Send Orders for Reprints to
[email protected] Combinatorial Chemistry & High Throughput Screening, 2015, 18, 000-000
1
Artificial Neural Network Methods Applied to Drug Discovery for Neglected Diseases Luciana Scotti*,1, Hamilton Ishiki2, Francisco J.B. Mendonça Júnior3, Marcelo S. da Silva1 and Marcus T. Scotti4 1
Health Sciences Center, Federal University of Paraiba, Campus I, João Pessoa, PB, Brazil
2
University of Western São Paulo (Unoeste), Presidente Prudente, SP, Brazil
3
Biological Science Department, State University of Paraiba, João Pessoa, PB, Brazil
4
Federal University of Paraíba, Campus IV, Rio Tinto-PB, Brazil Abstract: Among the chemometric tools used in rational drug design, we find artificial neural network methods (ANNs), a statistical learning algorithm similar to the human brain, to be quite powerful. Luciana Scotti Some ANN applications use biological and molecular data of the training series that are inserted to ensure the machine learning, and to generate robust and predictive models. In drug discovery, researchers use this methodology, looking to find new chemotherapeutic agents for various diseases. The neglected diseases are a group of tropical parasitic diseases that primarily affect poor countries in Africa, Asia, and South America. Current drugs against these diseases cause side effects, are ineffective during the chronic stages of the disease, and are often not available to the needy population, have relative high toxicity, and face developing resistance. Faced with so many problems, new chemotherapeutic agents to treat these infections are much needed. The present review reports on neural network research, which studies new ligands against Chagas’ disease, sleeping sickness, malaria, tuberculosis, and leishmaniasis; a few of the neglected diseases.
Keywords: artificial neural network, Chagas’ disease, chemometrics tools, drug discovery, leishmaniasis, malaria, sleeping sickness, tuberculosis. 1. INTRODUCTION Neural networks have application in many areas, including pharmaceutical research, engineering, psychology, and medicinal chemistry. Artificial Neural Networks (ANNs) are a group of statistical learning algorithms inspired by biological neural networks, being systems of interconnected "neurons" (or inputs). They are capable of pattern recognition, thanks to their adaptive nature, and are one part of a machine learning approach [1-5]. There are many algorithms in the training neural network models that aid a straightforward application of optimization theory and statistical estimation. These methodologies employ some form of gradient descent, using backpropagation to compute the actual gradients. This is done by simply taking the derivative of the cost function with respect to the network parameters, and then changing those parameters in a gradient-related direction [6-8]. ANNs are powerful tools, frequently used in drug discovery and for monitoring complex interactions between drugs (the substances), and the physiological system [9, 10] The Neglected Diseases affect (almost exclusively) poor and powerless people living in rural parts of low-income countries. The neglected diseases include leishmaniasis
(kalazar), onchocerciasis, Chagas disease, leprosy, tuberculosis, schistosomiasis, lymphatic filariasis, African trypanosomiasis (sleeping sickness), malaria, and dengue. Some neglected diseases are life-threatening, while others result in great morbidity and severe disabilities. The neglected diseases continue to cause significant morbidity and mortality in the developing world. Yet, of the 1,556 new drugs approved between 1975 and 2004, only 21 (1.3%) were specifically developed for tropical diseases and tuberculosis, even though these diseases account for 11.4% of the global disease burden [11-14]. Current therapies present compounds with questionable efficacy and serious toxicity, and they also face greater bacterial resistance. It has become important to review and explore the existing available resources that might aid in new drug development amid novel target identifications for neglected diseases [12, 13]. In new drug studies against such infections, chemometric methodologies, (which are comprised of machine learning methods, and among them ANN), are a powerful tools. They aid in rational drug discovery. The present review reports on new drugs studies against the neglected diseases that have used neural network methods. 1.1. Artificial Neural Networks
*Address correspondence to this author at the Health Sciences Center, Federal University of Paraiba, Campus I, 58051-970, João Pessoa, PB, Brazil; Fax 55-83-3291-1528; E-mail:
[email protected] 1386-2073/15 $58.00+.00
ANNs are a set of methods used extensively since the 1990s. They are computational models with structures © 2015 Bentham Science Publishers
2
Combinatorial Chemistry & High Throughput Screening, 2015, Vol. 18, No. 7
Scotti et al.
derived from a simplified concept of the brain, in which a number of nodes called neurons that are interconnected in a network-like structure (Fig. 1) [15-17]. The neural network has many interconnected nodes, acting as “neurons”, which are connected together according to a topology that can be explained with architectural and functional properties. The architecture is formed of various neurons usually situated on three or more layers; the input layer, the output layer, and one or more hidden layers between them. Artificial neural networks can be divided into two main categories: single-layer and multilayered. As with other chemometrics tools, supervised or unsupervised learning can be used [18]. The neurons on the input layer depend on input variables while on the output layer they depend on the number of functions to approximate. The selection of networks is not a simple task given the number of hidden layers and number of their neurons. The elements in the hidden layer are needed for the best approximation level, and often these are found based on a trial-and-error process. Each neuron receives an input signal vector from the input channels. The functional properties are concerned with information, learning takes place by decomposing the complex into its fundamental elements; which can be characterized as a dynamic system represented by equations [19-21].
Fig. (1). Artificial neural networks.
The initial weights and biases are set to random values for the training algorithm, the input, and target values must be previously prepared. Robustness, a key feature, reflects the ability of a neural network to ignore scattered failures [22]. Tolerance of partial inputs, or inputs with noise, and/or missing information, is another useful feature.
1.2. Neglected Diseases
A neural network presents learning ability and is known to solve problems through mathematical models when a complete formulation is not known. Using the learning algorithm in a neural network, one can derive useful information from the training data samples, and build up knowledge about the weights that connect different neurons together. Based in this information, the network is capable of solving certain complex tasks, including speech recognition, and computer vision [18]. There is off-line learning, and incremental or online learning. The optimization process updates the knowledge base of the neural network with respect to the training data samples in the off-line learning scenarios. Online learning attempts to update the knowledge base of the neural network incrementally after the presentation of each training sample [19]. Networks learn from the input data, therefore the prediction quality is dependent on the data used for training, and the query data point must be covered by the information space of the training data set. If the prediction for a query data point is outside the information space of the training, the prediction will often fail [18]. In drug discovery, artificial neural networks (ANN) are very useful for the prediction of properties of potential drugs. The input received generates output; data could be used from either a patient’s medical information to know what medicine to take, or data to predict structure, and structural parameters to predict biological activity [23].
k-nearest neighbor There are various neural network architectures including the following: feed forward neural networks, self-organizing maps, recurrent neural networks, fuzzy neural networks, back-propagation networks, counter-propagation networks, and other related networks [21, 24-26]. A supervised neural network with supervised learning is showed in the Fig. (2).
Neglected diseases refer to a set of diseases caused by various infectious and parasitic agents (helminths, protozoa, bacteria, and viruses), endemic in "under-developed" countries and low-income populations inhabiting Africa, Asia, and America (South and Central America). This set of diseases has not aroused the interest of big pharmaceutical companies, and receives no financing by governments. The treatment options are non-existent, weak, or outdated. And so, they became the overlooked or neglected diseases [27]. This concept has undergone minor modifications over the years, especially with regard to the exclusion of inappropriate or merely geographical characterizations, which are to some extent discriminatory, since they are diseases with a global range, and need to be addressed within the socio-political, and economic development dimensions of the countries most affected [28]. Currently, the World Health Organization (WHO) together with the Doctors Without Borders Organization define the neglected diseases as: "... a set of diseases associated with poverty, poor living conditions, and health inequities”. Although they account for almost half the burden of disease in developing countries, investments in R & D, especially by the private sector, are not traditionally prioritized this area. In order to demonstrate the degree of disinterest of the pharmaceutical industry regarding neglected diseases, Chirac and Torreele (2006) [29] made a survey of the number of new chemical entities (new active ingredients) sold worldwide between 1975 and 2004. They found for the period that of the 1,556 new chemical entities developed, only 21 (less than 1%), were for neglected
ANNs Applied to Drug Discovery for Neglected Diseases
Combinatorial Chemistry & High Throughput Screening, 2015, Vol. 18, No. 7
3
)$*&%+",&-'.&$/%01'.&/23%4,
!&&-53%20%-60,&-
8&9$%%&:/60,&-
!$;;/?&%'%&10/&-
73-&1,
73-&1,
73-&1,
73-&1,
!"#$%&'()*+ ,*+-*.$+/0 1"0-$%/0(# #%0230*$4/+25
8(9%(#3:(5%5 1"0-$%/0 6(7*#*$ 0*$4/+25
8*-"++*0$
1"==) !%0&!(@
;#
