Tissue-to-plasma Partition Coefficient Prediction by a Multi-channel Restricted Boltzmann Machine
Year: 2014
Type: dissertation
Abstract: The use of a modified restricted Boltzmann machine (RBM) on the multifactorial pharmacokinetic problem posed by the prediction of tissue-to-plasma partition coefficients (Kp) within a living organism is investigated. The RBM’s architecture is altered to allow for an association between subsets (i.e., channels) and the entirety of the visible layer through contrastive divergence. This new architecture is termed a ‘multi-channel RBM’ (McRBM). As a proof of concept, the McRBM is applied to a modified MINST data set. The modified MNIST data set contains handwritten digits obfuscated by pattern overlays to introduce a multifactorial problem. A McRBM is shown to be able to distinguish a digit and an overlay pattern when provided with an obfuscated image. The McRBM is then compared to a published Kp prediction model on a data set consisting of acidic, weak to strong basic, neutral, and zwitterion compounds. Results of McRBM’s predictions are analyzed for accuracy by the Mann-Whitney U test. At the 5% significance level, the statistical analysis suggests that McRBM predictions are more accurate when no separation of the data set, organized by compound type, occurs. Application of this novel machine learning technique will allow for a more accurate prediction of compound affinity for a given organ. Knowledge of compound affinity will reduce the number of in vivo and in vitro experiments required by the traditional pharmaceutical development model. In conclusion, an extension of the McRBM from a shallow to a deep architecture will be described that may allow for improved performance on multifactorial problems.
Author: Christopher JF Cameron
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W2586948180