Nalanda Institute of Pharmaceutical Sciences, Sattenapalli(M), Guntur district, Andra Pradesh-522438.
Artificial intelligence accelerates the drug discovery and development process and reduces the cost, with enormous amounts of successful applications from language modeling to improvement in the pharmaceutical sector. The deep-learning approach has been used throughout the drug discovery steps as the drug-related data increase. In this mini-review, I gave a general description of AI and its application in drug discovery and development. Computer-aided drug discovery and ligand-based quantitative structure activity and property (QSAR/ QSPR) and De Novo drug design, integration with single cell technology, drug metabolism, and excretion, and discuss recent advancement in colorectal cancer and tooth loss, integration of plant-based traditional medicine, and showing AI-assisted platform used to discover serotonin 5-HT1A drug, which is reaching the clinical trial in less than 12 months which is far less than conventional method that needs four years of drug discovery process.
The use of artificial intelligence (AI) in medicinal chemistry has gained significant attention in recent years as a potential means of revolutionizing the pharmaceutical industry. Drug discovery, the process of identifying and developing new medications, is a complex and time-consuming endeavour that traditionally relies on labour-intensive techniques such as trial-and error experimentation and high-throughput screening. However, AI techniques such as machine learning (ML) and natural language processing offer the potential to accelerate and improve this process by enabling more efficient and accurate analysis of large amounts of data. The successful use of deep learning (DL) to predict the efficacy of drug compounds with high accuracy has been recently described. AI-based methods have also been able to predict the toxicity of drug candidates. These and other research efforts have highlighted the capacity of AI to improve the efficiency and effectiveness of drug discovery processes. However, the use of AI in developing new bioactive compounds is not without challenges and limitations. Ethical considerations must be taken into account, and further research is needed to fully understand the advantages and limitations of AI in this area. Despite these challenges, AI is expected to significantly contribute to the development of new medications and therapies in the following few years.1
Commonly used AI techniques
To be a fit-for-purpose approach, the selection and application of AI techniques are problem-oriented. Two common types of AI techniques, namely, supervised and unsupervised learning, are used in the field of drug discovery. A supervised learning technique uses input-labelled data to train models that are capable of classifying or predicting outcomes of new data. By contrast, an unsupervised learning technique deals with unlabeled data and aims to develop models that are capable of identifying recurring patterns and clustering of the input data in a manner without prior knowledge. Supervised learning techniques can be further classified into classification and regression algorithms, and unsupervised learning techniques include clustering and dimensionality reduction algorithms. To facilitate users in applying these AI techniques, a series of opensource package sand frame works, such as Scikit-learn, PyTorch, and Keras, have been developed.
Multiple linear regression (MLR) is a modelling technique that aims to estimate the relationship between independent variables and the dependent variable by fitting a linear equation in to observed data. The ordinary least squares method is used to find the best-fit line by reducing the sum of squared errors, which are the differences between the observed value and the fitted value given by the model. A decision tree (DT) is a non-linear super vised learning technique that can be used in classification and regression tasks. The primary components of a DT model are nodes (including root nodes, internal nodes, and leaf nodes) and branches. The algorithm starts at the root node and selects a branch according to the decision rule of the root node. Subsequently, the algorithm reaches the internal nodes and further makes decisions on the basis of this node. Finally, the algorithm will reach leaf nodes that represent possible outcomes with in the dataset. Logistic regression (LR) is a supervised learning technique that can be used to estimate the probability of occurrence of an event on the basis of logo dds ratio.LR can be classified into three categories, namely, binary, nominal, and ordinal LR, in accordance with the categories of response variables.2
Support vector machine (SVM) is a classical supervised learning technique that is widely used in drug discovery. The basic idea of SVM is to cast data into higher-dimensional features pace by using kernel functions and find the optimal separating hyper plane that maximizes the margin of training data. Convolution neural network (CNN) is a deep learning technique with feed forward neural network architecture. The CNN model includes three types of layers: the convolutional, pooling, and fully connected layers. The convolutional layer aims to learn feature representations of the input. The pooling layer is used to reduce the number of trainable parameters. The fully connected layer aims to produce classification scores and perform reasoning. Compared with conventional machine learning methods, the advantages of CNN include automatically extracting non-hand crafter features from raw input. Recurrent neural network (RNN) Is a feed forward artificial neural network (ANN) that specializes in dealing with sequential data. RNN consists of numerous successive recurrent layers, and its information cycles through a loop. These features make RNN distinct from the traditional neural network. Hence, RNN has the ability to capture contextual content from input data. RNN has also been used in drug design and discovery given its great promise in handling sequential data. Generative adversarial network (GAN) is a deep learning framework with two components: the generator and the discriminator. The former is used to generate new data with the same characteristics as the training data. The latter is used to distinguish actual samples from the generated fake ones. Compared with conventional machine learning methods and other deep learning techniques, GANis good at solving problems with a small sample size.
k-means clustering is one of the most important and popular clustering algorithms. It aims to group similar data into clusters, such that samples in the same cluster are more similar to each other than to those in other clusters. This algorithm iteratively identifies a certain number of centroids (i.e., the arithmetic means of all data points assigned to a particular cluster) within a dataset and allocates every datum to the nearest cluster. These procedures are repeated until cluster assignments stop changing. Hierarchical clustering is another type of clustering algorithm that is used to group data into clusters on the basis of similarity measures. Distinct from k-means clustering, hierarchical clustering initially regards each datum as an individual cluster and then identifies the two closest clusters and merges them together. These procedures are iterated until all the clusters are merged together. The final result is presented in a dendrogram.
Principal-component analysis (PCA) is a linear dimensionality reduction technique that can transform a large dataset into a smaller one while maintaining most of the original information. The basic idea of PCA is to find principal components that explain a large portion of the variation in a dataset. The procedures for conducting PCA include standardizing data, computing the covariance matrix, computing the eigenvalues and eigenvectors, identifying the principal components, and remodeling the data. T-distributed stochastic neighbour embedding (t-SNE) is a nonlinear dimensionality reduction technique that is capable of visualizing high-dimensional data in 2D or 3D space. The t-SNE algorithm first converts similarities between data points into joint probabilities. Then, it minimizes the Kullback-Leibler divergence between the joint probabilities of high-dimensional data and low-dimensional embedding. Application of AI to pharma.3
Application of AI to pharmaceutical analysis
Pharmaceutical analysis involves the processes of identification, determination, quantification, and purification of pharmaceutical raw materials; it is an essential part of drug discovery. Qualitative and quantitative analyses are the two major types of experimental methods in pharmaceutical analysis. Although these techniques exhibit high accuracy, their cost for screening novel drug candidates from a huge amount of natural products is still expensive. Compared with experimental techniques, the costs required by computational methods are negligible. Hence, AI techniques have been used in pharmaceutical analysis to complement experimental techniques.4
Applications of AI in drug discovery
Toxicity is a measure of the unwanted or adverse effects of chemicals. Toxicity evaluation is one of the fundamental steps in drug discovery, and it aims to identify substances that have harmful effects on humans. However, the in vivo test requires animal tests and thus increases the costs of drug discovery. Computational methods exhibit the advantages of being able to predict a chemical’s toxicity with low cost and high efficiency. Accordingly, a series of AI technique-based methods have been developed to predict the toxicity of chemicals. To assess the performance of different computational methods for predicting the toxicity of chemicals, the scientific community proposed the “Toxicology in the 21st Century (Tox21)” challenges. Deep Tox is an ensemble model for predicting the toxicity of chemicals, and its fundamental framework is based on a three-layer deep neural network (DNN). After performing data cleaning and quality control, the remaining chemicals are encoded by using the aforementioned 0D to 3D molecular descriptors, which are used as input of DNN. The Deep Tox pipeline is obtained by tuning and optimizing a set of hyperparameters, such as number of hidden units, learning rate, and dropout rate. Comparative results based on the Tox21 data set demonstrate that Deep Tox outperforms its counterparts in toxicity prediction.
In reality, a large number of drugs derived from natural products are ineffective due to the lack of bioactivity. Hence, drug bioactivity assessment has become an active area in drug discovery. Although in vitro and in vivo experiments can mimic the functions of molecules in the human body, they are still time-consuming and expensive. Given their cost-effectiveness and time economy, AI techniques have been effectively applied to predicting drug bioactivities, such as anticancer, antiviral, and antibacterial activities. For example, Stokes et al. proposed a directed message passing neural network that is capable of predicting antibacterial activity. For each molecule, they first constructed a molecular graph in accordance with its SMILES and then obtained the feature vector based on atomic features (e.g., number of bonds for each atom and atomic number) and bond features (e.g., bond type and stereochemistry). By applying the message passing operation multiple times, the optimized feature vector was fed into the feedforward neural network that outputted the antibacterial probability of a molecule.
Physicochemical properties are intrinsic characteristics of drugs. Knowledge about physicochemical properties is required for understanding and modelling the action of drugs. Among the numerous types of physicochemical properties, solubility is important because it affects the pharmacokinetic properties and formulations of drugs. However laborious and costly experimental techniques have precluded rapid solubility prediction; hence, considerable effort has been devoted to develop AI-based solubility prediction models. Panapitiya et al. assessed different deep learning methods (i.e., fully connected neural networks, RNNs, graph neural networks, and SchNet) and molecular representation approaches (i.e., molecular descriptors, SMILES, molecular graphs, and 3D atomic coordinates) for solubility prediction. Based on the same test dataset, the authors found that the fully connected neural network achieved the best performance for solubility prediction by leveraging molecular descriptors. In addition, the authors analyzed the importance of different features for prediction and found that 2D molecular descriptors made the greatest contributions. To facilitate further research on solubility prediction. To be a fit-for-purpose approach, the selection and application of AI techniques are problem-oriented. Two common types of AI techniques, namely, supervised and unsupervised learning, are used in the field of drug discovery. A supervised learning technique uses input-labelled data to train models that are capable of classifying or predicting out comes of new data. By contrast, an unsupervised learning technique deals with unlabelled data and aims to develop models that are capable of identifying recurring patterns and clues tiring the input data in a manner without prior knowledge. Supervised learning techniques can be further classified into classification and regression algorithms, and unsupervised learning techniques include clustering and dimensionality reduction algorithms.5
Drug discovery is a process of identifying active compounds with therapeutic effects on the intended diseases. Although a high throughput screening technique can scan thousands of different com pounds one at a time, it is still time-consuming and costly. To address these challenges, AI techniques have been applied to nearly all aspects of drug discovery. The applications of AI to natural product-inspired drug discovery, such as de novo drug design, target structure prediction, DTI prediction, and drug-target binding affinity prediction.
De novo drug design refers to the process of generating novel drug like compounds without a starting template. Although conventional structure-based and ligand-based drug design methods have enhanced the discovery of small-molecule drug candidates, they respectively rely on knowledge about the active site of a biological target or the pharmacophores of a known active binder, hindering their applications to modern drug discovery. The boom of AI techniques has offered new opportunities to de novo drug design and accelerated the drug discovery process. In recent years, various deep learning-based models have been proposed for de novo drug design, such as the reinforcement learning based model Release, the encoder-decoder-based model ChemVAE the GAN-based model Graph INVENT, and the RNN-based model MolRNN Another key point of de novo drug design is molecular representation. SMILES, fingerprint, molecular graph, and 3D geometry have been used as input of deep learning algorithms. The fundamental framework of deep learning-based de novo drug design methods is shown in the left upper corner of Detailed information about deep learning-based de novo drug design models is provided recent reviews.6
Most drug targets are proteins that play important roles in enzymatic activities, cell signaling, and cell-cell transduction. The functions of proteins are determined by their structures. Although conventional experimental techniques, such as X-ray crystallography, cryogenic electron microscopy, and nuclear magnetic resonance spectroscopy, have been proposed to determine protein structures, they are still time-consuming and costly as reported, experimental techniques have only deciphered the structures of 100,000 unique proteins, which account for only a small part of known proteins. Therefore, developing novel methods to fill the gap between the number of protein sequences and known protein structures is an urgent need. With the rapid growth of computational power and the break throughs of AI techniques, many computational approaches have been proposed for protein structure prediction. The neural network-based AlphaFold method developed by DeepMind is the best-performing method, and it is able to predict the 3D structures of proteins from their amino acid sequences and achieve accuracies competitive with experiments.
AI Techniques for natural predict-inspired drug discovery
DTI prediction refers to the interaction between chemical com pounds and protein targets in living organisms. DTI prediction is an essential process for drug discovery. Hence, experimental methods have been used to determine DTI, such as co-immuno precipitation phage display technology, and yeast two-hybrid. However, these wet laboratory techniques are time-consuming when they are used to predict DTI. Recently, the ever-increasing biological data have paved the way for the in-silico prediction of DTI. Therefore, computational methods are being increasingly used in DTI prediction. These methods, which were summarized in a recent review, can be classified into the following categories: ligand-based methods, docking simulations, gene ontology-based methods, text mining-based methods, and network-based methods. Compared with other types of methods, deep learning-based methods frequently exhibit better performance in DTI prediction. First, compounds and proteins are encoded by using their corresponding features. Then, the feature embedding of the compounds and proteins issued as the input of deep learning methods. In accordance with this strategy, models based on deep belief neural network, CNN, and multiple layer perceptron have been proposed for drug-protein interaction prediction, considerably facilitating drug discovery. In real life, many diseases lack well-defined targets. Hence, finding drugs for these diseases is impossible by using the aforementioned methods. Zhu et al. recently proposed a deep learning-based efficacy prediction system (DLEPS) that can identify drug candidates in accordance with the changes in gene expression profiles rather than specific targets. First, compounds were encoded using SMILES and used as input of CNN to fit gene expression changes. Subsequently, the potential efficacy of compounds against diseases was evaluated on the basis of gene signatures specific to certain diseases and sorted using a method similar to gene set enrichment analysis. DLEPS provides novel insights into identifying new drugs for complex disease.7
In most cases, DTI prediction is regarded as a binary classification problem, but binding affinity between a drug and its target is disregarded. Binding affinity reflects the strength of drug-target pair interactions, and it is considerably informative for drug discovery. Although binding affinity can be experimentally determined by measuring dissociation and inhibition constants, the time cost and financial expenses of these procedures are extremely high. Therefore, developing computational methods for predicting binding affinity is necessary. In 2018, Öztürk et al. proposed the first deep learning model, called Deep DTA, for predicting binding affinity between drugs and their targets. In Deep DTA, the drug and the target were encoded using SMILES and amino acid letters, respectively, which were then used as input for CNN. The comparative results demon started that Deep DTA suppressed KronRLS and SimBoost for drug-target binding affinity prediction. Inspired by Deep DTA, a series of deep learning-based models has been sequentially proposed, such as Wide DTA and Deep Affinity, which have become useful tools in drug discovery.
Synergism and antagonism are the two categories of drug combination effects. The former can overcome primary and secondary drug resistance, and it is effective for the treatment of cancers, AIDS, and bacterial infections, whereas the latter reduces the effectiveness of drugs. With the ever-increasing number of drugs, their possible combinations are astronomical. Thus, experimentally investigating drug combination effect is costly and time-consuming. The advancements of AI techniques have made them applicable to exploring possible drug combinations at lower cost and with more efficiency. In 2015, Li et al. proposed a Bayesian network model for exploring and analysing drug combinations. In the same year, Wildenhain et al. developed a random forest-based model for predicting compound synergism from chemical-genetic interactions. Recently, Preuer et al. proposed Deep Synergy, a deep learning-based model for predicting the synergism of anticancer drugs. The inputs of Deep Synergy included the chemical information of drugs and the genomic in formation of diseases, which were then propagated through the network to the output unit. The comparative results from a publicly available synergy dataset demonstrated that Deep Synergy out per formed its counterparts in predicting drug synergism.8
Nanotechnology has been applied to design nanomedicines by using nanometric-scale materials in the clinical setting. Nanomedicines are developed by materials at the nanometric scale, and, thus, they can penetrate the barriers to interact with targets in the body. At pre sent, some nanomedicines have already been approved by the U.S. Food and Drug Administration, and they have exhibited better performance in the treatment of cancers and HIV-1 infection. However, the lack of quantitative and qualitative understanding of nanomaterial properties and biological responses precluded the wide application of nanomedicines.
A combination of nanotechnology and AI provides novel solutions to deal with this dilemma. For example, Li et al. proposed an ANN for the task of nanomedicine composition optimization. Muñiz Castro et al. developed a 3D printing nanomaterial formulation pipeline that can predict the extrusion temperature, filament mechanical characteristics, and dissolution time of nanomaterials. In addition, the effectiveness of a nanomedicine is affected by its cellular uptake. Hence, a cellular uptake prediction model will considerably help researchers in predicting nanomedicine effectiveness.
Case studies of successful AI-aided drug discovery efforts
The potential of AI in drug discovery has been demonstrated in several case studies. For example, the successful use of AI to identify novel compounds for the treatment of cancer has been recently reported by Gupta, R et al.9 These authors trained a DL algorithm on a large dataset of known cancer-related compounds and their corresponding biological activity. As an output, novel compounds with high potential for cancer treatment were obtained, demonstrating the ability of this method to discover new therapeutic candidates. The use of ML to identify small molecule inhibitors of the protein MEK has been recently described. MEK is also a target for the treatment of cancer, but the development of effective inhibitors has been challenging. The ML algorithm was able to identify novel inhibitors for this protein. Another example is the identification of novel inhibitors of beta-secretase (BACE1), a protein involved in the development of Alzheimer's disease by using a ML algorithm. AI has also been successfully applied in the discovery of new antibiotics. A pioneering ML approach has identified powerful types of antibiotics from a pool of more than 100 million molecules, including one that works against a wide range of bacteria, such as tuberculosis and untreatable strains. The use of AI in the discovery of drugs to combat COVID-19 has been a promising area of research during the last two years. ML algorithms have been used to analyze large datasets of potential compounds and identify those with the most potential for treating the virus. In some cases, these AI-powered approaches have been able to identify promising drug candidates in a fraction of the time it would take traditional methods.10
Ethical considerations in the use of AI in the pharmaceutical industry
As discussed in the previous section, it is important to consider the ethical implications of using AI in this field. One key issue is the potential for AI to be used to make decisions that affect people's health and well-being, such as decisions about which drugs to develop, which clinical trials to conduct, and how to market and distribute drugs. Another key concern is the potential for bias in AI algorithms, which could result in unequal access to medical treatment and unfair treatment of certain groups of people. This could undermine the principles of equality and justice. The use of AI in the pharmaceutical industry also raises concerns about job loss due to automation. It is important to consider the potential impact on workers and provide support for those who may be affected. Additionally, the use of AI in the pharmaceutical industry raises questions about data privacy and security. As AI systems rely on large amounts of data to function, there is a risk that sensitive personal information could be accessed or misused. This could have serious consequences for individuals, as well as for the reputation of the companies involved. The collection and use of sensitive medical data must be done in a way that respects individuals' privacy and complies with relevant regulations. Overall, the ethical use of AI in the pharmaceutical industry requires careful consideration and thoughtful approaches to addressing these concerns. This can include measures such as ensuring that AI systems are trained on diverse and representative data, regularly reviewing and auditing AI systems for bias, and implementing strong data privacy and security protocols. By addressing these issues, the pharmaceutical industry can use AI in a responsible and ethical manner.11
Limitation of AI in Drug Discovery
Although the efficacy of AI based methods in drug discovery are significant but their applications are limited in both capability and functionality. One major criticism of many AI techniques such as neural networks is that they are often regarded as black boxes that merely attempt to map a relationship between output and input variables based on a training data set. This also immediately raises some concerns about the ability of the tool to generalize to situations that were not well characterized in the data set. One of the limitations of the genetic algorithm methods is that they are never guaranteed to reach the "optimal" solution, though the solutions provided are highly useful. In ML technique, we cannot ensure that what the model learned in terms of derivitization or in terms of heuristic reasoning, the ML model itself learns a few factors from the data provided to it. It is difficult to ensure which factor of the supplied data was utilized to train which component of an ML model. A well-known drawback of deep learning is its poor performance where data size is low-to-medium.12
CONCLUSION
AI in drug discovery is not a new phenomenon. Machine learning has been an integral part of generating small molecule targets for decades. Recent and ongoing improvements in AI have allowed it to enter other parts of the drug discovery process, to cut costs and improve efficiency. Besides molecular docking and toxicity prediction, which were already staples of state-of-the-art drug discovery workflows, small biotech companies are piloting many new ways of using AI. This is accelerating the shift in the business model of big pharma. Rather than doing all the research in-house, these firms buy trial-ready compounds from external parties. While some of the needed steps are still unsolved, successful adoption of AI in the entire drug discovery pipeline could dramatically decrease drug development costs. This would enable the industry to make drugs for patient populations previously considered far too small to justify the expense.
ACKNOWLEDGEMENT
I am privileged to express my sincere gratitude to our principal sir Dr. D. Rama Brahma Reddy for providing us the opportunity to write this review under the guidance and support of K. Malleswari madam and also my institute Nalanda institute of pharmaceutical sciences. Deori for kind support and encouragement.
REFERENCE
Rama Brahma Reddy D.*, Malleswari K., Chetan M., Adarsh Babu B., Bhuvan Chandra Durga Eswar J., A Review On Role Of Artificial Intelligence In Drug Discovery, Int. J. of Pharm. Sci., 2024, Vol 2, Issue 8, 2913-2922. https://doi.org/10.5281/zenodo.13293031
10.5281/zenodo.13293031
