Improving accuracy scores of neural network driven QSAR models of mutagenicity

Multiple QSAR models of mutagenicity were created and compared, using knowledge graph approaches to train and test multi-layer perceptron classifiers, following dimensionality reduction from several thousand dimensions to hundreds of dimensions via principal component analysis. Such knowledge graphs were built in one case using molecular fingerprint based structural similarities, while in another case using molecular fragments found via application of the Girvan-Newman algorithm. A simple hybrid model was also explored. However, both competing QSAR models performed with comparable accuracies, with both sensitivity and specificity scores for each occurring within range of 70%. The predictions of both models were in agreement in an average of 71% of cases, meaning that each could offer a related yet notably different perspective of toxicological space; hence a simple hybrid model was trialed, which only output predictions agreed between both constituent models, which averaged at 78% accuracy.