Drug Design with Artificial Neural Networks

Artificial neuron An artificial neuron is a mathematical 18 function that simulates in a simplified form the func19 tions of biological neurons. Usually, an artificial neu20 ron has four computational functions, namely receives 21 signals through input connections from other neurons 22 or from the environment, sums the input signals, ap23 plies a nonlinear functions (transfer function or activa24 tion function) to the sum, and sends the result to other 25 neurons or as output from the neural network. 26 Counterpropagation neural network The counterprop27 agation neural network is a hybrid network that con28 sists of a self-organizing map as the hidden layer and 29 an output layer that has as output a computed value for 30 the modeled property. The network implements a su31 pervised learning algorithm that converges to a unique 32 solution. 33 Multilayer feedforward artificial neural network 34 A multilayer feedforward (MLF) artificial neural net35 work consists of artificial neurons organized in layers. 36 The MLF network has an input layer that receives the 37 structural descriptors for each molecule, an output 38 layer that provides one or more computed properties, 39 and one or more hidden layers situated between the 40 input and the output layers. Each neuron in a hidden 41 layer receives signals from neurons in the preceding 42 layer and sends signals to the neurons in the next layer. 43 Perceptron A perceptron is a linear classifier that consists 44 of a layer of input neurons and an output neuron. Each 45 connection between an input neuron and the output 46 neuron has a weight. Depending on the sum of the sig47 nals received by the output neuron, its output is +1 or 48 1. 49 Quantitative structure-activity relationships 50 Quantitative structure-activity relationships (QSAR) 51 represent regression models that define quantita52 tive correlations between the chemical structure of 53 molecules and their physical properties (boiling point, 54 melting point, aqueous solubility), chemical properties 55 and reactivities (chromatographic retention, reaction 56 rate), or biological activities (cell growth inhibition, 57 enzyme inhibition, lethal dose). The fundamental 58 hypotheses of QSAR is that similar chemicals have 59 similar properties, and small structural changes result 60 in small changes in property values. The general form 61 of a QSAR equation is P(i) D f (SDi ), where P(i) is 62 a physical, chemical, or biological property of com63 pound i; SDi is a vector of structural descriptors of i, 64 and f is a mathematical function such as linear regres65 sion, partial least squares, artificial neural networks, or 66 support vector machines. A QSAR model for a prop67 erty P is based on a dataset of chemical compounds 68 with known values for the property P, and a matrix of 69 structural descriptors computed for all chemicals. The 70 learning (training) of the QSAR model is the process 71 of determining the optimum parameters of the re72 gression function f . After the training phase, a QSAR 73 model may be used to predict the property P for novel 74 compounds that are not present in the learning set of 75 molecules. 76 Radial basis function network The radial basis function 77 (RBF) neural network has three layers, namely an in78 put layer, a hidden layer with a non-linear RBF activa79 tion function and a linear output layer. 80 Self-organizing map A self-organizing map (SOM) is an 81 artificial neural network that uses an unsupervised 82 learning algorithm to project a high dimensional input 83 space into a two dimensional space called a map. The 84 topology of the input space is preserved in SOM, and 85 points that are close to each other in the SOM grid cor86 respond to input vectors that are close to each other in 87 the input space. A SOM consists of neurons arranged 88 usually in a rectangular or hexagonal grid. Each neu89 ron has a position on the map and a weight vector of 90 the same dimension as the input vectors. 91 Structural descriptor A structural descriptor (SD) is 92 a numerical value computed from the chemical struc93 ture of a molecule, which is invariant to the number94