**: Yakudu
Written by: Gohar
Artificial intelligence (AI) has received widespread attention in drug design and development. Startups trying to apply AI to drug discovery have sprung up, with at least 350 companies currently active in the drug discovery space. These include early-stage companies, startups, as well as more established and IPO-stage companies. Recently, there has been a wave of breakthroughs in this field, with AI systems helping to quickly discover and develop best-in-class small molecules, and these small molecules have all entered clinical research, represented by companies such as Insilico Medicine, Exscientia, Benevolentai, Recursion Pharmaceuticals, Deep Genomics, and others.
It's important to note that most AI-driven startups focus on small molecule drug discovery rather than biologicsThis is not surprising. Historically, even non-AI computational methods (cheminformatics) have been used primarily for small molecules** because their molecular structures and interaction modes are much simpler.
AI has some unique advantages in small molecule drug discovery that make it a powerful tool in this field:
Chemical information availability
The discovery of small molecule drugs usually involves a large amount of chemical information, including molecular structure, drug activity, and toxicity. This information exists in digital form and can be easily processed and analyzed by computers. In contrast, the discovery and design of biologics, such as protein drugs, can involve more complex biological information, which is difficult to represent in numbers, and therefore more complex to process.
Virtual screening at scale
The discovery of small molecule drugs can be accelerated by large-scale virtual screening, which is a powerful advantage of AI. AI can process a vast chemical database of molecules' efficacy, toxicity, and other critical properties, allowing for more efficient screening of potential candidates. Biologics discovery typically involves fewer molecules, and virtual screens are relatively small.
Chemical space exploration
AI is able to effectively explore and understand the chemical space, driving the design and innovation of new compounds. In the field of small molecule drugs, by learning from large amounts of data on known drugs and related compounds, AI can help mine new chemical structures, thereby facilitating the development of new drugs. In contrast, the design of biologics is often more dependent on biological information.
Optimization for high permeability and bioavailability
Small molecule drugs need to have good oral absorption properties, which involves the physicochemical properties of the compound. AI has certain advantages in optimizing the permeability and bioavailability of small molecule drugs, as these properties can be modeled and ** through large amounts of experimental data.
These advantages enable AI to screen and design compounds more quickly and accurately in small molecule drug discovery, accelerating the discovery and development of new drugs.
Corresponding to this,Biologics in the process of drug discovery using artificial intelligence, as opposed to small molecule drugs,There may be the following disadvantages:
Complexity and variety
Biologics are typically large, complex proteins, antibodies, or other large molecules with more complex structures and functions. Dealing with this complexity of data requires more advanced models and algorithms, and current AI technologies may face some challenges when dealing with large-scale diversity of biologics data.
Data availability
The development of biologics involves the complex interaction of databases in different biological systems. Despite the increasing number of biologics data, the data available for large-scale biologics to train AI models is still limited compared to small molecule drugs, limiting the learning and learning capabilities of the models.
Customization and personalization
Biologics are often highly customized and personalized because they are designed for specific diseases or patient needs. This increases the difficulty of design because each biologics may have different structures, functions, and modes of interaction.
Biological diversity
Biologics often interact with the body's biological systems to achieve optimal results, and the biological diversity and complexity of the human body make the development of biologics more challenging. AI models need to better account for this diversity in order to achieve more accurately** the effects of biologics in different individuals.
Artificial intelligence has been widely used in the process of drug molecule discovery, second only to disease modeling and target discovery. AI-driven drug design is divided into three main categories: de novo drug design, virtual screening of existing databases, and drug reuse.
Figure 1The three main categories of AI drug molecule discovery.
De novo drug design
De novo drug design is mainly achieved through deep learning models, such as generative adversarial neural networks (GANs), which are a class of deep learning models that consist of generators and discriminators. The training process of GaN is a game in which the performance of generators and discriminators is continuously improved through adversarial methods). Some examples of generative AI platforms include Chemistry42 software from Insilico Medicine, Makya from Iktos, and De Novo Platform from Ro5. Also included are Recursion Pharmaceuticals, Deep Cure, Standigm, and more.
Simply put, de novo drug design is a computer-aided approach to the design of entirely new drug molecules. Artificial intelligence plays a key role in de novo drug design, which can be done by following these steps:
Data Collection:First, the system requires a large amount of biochemical and pharmacological data, including information on the structure, activity, toxicity, and other information of known drugs. This data is used to train machine learning models.
Feature Extraction:Before training a model, features need to be extracted from the collected data, which may include the structure of the molecule, charge distribution, solubility, and more. The purpose of this step is to convert the chemical information into a digital form that can be understood by a computer.
Machine Learning Model Training:Using a variety of machine learning algorithms, such as deep learning or rule-based approaches, models are trained to understand the relationship between the structure and activity of drug molecules. This allows the model to learn general patterns from known data.
Generation of new molecules:Once the model is trained, it can be used to generate new, unseen molecular structures. This can be done by starting with a random molecular structure and then continuously optimizing through the generative capacity of the model until a drug molecule meets a specific target.
Assessment and Screening:The resulting molecular structure needs to be evaluated to ensure that it has potential medicinal properties. This may involve changes to bioactivity, toxicity, bioavailability, etc.
Optimization and Synthesis Planning:The resulting molecules often require further optimization to improve their synthesizer feasibility and biological activity in experiments. AI can also provide synthesis planning to help determine how these new molecules will be prepared in the lab.
The whole process is an iterative cycle, through continuous optimization of models and experimentation with new molecular designs, with the ultimate goal of finding new drug molecules with good biological activity and clinical potential. This approach can accelerate the drug discovery process, especially when exploring a large number of potential molecular structures, especially when exploring a large number of potential molecular structures.
Virtual screening
The second avenue for applying AI for drug molecule discovery is hyperscale virtual screening, which screens billions of molecules to find successful targets. In August 2022, Sanofi partnered with Atomwise on a drug design deal that could be worth up to $1.2 billion. Sanofi made an upfront $20 million investment focused on using Atomwise's AtomNet platform to study small molecules of up to five drug targets of Sanofi's choice. According to the announcement, Convolutional Neural Network-based AtomNet excels at structure-based drug design and is able to "quickly search for more than 3 trillion synthesizable compounds through AtomWise, AI."
Virtual screening is a method that uses computer models and algorithms to evaluate and evaluate potential drug molecules in order to screen out candidate molecules with potential biological activity from a large compound library. This screening process is done by conducting simulations and ** in a computer, rather than conducting physical experiments in a laboratory, hence the name"Virtual"Sift.
The primary goal of virtual screening is to identify molecules that may be biologically active for specific disease targets in the early stages of drug discovery, from millions to tens of millions of potential drug candidates. This helps to speed up the drug development process and reduce the time and cost of laboratory experiments. In virtual screening, the application of artificial intelligence is mainly achieved by building ** models:
Data preprocessing:Data cleansing, denoising, and normalization. This ensures that the data used for model training is accurate and consistent, improving the model's performance.
Molecular Characterization:Convert molecular structures into computer-processable representations of features. This can be achieved through methods such as molecular descriptors, molecular fingerprints, graph neural networks, etc. Proper molecular characterization is critical to model performance.
Model Selection:Choose the appropriate machine learning or deep learning model. Commonly used models include support vector machines, random forests, deep neural networks, and more. The choice of model should be based on the nature of the task and the characteristics of the data.
Model Training:Selected models were trained using a dataset of compounds with known biological activity. In this way, the model is able to learn the relationship between the structure of the drug molecule and the biological activity.
Model Evaluation:The validation set is used to evaluate the model to test its ability to generalize to unseen data. Metrics may include accuracy, sensitivity, specificity, and more.
Virtual screeningUse a trained model to perform a ** on potential drug molecules. This can be a known molecule obtained from a public database, or a new molecule generated by computation or synthesis. The model gives each molecule a bioactive value, which is sorted according to those values.
Molecular Optimization:For molecules that rank higher in the virtual screen, further chemical optimization may be required. This can be achieved by tweaking the molecular structure to increase biological activity, improve pharmacokinetic properties, etc.
Experimental Validation:The results of the virtual screening need to be validated in the laboratory. Experimental validation helps confirm the accuracy of virtual screens and validates the biological activity and other critical properties of potential drug molecules.
Drug repurposing
Finally, many companies are using drug repurposing strategies for AI drug discovery. Such companies include Healx, Benevolentai, Bioxcel Therapeutics. Using natural language processing (NLP) models and machine learning, they analyze large amounts of unstructured text data (research articles and patents, electronic health records, and other types of data) to construct and search for populations of drugs that can be reused.
Data Integration & Mining:AI can integrate and mine large amounts of biomedical data, including genomics, proteomics, drug interactions, and more. This helps to spot signs that existing drugs may be active in new** areas.
Network Analysis:Using network analysis techniques, AI can establish a network of associations between drugs, proteins, and diseases. By analyzing these networks, potential drug repurposing opportunities can be identified, for example, by discovering pre-existing drugs or compounds associated with the target disease.
Drug Similarity and Feature Learning:AI can analyze the similarity between existing drugs and new ** targets using drug similarity and feature learning methods. This helps to determine if an existing drug is potentially active against a new target.
Machine Learning**:Using machine learning algorithms, AI can build models of the effects of these drugs in new fields based on the biological activity and pharmacological properties of existing drugs. This approach facilitates efficient screening of drug candidates.
Text Mining and Knowledge Graph:AI technology can automatically extract information about drugs from literature, patents, and clinical trial databases through text mining and knowledge graph construction. This helps to discover new uses and associations for the drug.
Cell and Genomics Data Analysis:Using cellular and genomic data, AI can identify the effects of existing drugs on cell or gene expression, thereby discovering possible mechanisms and inferring their potential effects on new targets.
For example, Lantern Pharma, a U.S.-based clinical-stage biotechnology company, is focused on using advanced genomics, machine learning, and artificial intelligence to innovate the cancer drug development process. The Company's AI platform, RADR, currently contains more than 25 billion data points and uses big data analytics and machine learning to quickly discover biogenomic signatures associated with drug response, and then identify relevant subgroups of cancer patients to benefit from Lantern's drug candidates. Lantern and its collaborators also use RADR to develop and position new drugs as well as drug repurposing.
Prospects for the application of AI in small molecule drug research
Based on data collected by Biopharmatrend**, the chart below shows the use of AI by 319 drug discovery startups. Nearly half of the companies (49%, 156 startups) are focused on the discovery of small molecule drugs, while only 20% (64 startups) are involved in the discovery and development of biologics (antibodies, vaccines, etc.).
Figure 2Business distribution map of pharmaceutical AI startups, **biopharmatrend
AI has a bright future in the field of small molecule drug discovery, and although there is currently no achievement in bringing small molecule drugs to market through AI, it is only a matter of time.
Perhaps the biggest advantage of AI is that it dramatically shortens the drug design cycle. The traditional drug discovery process is time-consuming and expensive, and AI is able to speed up the entire process. Through efficient virtual screening, molecule design, and optimization, AI can generate and evaluate a large number of drug candidates in less time. According to the Nature Review Drug Discovery report, the investigators found:Multiple AI projects have completed the entire discovery and preclinical process in less than four years, compared to the five to six years it typically takes.
Based on existing biological and pharmacological knowledge, AI can design molecules with targeted biological activity and better bioavailability. This includes generating new molecular structures, optimizing drug properties, and more. AI can also analyze complex biological networks, the interactions of different drugs, and thus help discover more effective drug combinations. This is especially important for complex diseases and the issue of drug resistance. By analyzing biological and clinical data at scale, AI can help patients personalize their medicines**.
According to the patient's genetic information, biomarkers and disease characteristics, the drug regimen is customized to improve the effect. AI models can help identify potential safety issues at an early stage and reduce the loss of drug candidates by ** potential toxicity and adverse reactions of drugs. Combined with structural biology data, AI can be used to more accurately ** the interaction of small molecules with proteins, thereby guiding the direction of drug design. In terms of the identification of new drug targets, artificial intelligence can provide more options for drug discovery through the identification of new targets in the process of analyzing large-scale biological data.
These prospects suggest that the application of AI in the field of small molecule drug discovery is expected to achieve more breakthroughs in the future, providing more powerful and efficient tools for the discovery and development of new drugs.
References:
1.buvailo, a. will biologics surpass small molecules in the pharmaceutical race? biopharmatrend. 03. 11. 2023.
2.the landscape of artificial intelligence (ai) in pharmaceutical r&d. biopharmatrend. retrieved on 09. 11. 2023.
3.buvailo, a. ai drug discovery: key trends and developments in pharmaceutical industry. biopharmatrend. 26. 07. 2023.
4.jayatunga, m. k. p. et al. ai in small-molecule drug discovery: a coming w**e? nature review drug discovery. 07. 02. 2022.
*Disclaimer: This article is only to introduce the research progress in the field of medical diseases or briefly describe the research overview or share relevant information about medicine, and is not and will not make a recommendation of ** or diagnostic solutions, nor does it constitute any recommendation for related investment. If there is any omission in the content, please feel free to communicate and point it out!