Explore the Use of Artificial Neural Networks in Medicine, Aiming to Understand the Classification, Mechanism, Current Trends, and Applications of Neural Networks in The Pursuit of Smart Health.

The term "neural network" is derived from its conceptual similarity to the human brain. Artificial Intelligence (AI) aims to create advanced computational systems that mimic the human brain's complex neural networks, allowing them to perform complex tasks effortlessly.

AI has significantly transformed the medical and pharmaceutical sectors by utilizing Artificial Neural Networks (ANNs) to assist in various research and development tasks. These networks facilitate innovation in many applications, saving time, reducing costs, and streamlining efforts. However, there are still many untapped opportunities for AI in the fields of science and technology. This review aims to provide insights into the functioning of neural networks in medicine and related fields, as well as highlight potential future research directions.

Neural networks were initially proposed by Alexander Bain and William James separately in 1873 and 1890. In 1943, McCulloch and Pitts developed threshold logic, a computational model based on mathematical algorithms for neural networks, outlining two research approaches: one focused on biological brain processes and the other on AI applications.

In the past decade, artificial intelligence has played an increasingly critical role in medicine. Artificial Neural Networks (ANNs) have revolutionized our world by utilizing complex computational algorithms to find cures for incurable and rare diseases, as well as to optimize and enhance the use of medications and medical procedures. Among the various analytical tools used in AI and machine learning, Artificial Neural Networks (ANNs) have emerged as the most prominent due to their wide range of applications. In addition, during the recent COVID-19 pandemic, AI-based technologies, particularly the Artificial Neural Network (ANN), contributed to potential treatment options.   

Fig. No. 1: Hierarchy of Neural Networks

Introduction to ANNs:

The term "neural network" originates from the brain, and it is the goal of AI to develop intricate computational systems that can rival the human brain's performance. A neural network consists of multiple interconnected nodes or neurons, intended to solve AI-related problems. The connections in artificial neural networks (ANNs) are quantified by weights, with positive weights signifying excitatory connections and negative weights representing inhibitory connections. These weights are utilized in a linear combination of inputs, and the output is then subjected to an activation function to limit its value within a specific range. ANNs find extensive applications in predictive modelling, adaptive control, and various other tasks that involve training on datasets, learning from experience, and deriving insights from complex data.

The concept of the neural network was first conceived by Alexander Bain and William James in the late 19th century. However, it was not until 1943 that a computational model based on mathematical principles was developed by McCulloch and Pitts. This development bifurcated the future research on neural networks into two streams: one focusing on the biological processes of the brain, and the other on applications in AI.[1]

Classification/Types of Neural Network:

Artificial Intelligence (AI), particularly Artificial Neural Networks (ANNs), have garnered significant attention in the medical field in recent years. AI has radically transformed healthcare by uncovering treatments for previously considered incurable diseases and optimizing medication applications using highly complex algorithms. Among the plethora of analysis tools in AI and machine learning, ANNs have been prominently highlighted for their paramount role in treatment strategies during the COVID-19 pandemic.

Neural networks are a foundational technique in machine learning, empowering computers to learn tasks from examples. This network comprises numerous interconnected processing units with each node linked to specific nodes before and after it, enabling seamless information exchange. The network processes information from the training dataset through the input layer, meticulously transforming it into a complex representation as it traverses all layers, culminating at the output layer. With its architecture comprising input, hidden, and output cells, it stands as an incredibly potent tool applicable to a diverse array of purposes.[2]

Fig. No.2: Classification of Neural Networks

AI Algorithms

The following is just a brief overview of some of the algorithms in artificial intelligence with their descriptions:

1. Artificial Neural Networks (ANNs)

Interconnected processing units processing information using a dynamic state response to external inputs.[3]

2. Logistic Regression

 A mechanism to describe the probability of a categorical outcome using a logistic function consisting of one dependent binary variable and one or more independent variables.[4]

3. Hidden Markov Models (HMMs)

 Models the evolution of observable events which may depend upon internal unobservable factors.[5]

4. Clustering Algorithms

Procedures to assign data points to clusters in such a way that items in the same cluster are as similar as possible and items in different clusters are as dissimilar as possible.[6]

5. Fuzzy C-Means

A form of clustering where data points may belong to more than one cluster.

6. K-Means

creates K partitions iteratively, assigning data points to each partition with similarity to a representative centroid. Sparse Autoencoders create deep network structures for feature learning and data clustering. Deep Learning Algorithms are the new version of ANNs that uses plentiful cheap computation for building larger and complex neural networks.[7]

7. Convolutional Neural Networks (CNNs)

 contain an input scanner at the beginning and are typically applied in image processing, but they can deal with other analog inputs such as audio.[8]

8. Recurrent Neural Networks (RNNs)     

Model the temporal dynamic behaviour, conditioning on previous time instants, by interconnecting lower and higher levels.[9]

9. Random Forests (RFs)    

This are collections of many individual decision trees working as an ensemble, and the model's prediction is the class that gets the most votes.[10]

10. Linear Discriminant Analysis (LDA)

Dimensionality reduction of n-dimensional samples is done keeping the class-discriminatory information intact.[11]

11. Support Vector Machines (SVMs)       

Optimum separating hyperplane with the maximum margin, where the nearest point is called.[12]

12. Bayesian Algorithms

These algorithms explicitly apply Bayes' Theorem to problems such as classification and regression.[13]

13. Regression Algorithms

Model the relationship between variables using iteratively-refining a model by minimizing a measure of error in its predictions.[14]

Mechanism of AI algorithms          

1. Data Collection

Gathering relevant data: labelled in the case of supervised, and unlabelled in unsupervised learning.[15]

2. Data Preprocessing

Clean and organize the data for analysis.

3. Model Selection

The selection of an appropriate algorithm depending on the type of problem and data characteristics. Typical models are decision trees, support vector machines, random forests, and artificial neural networks.[16]

4. Training

The model learns patterns or relationships from the training data, and its parameters are changed accordingly.[17]

5. Evaluation

 The model performance is mainly measured using, among other things, accuracy, specificity, and sensitivity. This normally includes cross-validation that guarantees generalization on other data not used in the training stage.[18]

6. Prediction

The application of the trained model to new data to make a prediction or classification.

7. Feedback Loop

It enables the incorporation of feedback from the model's predictions in order to enhance its accuracy over time; this is very important in reinforcement learning scenarios in which a model learns from the outcome of its actions.

Artificial Neural Networks and Their Implementation in the Medicinal and Pharmaceutical Domains:

Artificial neural networks (ANNs) have extensive applications in the medicinal and pharmaceutical industries alongside their widespread use in engineering, computer science, and technology. In drug discovery, ANNs empower the use of various AI and ML architectures for drug screening, design, in-line quality control measurements, and clinical trials. Deep convolutional neural networks are specifically employed in drug development to predict the bioactivity of small compounds, utilizing local convolutional filters to obtain input data on the structural target. It is crucial for sparse ML algorithms, due to their local nature, to exercise caution when dealing with biochemical interactions, whether it involves seeking them out or evaluating them.[19]

Fig. No.3: Application of Artificial Neural Networks

1.3 D Pharmacophores and Neural Networks for Hit-Hopping Beyond the Patent Space of Drug Discovery

This is a very important issue in the area of drug discoveries: the search for hits with new scaffolds, also called scaffold hopping. In this respect, machine learning techniques, especially artificial neural networks (ANN), have been efficient in focusing the research on preferred regions of chemical space. A study was recently published reporting on the approach that uses a three-dimensional pharmacophore descriptor in combination with ANN in the virtual search for NAMs of the metabotropic glutamate receptor 5 (mGluR5) in the treatment of pain, anxiety, and Parkinson's disease. Since there was no available data for the binding of mGluR5 NAMs, a ligand-based virtual screening strategy described the enhancing effect of ANN in virtual screening and scaffold hopping.[20]

Supervised feed-forward networks to generate a mGluR5 focused library; the SOMs for the selection of structurally diverse compounds for bioactivity testing on mGluR5. The supervised networks showed high predictive accuracy in discriminating mGluR5 allosteric antagonists from molecules for other targets and could distinguish mGluR5 from mGluR1 antagonists with lower predictive accuracy. Network ensembles showed higher prediction accuracy than the individual network classifiers.

Commonly used descriptors include,

Hydrophobic descriptors-these descriptors identify hydrophobic regions crucial for binding NAMs to the allosteric sites.[21]

Hydrogen bond donors and acceptors-these are represented as vector indicating possible hydrogen bond formation locations and influence receptors conformation.

Aromatic rings-properties of aromatic rings in 3D space helps to define binding affinity and specificity of potential NAMs.               

2. Neural Network Prediction of Octanol-Water Partition Coefficients

Lipophilicity plays a significant role in the process of drug absorption and its distribution and, thus, is a vital aspect of drug design. Lipophilicity is also described by the partition coefficient in 1-octanol-water, logP, and these values are implemented in use QSAR methods for prediction.[22] Among all, the following ways are in use for predicting logP: hydrophobic fragment values, relationships between the Physico-chemical parameters and logP, and molecular descriptors combined by means of algorithms.[23] Traditional linear methods have often fragmented molecules into larger fragments, while newer methods focus on atomic contributions which then require correction terms due to non-linearity between structure and logP.[21]

A non-linear approach based on neural networks for higher accuracy of logP prediction has been reported to be better than the traditional models. The erroneous structures were filtered out, and the remaining compounds resulted in three sets of 12,729 in total, using a set of organic molecules from the PHYSPROP database. Training for neural networks was conducted using the atomic5 descriptors with 130 input neurons, 12 hidden neurons, and one output neuron. Test sets were used to check generalization and the best model was saved for external validation.[24, 25]

3.Cytotoxicity Prediction Classification by Neural Networks

Cytotoxic compounds can have their discovery costs and resources reduced because such compounds can be eliminated very early in the process. Hence, it is imperative to develop an ANN-based method using atomic fragmental descriptors in the classification of compounds showing in vitro human cytotoxicity.[26] The fragmental descriptors were obtained using Atomic7 linear logP calculation methods. Cytotoxicity data were gathered using in-house screening efforts where 30,000 drug-like molecules were screened for cytotoxic behaviour under pre-defined criteria for the human cell line of interest.[27]

A three-layer feed-forward neural network was set up using the Stuttgart Neural Network Simulator ?(SNNS). There were 164 input neurons,  13 hidden neurons, and 1 output neuron to predict cytotoxicity in human fibroblast cells. For all those compounds that were non-toxic the assigned viability was 0.1. A value of 0.9 was assigned to the toxic compounds. The training was set up based on inputting the Atomic7 descriptors and the output neuron using the scaled experimental viability values.[28]

4. Virtual screening of cytochrome P450 3A4 inhibitors using neural networks

Of the drugs coming under cytochrome P450 3A4, nearly 50% passes through its main isoenzyme 3A4 of cytochrome P450. Its inhibitors lead to drug-drug interactions. Because this enzyme is crucial for so many drugs in the market, the detection of potential 3A4 inhibitors at an early stage is necessary. Several methods of high throughput screening (HTS) have been developed for screening 3A4 inhibitors. Although HTD is more effective in determining potential substrates, it is sometimes incapable of discriminating between substrates and inhibitors. A three-dimensional (3D) QSAR study necessitates good quality IC50 information, which.[29]

To address this issue, a fast neural network model was designed for 3A4 inhibitory activity prediction using 2D structural data. Inhibitory and Gene test of a non-inhibitory were retrieved from Human P450 Metabolism Database tested by the company. Data contained 145 inhibitors and 145 non-inhibitors, which were then split into training and test sets.[22] Under the same method of fingerprinting, 992 numbers of input neurons were created, and the network was structured with 31 hidden neurons and one output neuron. It gave a classification accuracy of 97% and 95% for inhibitors and non-inhibitors, correspondingly. Therefore, it further indicated better prediction regarding those compounds that previously never took any single consideration.[30, 31]

5. Pattern Recognition and Analytical Data Modelling

(Neural networks (NNs) are extremely well suited to pattern recognition in complex or noisy data and hence very effective tools for estimating non-linear connections. No doubt that this is highly important when it comes to research and data analysis, key in identifying the signals appearing as peak-shapes in analytical data such (spectral) areas of study.[32]

For instance, artificial neural networks (ANNs) can differentiate an unknown sample by its complete spectrum and deconvolute overlapping known spectra. Traditional methods consider only one single peak in isolation, whereas the spectrum is used as a whole by ANNs for identification. In contrast, conventional multiple linear regression (MLR) approaches are very labor-intensive iterative procedures requiring a spectrum decomposition and polynomial regression for each peak fitting that can be time-consuming)[33]

In a specific application, ANNs were expertly combined with diffuse reflectance infrared (IR) spectral analysis and X-ray diffraction to swiftly and highly sensitively develop a method for the qualitative and quantitative control of ranitidine hydrochloride, an antihistaminic drug that exists in two polymorphic forms.[34, 35]

6. Modeling the Response Surface

Recent researches have proven clearly the ability of Artificial Neural Networks to model response surface in high performance liquid chromatography (HPLC) optimization. Including retention mapping, i.e. detailed knowledge on how solutes behave chromatographically in relation to the mobile phase components ANNs help a lot in estimating the capacity factors of solutes, which is very advantageous for optimization and it used to be quite an elaborate work requiring large number of experimental separations.[36] Again, several studies have been conducted on this subject and it has uniformly resulted in conclusion that ANNs are able to predict capacity factors more accurately than multilinear stepwise regression models.[37]

7. Relationship between structure and retention

Structure-retention relationship (SRR) methodology was used which is designed to predict chromatography behavior on the basis of structural information about solutes. For example, predicting retention times for the changes in mobile phase pH and composition, as well as solute molecular parameters are related to ANNs.[38] That is, a neural network model accurately related the liquid chromatographic retention of a series of chemically diverse diuretics to their structural descriptors such that new and untested molecules could be assigned predicted retention times. In addition, we introduced a new and effective method to de-convolute overlapping peaks of chromatograms by cross correlating parameters that describe peak shapes with the constituent area thereby achieving remarkably high accuracy and speed over traditional methods.[39]

8. Applications in the Development of Pharmaceutical Products

Development of pharmaceutical products involves solution of complex multivariate optimization problems. The functional relationships among formulation and process variables cannot be easily modelled using conventional techniques. ANNs will find a good fit for such tasks since they can learn and identify patterns in input-output data pairs. After training, ANNs can predict results for new data sets; hence, they are valuable in optimizing pharmaceutical formulations.[40]

In contrast, the Response Surface Methodology, which mainly utilizes second-order polynomial equations, is often deficient in estimation power for predicting the best formulations. ANN models, on the other hand, have been proven to fit and predict better candidates in developing solid dosage forms. They consider many formulation variables and compression factors affecting tablet properties, such as dissolution. In another application of ANNs, it made possible the development of microemulsion-based drug delivery systems with minimal paperchase behaviour using a few inputs alone.[41]

9. QSPR and QSAR Methodologies

The SPR modelling method is based on the basic principle that a compound's molecular structure is very critical to its behaviour. The working of this methodology involves calculating a varied range of physicochemical descriptors, which may include hydrophobicity, topology, electronic properties, steric effects, and others that can be empirically or through modern computational techniques generated.[30]

At the initial periods of QSPR studies, a number of structural descriptors have been calculated with great care in order to precisely describe the chemical structure mathematically. The proper challenge, however, lies in the identification of the best subset of descriptors that would robustly encode the property of interest. The conventional methods of testing of all combinations can be so time-consuming.[24] To attain much better efficiency, genetic algorithms have been very effective optimization systems, using selection and recombination processes to create new sample points with improved fitness. While an effective subset of descriptors is identified, these can then be cleverly mapped to the property of interest using nonlinear computational neural networks.

For instance, it has been shown that back-propagation ANNs could be successfully trained with various descriptors for modelling the structure-activity relationships of capsaicin analogues, where very high correlation was obtained between experimental and predicted results. In one such interesting study, the ANN model correctly classified 34 out of 41 inactive compounds and 58 out of 60 active compounds among 101 capsaicin analogues.)

In conclusion, neural networks are exceedingly versatile and efficient, rendering them an invaluable tool across a multitude of domains, especially in analytical chemistry and pharmaceutical development.

10. Applications in Drug Absorption and Transfer

It has been shown that a more sophisticated neural network model can accurately predict the HIA% from a drug compound's molecular structure. A three-layer feed-forward neural network was effectively used for prediction with six descriptors in an extensive study of 86 drug and drug-like compounds.[42]

In addition, an accurate prediction of the degree of drug transfer into breast milk using a four-layer GNN model will be presented, whose results will be compared with the previous models. In the present study, in-depth research has been carried out on 60 drug compounds and experimentally derived M/P values taken from the literature.[43]

It has also proven its worth in neural network technology, especially in molecular sequence data, ranging from gene identification and protein structure predictions to sequence classification. Such methodologies are very important in understanding the relationship between protein structure and function because unstructured similarities may indicate evolutionary relationships not detectable through sequence analysis.[44]

For example, one back-propagation ANN recognized patterns in protein side-chain contact maps and yielded a remarkable 84.5% accuracy in differentiating the original from the randomized patterns. Even more recently, ANNs trained using genetic algorithms have enjoyed success in the alignment of RNA and DNA sequences and in the determination of RNA folding and secondary structure, firmly establishing the efficacy of this method within biological sequence analysis.[45]

11. Pharmacokinetics and Drug Design

An understanding of pharmacokinetics and pharmacodynamics is indispensable in determining proper drug dosage and selection. ANNs can give a model-independent approach to the analysis of PK-PD data that allows the exact prediction of PD profiles for multiple PK-PD relationships without recourse to detailed structural information. This flexibility gives ANNs a very distinct advantage over traditional model-dependent methods.

Very clearly, structure-based methods have emerged in drug design as superior to traditional methods and save enormous amounts of time and resources. Computational chemistry aspires to quantitative models that predict compound activities based on parameters such as hydrogen bonds, hydrophobic surface area, interaction energies, and desolation.

Two major approaches to finding molecules that fit active sites are:  Using libraries of molecular fragments. In this case, because very large numbers of combinations of fragments can be made, very large numbers of distinct molecules can be produced quickly; the limitations imposed by database size and conformational flexibility are sidestepped.[46]

Conclusion: The quantitative structure-property relationship integrated with genetic algorithms and neural networks has immensely extended molecular modelling, drug design, and biological analysis into areas such as easier and more accurate predictions of chemical behaviours and interactions.[20]

Artificial neural network for Different Physiological Disorders:

Artificial intelligence has made significant strides in the field of medical diagnosis and treatment, revolutionizing the management of a wide range of human ailments including severe cardiovascular conditions, central nervous system disorders, renal disorders, urogenital difficulties, and respiratory illnesses. Predictive modelling using various AI and machine learning tools not only accurately diagnose diseases and recommend treatment options but also forecast potential treatment outcomes, along with the impact of prescribed medication and potential adverse effects once the drug is available to the public.[47] This section aims to debunk exaggerated claims about the functions of different AI algorithms in treating diseases. It endeavours to demystify the mechanisms of various AI tools, such as artificial neural networks (ANNs) and convolutional neural networks (CNNs), to achieve desired outcomes for curing ailments and pathological conditions, thereby vastly improving overall health conditions. The goal is to make intricate AI algorithms more accessible to the general public and ensure transparency in their functioning.[48, 49]

Fig. No. 4: AI in the diagnosis, detection and treatment of human diseases