preface
In the past few years, the digitization of data in the pharmaceutical industry has increased greatly. However, the challenge of digitalization is how to apply these data to solve complex clinical problems. This has stimulated the use of AI because it can process large amounts of data through enhanced automation. AI is a technology-based system, including various advanced tools and networks, which can imitate human intelligence. At the same time, it will not threaten to completely replace human existence. AI uses systems and software that can interpret and learn input data to make independent decisions to achieve specific goals. The application of AI in the field of medicine is expanding.
artificial intelligence
AI involves many methodological fields, such as reasoning, knowledge representation, and solution search, including the basic paradigm of machine learning (ML). A subfield of ML is deep learning (DL), which involves artificial neural network (ANN). They include a set of interrelated complex computing elements, involving “perception” similar to human biological neurons, and simulating the transmission of electrical pulses in the human brain. Neural networks involve various types, including multilayer perceptron (MLP) networks, recursive neural networks (RNNs), and convolutional neural networks (CNNs). More complex forms include Kohonen network, RBF network, LVQ network, backpropagation network and ADALINE network. The following figure summarizes the example of AI method domain.
Artificial intelligence helps drug screening
The process of discovering and developing a chemical drug may take more than 10 years, with an average cost of 2.8 billion dollars. Even so, 90% of therapeutic molecules failed to pass the phase II clinical trial and regulatory approval. Nearest neighbor algorithm, RF, limit learning, SVMs and deep neural networks (DNNs) can be used for virtual screening (VS) based on synthetic feasibility, and can also predict the activity and toxicity in vivo. Some large biopharmaceutical companies, such as Bayer, Roche and Pfizer, have cooperated with IT companies to develop artificial intelligence platforms for the discovery of therapeutic methods in the fields of tumor immunology and cardiovascular diseases.
Prediction of physical and chemical properties
The physical and chemical properties of drugs, such as solubility, partition coefficient (logP), ionization degree and internal permeability, will indirectly affect the pharmacokinetic characteristics and targeted receptors of drugs. Therefore, it must be considered when designing new drugs. Different AI tools can be used to predict physical and chemical properties. For example, ML uses the large data set generated in the composite optimization process to train the program. The algorithm of drug design includes molecular description, potential energy measurement, electronic density around the molecule and three-dimensional atomic coordinates. Through DNN, feasible molecules are generated to predict their properties.
Bioactivity prediction
The efficacy of drug molecules depends on their affinity to target proteins or receptors. Drug molecules that have no interaction or affinity with the target protein will not provide a therapeutic response. In some cases, the developed drug molecules may interact with unexpected proteins or receptors, resulting in toxicity. Therefore, drug targeting binding affinity (DTBA) is the key to predict the interaction between drug and target. The method based on artificial intelligence can measure the binding affinity of drugs by considering the characteristics or similarities of drugs and their targets. Based on the interaction of features, the chemical components of drugs and targets are identified to determine the feature vector. On the contrary, the interaction based on similarity considers the similarity between drugs and targets, and assumes that similar drugs will interact with the same target.
Network applications, such as ChemMapper and Similar Integration Method (SEA), can be used to predict the interaction between drugs and targets. Many strategies involving ML and DL have been used to determine DTBA, such as KronRLS, SimBoost, DeepDTA and PADE. ML-based methods, such as KronRLS, evaluate the similarity between drug and protein molecules to determine DTBA. Similarly, SimBoost uses regression trees to predict DTBA, taking into account feature-based and similarity-based interactions.
Toxicity prediction
Predicting the toxicity of drug molecules is essential to avoid toxic effects. Cell based in vitro tests are usually used as preliminary studies, followed by animal studies to determine the toxicity of compounds, increasing the cost of drug discovery. Some network-based tools, such as LimTox, pkCSM, admetSAR and Toxtree, can help reduce costs. Advanced methods based on artificial intelligence find the similarity between compounds or predict the toxicity of compounds according to the input characteristics. The Tox21 Data Challenge, organized by the National Institutes of Health, the Environmental Protection Agency (EPA) and the United States Food and Drug Administration (FDA), is an initiative to evaluate several computational techniques for predicting the toxicity of 12707 environmental compounds and drugs. The ML algorithm named DeepTox stands out. It can effectively predict the toxicity of molecules according to the predefined 2500 toxic group characteristics by recognizing the static and dynamic characteristics in the molecular chemical description, such as molecular weight (MW) and van der Waals force. The different AI tools used in drug discovery are shown in the following table.
AI helps drug design
Prediction of target protein structure
In the process of developing chemical drugs, it is important to predict the structure of target proteins for designing drug molecules. AI can help structure-based drug discovery by predicting 3D protein structure, because the design should conform to the chemical environment of the target protein site, thus helping to predict the impact of compounds on the target site and safety considerations before synthesis or production. AlphaFold, an artificial intelligence tool based on DNNs, analyzed the distance between adjacent amino acids and the corresponding angle of peptide bonds, predicted the three-dimensional structure of target proteins, and correctly predicted 25 of 43 structures.
Prediction of drug-protein interaction
The interaction between drugs and proteins plays a crucial role in the success of treatment. Predicting the interaction between drugs and receptors or proteins is critical to understanding the efficacy and effectiveness of drugs, allowing drug reuse, and preventing multi-pharmacology. Various artificial intelligence methods are very useful in accurately predicting the ligand-protein interaction, ensuring better therapeutic effects.
The ability of AI to predict drug-target interactions has also been used to help change the use of existing drugs and avoid multi-pharmacology. Changing the use of existing drugs can be directly used in the second phase of clinical trials. This also reduced expenditure, because compared with newly developed drug entities ($41.3 million), the cost of restarting existing drugs is about $8.4 million. The “evil association” method can be used to predict the innovative association between drugs and diseases, which is a network driven by knowledge or computing. In computationally driven networks, ML method is widely used. It uses support vector, neural network, logical regression, DL and other technologies.
The drug-protein interaction can also predict the opportunity of multi-pharmacology, which is the trend of interaction between drug molecules and multiple receptors, resulting in non-targeted adverse reactions. AI can design a new molecule based on the basic principle of multi-pharmacology and help to produce safer drug molecules. Artificial intelligence platforms such as SOM, together with existing huge databases, can be used to link several compounds with many targets and non-targets. Bayesian classifier and SEA algorithm can be used to establish the relationship between pharmacological characteristics of drugs and their possible targets.
Ab initio drug design
In the past few years, the method of de novo drug design has been widely used in the design of drug molecules. The traditional ab initio design method is being replaced by the evolutionary DL method, which has the disadvantages of complex synthetic routes and difficult to predict the biological activity of new molecules. Popova et al. developed reinforcement learning of structural evolution strategies for de novo drug synthesis, including generation and prediction of DNN to develop new compounds. Merk et al. used the generative AI model to design retinoic acid X and PPAR agonist molecules at the same time, which has ideal therapeutic effect without complex rules. The author successfully designed five molecules, four of which showed good regulatory activity in cell detection. It is beneficial for the pharmaceutical industry to participate in molecular design from scratch, because it has various advantages, such as providing online learning and simultaneously optimizing the learned data, as well as suggesting possible synthetic routes of compounds, so as to achieve rapid pilot design and development.
AI boosts the development of pharmaceutical products
The discovery of a new drug molecule requires its subsequent combination with a suitable dosage form and desired drug delivery characteristics. In this regard, AI can replace the old trial and error method. With the help of QSPR, various calculation tools can solve problems encountered in the field of formula design, such as stability, solubility, porosity, etc. The decision support tool uses a rule-based system to select the type, nature and quantity of excipients according to the physical and chemical properties of drugs, and operates through the feedback mechanism to monitor the whole process and modify it intermittently.
AI helps pharmaceutical manufacturing
With the increasingly complex manufacturing process and the continuous improvement of efficiency and better product quality requirements, modern manufacturing systems are trying to transfer human knowledge to machines and constantly change manufacturing practices. The application of AI in the manufacturing industry can prove to be a boost to the pharmaceutical industry. Fluid dynamics calculation (CFD) and other tools use the Reynolds average Navier-Stokes solver technology to study the effects of stirring and stress levels in different equipment (such as stirring tanks), so as to automate pharmaceutical operations. Similar systems, such as direct numerical simulation and large eddy simulation, involve advanced methods to solve complex flow problems in pharmaceutical production.
AI helps quality control and quality assurance
The balance of various parameters is included in the production of products required from raw materials. The quality control test of products and the maintenance of consistency between batches require manual intervention. In many cases, this may not be the best method, indicating the necessity of AI implementation at this stage. FDA revised the current Good Manufacturing Practice (cGMP) and introduced a “quality by design” method to understand the key operations and specific standards for controlling the final quality of drugs.
AI can also be used to monitor the online manufacturing process to meet the expected standards of products. The freeze-drying process monitoring based on artificial neural network is adopted, and the adaptive evolutionary algorithm, local search and back-propagation algorithm are combined. This can be used to predict the future time point (t) under specific operating conditions+ Δ t) The temperature and the thickness of the dried filter cake will ultimately help to check the quality of the final product. In addition, data mining and various knowledge discovery technologies in the total quality management expert system can be used as valuable methods to make complex decisions and create new technologies for intelligent quality control.
AI Assists Clinical Trial Design
The purpose of clinical trials is to determine the safety and effectiveness of a drug under specific human disease conditions, which requires 6-7 years and a large amount of financial support. However, only one of the ten small molecules entering the clinical trial may succeed, and the low success rate is a huge loss to the industry. These failures may be caused by improper patient selection, insufficient technical requirements and poor infrastructure. However, with a large amount of digital medical data available, these failures can be reduced through the implementation of artificial intelligence.
The registration of patients requires one third of the clinical trial time. The success of clinical trials can be guaranteed by recruiting suitable patients, otherwise it will lead to about 86% of failed cases. AI can use patient-specific genome-exposure group analysis to help select specific disease populations for recruitment in the second and third phases of clinical trials, which helps to predict the available drug targets of selected patients at an early stage. Pre-clinical discovery of molecules and prediction of lead compounds by using other aspects of artificial intelligence (such as predictive ML and other reasoning techniques) before the start of clinical trials can help to predict the lead molecules passing clinical trials at an early stage and consider the selected patient population.
The patients who withdrew from the clinical trial accounted for 30% of the clinical trial failures, creating additional recruitment requirements for the completion of the trial, resulting in a waste of time and money. This can be avoided by closely monitoring patients and helping them follow the expected protocol of clinical trials. The mobile software developed by AiCure monitors the routine drug intake of schizophrenics in the second phase of the trial, which improves the compliance rate of patients by 25% and ensures the successful completion of the clinical trial.
Market prospect of AI in pharmaceutical industry
In order to reduce the financial costs and failure probability associated with pharmaceutical development, pharmaceutical companies are turning to artificial intelligence. The AI market has increased from US $200 million in 2015 to US $700 million in 2018, and is expected to increase to US $5 billion by 2024. From 2017 to 2024, it is expected to increase by 40%, which indicates that AI may completely change the pharmaceutical and medical industries. Many pharmaceutical companies have invested and are continuing to invest in AI, and cooperate with AI companies to develop necessary healthcare tools. One example is the cooperation between DeepMind Technologies, a subsidiary of Google, and the Royal Free London NHS Foundation Trust to assist acute renal injury. The main pharmaceutical companies and AI participants are shown in the figure below.
Continuous challenges of adopting AI
The whole success of AI depends on the availability of a large amount of data, because these data are used for the follow-up training provided for the system. Accessing data from different database providers may bring additional costs to the company, and the data should also be reliable and high-quality to ensure accurate result prediction. Other challenges that hinder the full adoption of AI in the pharmaceutical industry include the lack of skilled personnel to operate the AI platform, the limited budget of small organizations, the fear of replacing people and causing unemployment, skepticism about the data generated by AI and the black box phenomenon.
Nevertheless, AI has been adopted by many pharmaceutical companies. It is estimated that by 2022, the pharmaceutical industry will generate $2.199 billion in revenue through AI based solutions. Pharmaceutical organizations need to understand the potential of AI technology in solving problems and understand the reasonable goals that can be achieved. Only skilled data scientists, software engineers who have a full understanding of AI technology and a clear understanding of the company’s business objectives and R&D objectives can fully utilize the potential of the AI platform.
expectation
The progress of AI is constantly committed to reducing the challenges faced by pharmaceutical companies, affecting the drug development process and the entire life cycle of products, which is reflected in the increase in the number of start-ups in the industry. The current health care sector is facing some complex challenges, such as the increase in the cost of drugs and treatment, and the society needs to make concrete and major changes in this field. With the application of artificial intelligence in the manufacturing of pharmaceutical products, personalized drugs with required dose, release parameters and other required aspects can be manufactured according to the individual needs of patients. The use of the latest AI based technology can not only accelerate the time required for the product to market, but also improve the product quality and the overall safety of the production process, and make better use of existing resources while improving cost effectiveness.
For the application of these technologies, the most worrying is the consequent unemployment and the strict regulations required to implement artificial intelligence. However, these systems are only designed to simplify the work
