Artificial neural networks (ANNs) represent a variety of information technology–based algorithms and systems designed with the intention to mimic biological organization of neurons in animal and human central nervous systems [[1], [2], [3]]. They are essentially groups of connected nodes, or neurons, that may receive, process and transmit signals. One neuron may receive one or multiple input signals which are then weighted and processed using usually a non-linear function to produce an output. Weights are adjusted during the learning process during which the network is trained to produce the most accurate result from the given inputs. The training of the ANN usually refers to providing the network with the large number of examples with the known input and result data. The goal of the training is to adjust the weights in order to minimize the error or the difference between the true result and the predicted result. The trained network is then tested on another batch of data in order to establish the accuracy. Artificial neural networks belong to the family artificial intelligence supervised machine learning (ML) methods [[1], [2], [3]].

There are several types of artificial neural networks and most of them have at least some applications in medical and biological sciences [[4], [5], [6]]. Feedforward ANNs have unidirectional flow of information while feedback ANNs return information back. Single-layer perceptrons (SLP) are simple feedforward ANNs and they are commonly used for linear binary classification of data. More complex multi-layer perceptrons, apart from an input and an output layer, have one or more hidden layers of fully connected neurons, and contrary to SLPs, apply nonlinear activation function. Supervised machine learning using MLPs belongs to the family of deep learning (DL) techniques. Another deep learning ANN is a convolutional neural network (CNN) which is basically a regularized form of MLP with a layer (convolutional layer) that usually applies a rectified linear unit activation function. Unlike MLPs with fully connected neurons, in CNNs (Convolutional deep neural networks) the input data are convolved with individual neurons in convolutional layer receiving data only for a specific receptive field. This reduces the probability of data overfitting which is the major disadvantage of MLPs. In medicine and biology, CNNs are commonly used for analysis and classification of two-dimensional signals [7]. Contemporary alternatives to these classic neural network approaches include the use of network models that use Bayesian inference, or Bayesian neural networks (BNN). In comparison with other machine learning methods, by performing weight marginalization instead of optimization, BNNs in some circumstances may reduce the probability of data overfitting and improve the prediction accuracy [[8], [9], [10]].

In recent years, artificial neural networks have been used on several occasions to classify and predict toxic effects of various biologically active compounds. Application of deep learning in toxicology is an innovative and rapidly evolving approach, and with the further development of artificial intelligence, it can be assumed that DL will in the future become a very important part of various diagnostic and research protocols in toxicology. In this concise review article, we discuss recent findings and research strategies on the use of ANNs for analysis of complex toxicological data. We review the use and potential scientific value of single and multilayer perceptrons, convoluted neural networks and Bayesian neural networks.