gene, mutations confer differing degrees of resistance, depending on the specific amino acid to which the sequence has mutated. To test for these positions, an Amino Acid Space is introduced to represent genes in a vector space that faithfully captures the functional similarity between amino acid pairs.{09}Feature selection identifies several new positions, including 36, 37, and 43, with amino acid-specific contributions to resistance. Accordingly, Analog ARTMAP networks applied to inputs that represent specific amino acids at these positions perform better than networks that use only mutation locations. The coefficients of correlation between predictions and ground truth increase by 1--9%. |
