Artificial Neural Networks and Their use in Chemistry

Abstract

INTRODUCTIONChemists are being confronted with a confounding array of data at an ever increasing pace. The advent of new experimental techniques, the development of cheaper, faster, and more precise instrumentation, and the availability of desktop computing power that only a decade ago would have filled a small house have all contributed to this situation. Medicinal chemists using combinatorial chemistry methods have generated huge libraries of chemical compounds that must be assessed for pharmaceutical activity. Spectroscopists search and analyze huge databases of spectra. Computational chemists generate vast numbers of points describing potential energy surfaces in n-dimensional spaces.The issue facing chemists today is not how to generate data, which not so long ago was actually quite difficult and time-consuming, but how to extract useful information from the data generated. As a consequence, a branch of chemistry known as chemometrics began to evolve in the late 1960s.l Chemometrics is broadly concerned with the extraction of useful information from chemical data. Until about 1990, chemometrics primarily involved appli-

The Challenges Involved in Modeling Toxicity Data In Silico: A Review

“…Learning may be "unsupervised" (no information about the experimental bioactivities is used to update the weights), "supervised" (the difference between the output and the desired output, based on the experimental bioactivity, for the current training set compound, is used to adjust the weights) or, as an alternative/in addition to supervised learning, "reinforcement learning" uses information about how well the network is currently performing (e.g. via an error rate) in order to adjust the weights [188].…”

“…As comprehensively discussed by Peterson [188], Artificial Neural Networks (ANNs) are comprised of a series of interlinked layers of "neurons" which transform weighted input variables (signals) into a new signal which may be passed to subsequent neurons. The original input variables are descriptor values and the final output signal(s) are used to make predictions.…”

“…In "counterpropagation" networks [188,189], the weights in the Kohonen layer are adjusted on the basis of the difference between the descriptor values and the weights for a "winning" neuron, and the weights for the output layer are adjusted on the basis of the experimental bioactivities. Counterpropagation networks tend to be faster to train compared to backpropagation networks [188]; however, care must be taken to optimize the number of neurons in the Kohonen layer [189]. If this is greater than or equal to the size of the training set, considerable overfitting may occur [188].…”

“…In the commonly used "backpropagation" class of networks, the difference between the output and the desired output is not only used to adjust the weights associated with the inputs to the neurons in the output layer, but also to calculate, by computing partial derivatives of the output error with respect to the weights, how the input weights in preceding layers should be updated to minimize this difference [169,188]. In "counterpropagation" networks [188,189], the weights in the Kohonen layer are adjusted on the basis of the difference between the descriptor values and the weights for a "winning" neuron, and the weights for the output layer are adjusted on the basis of the experimental bioactivities. Counterpropagation networks tend to be faster to train compared to backpropagation networks [188]; however, care must be taken to optimize the number of neurons in the Kohonen layer [189].…”

Improvement of Virtual Screening Predictions using Computational Intelligence Methods

“…In next decade, supervised and unsupervised neural models were employed to model QSAR, predict molecules activities and structure, clustering and many more [17][18]. More recently the problem of drug solubility prediction from structure has been revisited [19].…”