Uses of Artificial Neural Networks in Macromolecular Structure and Function Predictions

An artificial neural network (NN) is a computer programmed model that attempts to mimic our understanding of the information processing and pattern matching that occurs in the brain. Our biological learning process centers around receiving certain input from our environment and shaping our response (the output) based upon positive and negative feedback supplied during our training. For example, we may be trained to answer the phone at certain times (such as weekends) when the person calling is more likely to be a family member or friend, or we may choose to let the answering machine pick up at another time (such as dinner time on a weeknight) when the caller is often a stranger who wants us to buy something. In such an example, our output (to either answer the phone or let our machine pick it up at some time of day) is trained based upon positive and negative feedback we might receive during our first few months living in New Haven. The same theory is applied to an artificial neural network computer program. Since our brains contain millions of neurons and connections, it is currently infeasible to achieve a similar NN architecture in the computer program. However, as has been discovered, the task of applying NNs to the problem of protein structure prediction requires many fewer artificial neurons and therefore far less total connections. The idea is to create a certain number of input "nodes" and connect each one to every node in a hidden layer. Each node in the hidden layer is then connected to every node in the final output layer. The connection strength between each and every pair of nodes is initially assigned a random value and is then modified by the program itself during the training process. Each node will "decide" to send a signal to the nodes it is connected to based on evaluating its transfer function after all of its inputs and connection weights have been summed (Figure 1). Training proceeds by holding particular data (say from an entry in the Protein Data Bank) constant onto both the input and output nodes and iterating the network in a process that modifies the connection weights until the changes made to them approach zero. When such convergence is reached, the network is ready to receive new experimental data. Now the connection weights are not changed and the values of the hidden and output nodes are calculated according to the functions in Figure 1 to achieve a prediction. Selection of unbiased and normalized training data, however, is probably just as important as the network architecture in the design of a successful NN.