AI Drug Discovery
In early 2020, the first AI-designed drug entered human clinical trials. Less than three years later, in February 2023, the FDA granted its first Orphan Drug Designation to Insilico Medicine’s AI-designed drug for idiopathic pulmonary fibrosis (1). With continued market adoption and technological advancement, AI is poised to revolutionize drug discovery. AI in drug discovery is projected to reach $11.9B by the end of 2030, with an annual compounded market growth rate of 30% through this period (2). Reflecting this market potential, pharmaceutical companies like AstraZeneca, Pfizer, Sanofi, and Eli Lilly have all made substantial investments in AI for drug discovery through acquisitions, partnerships, and internal AI development. Notably, AstraZeneca entered a multi-year collaboration with BenevolentAI focused on heart failure and Sanofi entered a $100M partnership with Exscientia for 15 new small-molecule medicines. As a result, there has been a surge of new AI-centric biotech companies entering the sector. These companies hope to increase drug discovery capabilities and accuracy, while simultaneously cutting the cost and length of the process. Many of these startups are emerging in accelerator programs such as Y Combinator and Creative Destruction Lab and disclosed funding and deals in AI drug discovery reached $2.1B in 2022.
The AI technology used in drug discovery is based on machine learning (ML) and generative AI. ML is a subset of AI that focuses on algorithms and models that allow computer systems to learn and make predictions or decisions without explicit programming. Generative AI refers to AI that generates new content from existing data that the model is trained on. AI drug discovery companies often develop complex multi-layer deep neural networks and recurrent neural networks that combine to create algorithms that can analyze large amounts of data, identify patterns, and create new, unique content, while also improving their performance over time.
In the drug discovery setting, supervised AI models can be trained on large datasets such as genomics, biological profiles, chemical properties, and wet lab results to generate their outputs. The data that each company chooses to train their AI model can significantly differentiate their technology. While some inputs, like basic molecule information and structures, are derived from open access sources such as ChemBank, DrugBank, and ChemDB, many companies primarily train their models on propriety data from their own experiments, ensuring internal validity and consistency. With the embedded recurrent frameworks in the AI software, these datasets are constantly being updated and built up over time through repetitions of use, thereby increasing accuracy and performance with each output.
Companies solely training their AI software on public data will have relatively little ability to differentiate themselves in the market. Instead, companies integrating public data with their own proprietary knowledge and data from partnerships will have greater impact with their cutting-edge discoveries. Further some companies, such as PolarisQB, are even exploring the intersection of AI and quantum computing to accelerate the volume of runs their software can execute to increase outputs. This leads to more efficient and accurate models. As both AI and quantum computing technology develop over the next few years, this combination could accelerate drug discovery even further and will be an exciting area to watch.
AI is being applied to every step in the drug discovery and development pipeline, from target discovery to clinical testing and clinical trial design and management. Traditional drug development represents an exorbitant time and capital investment, taking an estimated 12 years from drug discovery to approval, with costs averaging $2.6 billion. Yet, 90% of drugs fail in clinical trials and never make it to market (4). Leveraging AI in drug discovery has been shown to decrease costs by 70%, while also cutting the traditional time requirement in half (5). Additionally, with candidate drug prioritization based on safety, toxicity, and efficacy screening, AI can predict drugs that have the best outcomes and highest likelihood of succeeding in clinical trials. Further, with companies like Israeli-based Quant Health using AI for in silico clinical trial simulations, pharmaceutical companies can generate synthetic evidence of how a therapy will perform across thousands of clinical trial variations to evaluate drug candidate success and viability prior to human tests. Currently, the key areas for innovation with AI are target discovery, drug design, and candidate screening.
Target discovery is the first critical step in the drug discovery process. Therefore, unsurprisingly, this is an emerging sector within AI-powered drug discovery. AI platforms have the power to integrate genomic information from patients with proteomics and clinical data from biobanks to identify disease drivers and potential drug targets rapidly. These platforms can also ensure the target is ‘druggable,’ or able to be targeted by a drug and produce a therapeutic effect once it is bound. The enhanced ability to choose the most efficacious druggable target can increase translation to clinical settings and decrease the likelihood of drug failure and. The field of AI enabled target discovery is also continues to demonstrate great promise, as AI allows for the prediction, analysis, and validation of the chemical makeup and 3D structure of previously unreachable targets. These suddenly reachable targets, such as cell membrane proteins, GPCRs, specific enzyme reactions, and antigens, open the door to discover drugs for previously untreatable diseases and develop new treatment modalities.
Given this transformative potential, the number of companies offering target discovery as-a-service is growing. Companies like NonExomics are exploring the cutting edge of target discovery by investigating the “Dark Genome” to find novel disease-causing targets within non-coding genes and noncanonical proteins. Denmark-based Dianox approaches target discovery from another angle, focusing on novel targets that were previously considered ‘undruggable’ by small molecules and antibodies and instead exploring nucleic acid-protein interactions of these targets. Taking a third approach, CardiaTech analyzes the genomics, proteomics, and metabolomics behind cardiovascular disease to uncover strongly implicated gene variations and targets solely within this indication. Target discovery companies have strong market potential given the vast volume of newly reachable druggable targets and the sheer size of the genome that can now be targeted by AI. This area is only beginning to be discovered and researched, leaving lots of space for startups to differentiate themselves and fill a niche. Propelled by the possibility of partnerships across the pharmaceutical industry, target discovery as-a-service is ripe for growth.
AI promises to transform the ability of companies to design drugs faster and more effectively. Within AI drug design, key areas include de novo design, drug optimization, and drug repurposing. De novo AI algorithms have the power to generate entirely novel drug molecules based on the 2D and 3D structures of specific targets. For example, Decoy Therapeutics utilizes generative AI to design anti-viral drugs based on a given viral sequence in just days. AI platforms can design drug molecules with predefined chemical properties, such as stability, molecular weight, and binding profiles. Furthermore, these systems can predict synthesis routes for the molecules. Within de novo design and optimization of previously designed drugs, some companies use AI to advance drug delivery characteristics. Dandelion Medicine does this by leveraging ML to create novel lipid nanoparticle formulations that function as cell-specific targeting systems, while PersistAI uses ML to design and optimize long-acting microsphere formulations for extended-release drugs.
Lastly, drug repurposing and polypharmacology are areas poised for exciting advancements. Repurposed drugs do not require phase 1 clinical trials, making them attractive compounds to investigate. AI can be used to predict whether known compounds have multiple possible biological targets with therapeutic potential. While de novo drug design and drug optimization are both exciting areas with growth potential, drug repurposing seems especially poised for further development, as relatively few companies have focused here. Through effective analysis of potential secondary uses of drugs, AI has the potential to save research costs, development time, and increase treatments for multiple disorders. This would not only benefit pharmaceutical companies but also the future of patient care.
Lastly, the candidate screening stage, whereby the pharmacodynamics and pharmacokinetics of a novel compound are predicted, continues to be transformed by AI. AI algorithms can rapidly and accurately predict the efficacy of drug-target binding and screen off-target interactions of drugs to ensure candidates with the greatest efficacy and fewest side effects are prioritized. In silico screening also allows for a deeper understanding of a drug’s potential toxicity that would not be discovered until trials, improving the efficiency and safety of drug discovery. For instance, some companies, such as Lavo Life Sciences, leverage AI to simulate the behavior of drugs at a molecule scale, studying molecular crystal polymorphisms to determine bioactivity and safety. Others, like EtCembly and Biolexis, integrate both ML processes and wet lab data into their AI to ensure the highest confidence of candidate binding affinity and lowest probability of toxicity. Yet another emerging advantage of AI candidate screening is the ability to predict interactions that a novel compound could have with other drugs. This proves particularly critical for indications like oncology, where interactions between drugs in a treatment plan must be considered to reduce potential adverse reactions and ensure the best outcome for the patient.
Given that pharmaceutical companies intuitively want to pursue drugs with the greatest efficacy and lowest toxicity, the majority of de novo AI drug design companies also have candidate screening abilities. Therefore, in the AI drug discovery market currently, most companies that are focused on de novo drug design also have candidate screening abilities. Companies that can integrate drug design with enhanced bioactivity and toxicity prediction have the best market application and greatest room for growth in the sector. This combination of capabilities is advantageous as these drugs are less likely to have surprising side effects in animal or human trials, saving both time and money. Very few companies appear to only offer candidate screening as-a-service and, if they do, it is often a software that can be integrated into a lab’s pre-existing system. While a company solely offering candidate screening is theoretically positioned for acquisition by one focused on drug design, the market for a single-offering platform is limited because most companies can already do both.
Differentiation by Treatment Modality
It is also important to evaluate the drug modalities that larger companies and startups are focusing on. Many of the big players in the AI drug discovery space, such as Insilico Medicine and Atomwise, are focused on small molecules that target proteins. However, many of the databases for these small molecules are entirely public, resulting in less intellectual protection and room for differentiation between companies. While these larger companies had the first-mover advantage, the field is now relatively saturated, with fewer areas for transformation. Therefore, many new startups are veering away from small molecules.
Although small molecules still make up most of the drugs on the market, biologics are becoming increasingly attractive to pharmaceutical companies. Alongside higher specificity of treatment--for example, targeted antibodies or RNA-based therapies--biologics face less competition from biosimilars. Additionally, some of the leading companies in the pharmaceutical industry, like Moderna, are focused on RNA, suggesting that the future of therapy lies in nucleic acids. On top of that, tech companies like Microsoft, Meta, and Invidia are also investing in RNA and DNA therapy research, indicating that discovery in this area is ripe for disruption. Interestingly, some startups are now leveraging small molecules that target RNA, thereby taking advantage of this well-researched modality in a novel context. More innovative approaches like these are likely to emerge as the balance between small molecules and biologics continues to shift.
AI is transforming the drug discovery process by identifying novel targets and increasing the accuracy and efficiency of drug designs, while also decreasing cost and time of traditional drug development. AI-enhanced drug discovery has the potential to usher in life-changing treatment for patients across diseases, with initial research even indicating that AI can design new antibiotics. While potential challenges include public skepticism of AI-designed drugs and lack of standardized data inputs, continuous advancements in AI technology by startups and tech companies will likely overcome these barriers. With AI capabilities advancing and pharmaceutical traction building, startups in the AI drug discovery space are poised to transform the status quo of how treatments are developed.