In recent years, Artificial Intelligence (AI) has been widely used in a variety of industries, revolutionizing many areas of social life.
In the traditional industry of pharmaceuticals, AI has also been used in a number of processes, including target discovery, virtual screening, compound design and synthesis, prediction of ADME-T properties and physicochemical properties, clinical trial design, management, patient recruitment, pharmacovigilance applications, and real-world evidence generation.
So what is the logic of AI applied to pharma, and how will AI change drug development? How to deal with the efficiency challenges of the pharmaceutical industry? This article is divided into the upper and lower parts, this one focuses on the unfolding of AI in multiple scenarios in the pharmaceutical industry and the challenges it faces.
Pharma in distress
Starting with the plight of the pharmaceutical industry.
Over the past few decades, many scientific, technological, and managerial factors have made great strides that have helped increase the productivity of drug development (R&D). However, since 1950, the number of new drugs approved per $1 billion of R&D investment has nearly halved every nine years, a trend that has been so stable over a 60-year period that it has become known as the pharmaceutical industry's anti-Moore's Law (Eroom's Law). New drugs are becoming increasingly expensive to develop, and drug development is facing a serious productivity crisis.
There are three main explanations for the Anti-Moore's Law, namely, the low-hanging fruit hypothesis (the well-picked fruit has been picked away), the regulatory barrier hypothesis (the increasing regulatory requirements for the filing of new drugs), and the R&D model problem. The first two explanations are objective facts that are difficult to change, so is there a better model for drug development? This is a question that the pharmaceutical industry has been thinking about.
The pharmaceutical industry is facing a data dilemma along with a productivity dilemma.
With the rapid advancement of digital informatization throughout society, the upgrading of drug development equipment, and the accumulation of long-term, available drug development data is increasing, so much so that it is not possible to analyze and process all the data within a certain timeframe using conventional methods and software tools. Traditional statistics are becoming increasingly overwhelmed by the vastness of big data. Pharmaceutical companies are going through a digital transformation and a huge amount of data is being generated. Thus, the contradiction between the growing demand for data processing and the existing data analysis capabilities is driving the pharmaceutical industry to seek a new way out.
The Olive Branch of AI
In March 2016, the AI program AlphaGo's big win over Lee Sedol, the famous Korean chess player, was a milestone event in the history of AI development. This event accelerated the exploration and application of AI in several areas of social life, and gave the pharmaceutical industry hope for increased productivity in drug development.After 2016, there was a large number of technical tests of AI in the pharmaceutical industry. Experimental science is no longer the only option, and data-centered drug discovery is gradually taking the stage.
In the ensuing years, AI pharma has gradually "heated up," with continued proof-of-concept studies, a massive influx of capital into AI-driven biotech startups, and increasing collaboration between pharma companies and AI biotechs and AI technology vendors. Executives at some leading pharma companies see AI not just as a tool for lead compound discovery, but as a more generalized tool for facilitating biological research, discovering new biological targets, and developing new disease models.
AI Unfolds in Multiple Scenarios in Pharmaceuticals
Over the years, AI has been attempted to be applied to almost every process and aspect of drug development, mainly in the following areas:
/ / Target Identification
Target identification is a key step in drug development and one of the most complex. The vast majority of known drug targets are proteins, and it has become an important means of target research to extract features from the raw information of proteins and construct accurate and stable models for function inference, prediction and classification through machine learning methods. Genomics, proteomics, metabolomics, and other multi-omics data are extracted from patient samples and massive biomedical information, and deep learning is used to analyze the differences between non-disease and disease states, and can also be used to discover proteins that have an impact on disease.
/ / Phenotype-based drug discovery
Target-based drug discovery has been the primary approach to drug discovery for over three decades. In recent years, phenotype-based drug discovery (the direct use of biological systems for screening new drugs) has gained attention. Machine learning can link cellular phenotypes to compound modes of action in phenotypic screening to obtain clustering of targets, signaling pathways, or genetic disease associations. The powerful image processing capabilities of AI can integrate all morphological features of biological systems to systematically study potential drug modes of action and signaling pathways, and expand the biological understanding of diseases.
/ / Molecule Generation
Machine learning methods can generate new small molecules, and AI can generate many compounds that have never existed before in nature as drug candidates based on the laws of molecular structure and pharmacokinetics by learning a huge number of compounds or drug molecules, effectively building large-scale, high-quality molecules. This allows us to build a large and high-quality library of molecules.
/ / Chemical Reaction Design
One of the areas of chemistry where AI is making progress is in the modeling and prediction of chemical reactions and synthetic routes, where it can map molecular structures into a form that can be processed by machine-learning algorithms that can form multiple synthetic routes based on the known structure of a compound and recommend the best synthetic route. route. In turn, deep, transfer learning can predict the outcome of a chemical reaction given the reactants, and AI can be used to explore new chemical reactions.
// Compound Screening
AI can model the relationship between a compound's chemical structure and its biological activity, and predict the compound's mechanism of action. A prime example is the discovery of new antibiotics based on deep learning by researchers at MIT. The researchers trained a deep neural network capable of predicting molecules with antimicrobial activity, screened more than 100 million compounds in a few days, ranked the compounds based on the model's prediction scores, and ultimately identified eight antibiotics that were structurally different from known antibiotics.
/ / ADMET Property Prediction
Suboptimal pharmacokinetic properties are one of the main reasons for drug development failures in the clinical research phase. Deep learning can automatically identify relevant features of compounds, evaluate hidden relationships and trends between multiple ADMET parameters in a dataset, and predict properties such as cell permeability and solubility of compounds.
/ / Drug Clinical Trials
The most heavily funded phase of new drug development is the clinical trial phase, and AI has the potential to be used in the design, management, and recruitment of patients for clinical trials. Natural language processing technology can be used to extract information from a variety of structured and unstructured data types to find subjects that meet the enrollment criteria for a clinical trial; it can also be used to correlate large datasets to find potential relationships between variables and improve patient-trial matching. Novartis already uses machine learning algorithms to monitor and manage all of its clinical trials.
/ / Pharmacovigilance
AI will have an impact on traditional pharmacovigilance. As regulatory requirements become stricter and patient safety awareness increases, the workload and cost of pharmacovigilance has increased dramatically. ai can automate the entire process of adverse drug reactions from receipt to reporting, optimizing pharmacovigilance efforts and reducing costs. AI-based systems also have the potential for predictive capabilities for drug risk assessment.
/ / Real-world research
Advances in AI provide new strategies for analyzing large-scale, multidimensional RWD (real-world data).AI can identify intrinsic correlations in real-world data, generate new hypotheses, and also provide new information for clinical trials. In one of the latest cases, by analyzing real-world data, AI can identify enrollment criteria that do not affect the risk ratio of a trial's overall survival, thus expanding the population scope of a clinical trial.
Other applications of AI in drug discovery include physicochemical property prediction, drug redirection, applications in formulation development, and more.
Problems emerge
The application of AI in drug discovery and development is far from smooth sailing, and it boils down to the question of how AI can be adapted to the pharmaceutical scene.
For the pharmaceutical industry, the road of AI, we have to wear AI shoes, AI method for its application of the relevant conditions of the object has many requirements. Just as traditional drug development needs to be equipped with the necessary hardware equipment and the necessary environmental facilities (such as scientific instrumentation, laboratories, etc.), AI-based drug development needs to be equipped with data, algorithms, arithmetic, of which the most stringent requirements for data.
Traditional drug R&D is dominated by experimental science. For decades, the recording, governance, and storage of drug R&D data have been centered on experiments, and adjusted according to the needs of the experiments, with the data being an "adjunct" to the experiments. AI, as a method within the scope of virtual science, computational science and data science, starts directly from the data, puts the data in the first place, and has intrinsic requirements for the format, standard, quality and quantity of the data. In such cases, AI often encounters difficulties in using data directly from traditional drug development models.
For AI to enter the home turf of pharma, it should follow the laws of pharma. For example, drug development is a multi-dimensional simultaneous optimization process, given the scale and complexity of the data, AI-based drug development often requires rewriting machine learning algorithms, rather than simply calling. the deep integration of AI with the core business of the traditional industry of pharmaceuticals, which requires a deeper understanding of the industry and a higher rate of technical accuracy. aI, although it has been possible to mine from a large number of known papers, experimental data for new knowledge, changing the traditional research based on academic experience, however, the accuracy, interpretability and repeatability of the method still need to be improved.
In addition, the traditional drug R&D model has relatively sound regulatory policies and industry systems. As a new model, the application of AI in the pharmaceutical industry to explore, but also need the corresponding industry policy and system to regulate and guide.
Word: Hou Xiaolong
Source: China Food and Drug Administration