NVIDIA Bets $50M on Recursion as AI Drug Discovery Makes a Comeback

On July 12, NVIDIA announced a $50 million investment in Recursion to accelerate breakthrough foundational models in the field of AI drug discovery. This move has drawn significant attention within the industry and triggered a surge in the stock prices of related companies in the secondary market.

In fact, NVIDIA's pace in AI pharmaceuticals has been somewhat hesitant. As early as 2018, NVIDIA launched the Clara platform specifically for medical scenarios. Subsequently, Clara expanded its boundaries from being a research tool for imaging AI to entering the field of genomics. The Clara platform quickly became an efficient tool in new drug development, capable of being used for drug design by generating molecules through various AI methods to accomplish tasks such as protein generation, molecular generation and docking, and even predicting the three-dimensional interactions between proteins and molecules to optimize how drugs function within the body.

By March 2023, NVIDIA had already collaborated with over 100 companies worldwide on the Clara model, including those in new drug development. However, the $50 million investment in Recursion marked NVIDIA's first direct investment in the global AI pharmaceuticals sector. Founded in 2013, this veteran AI pharmaceutical company primarily uses cellular fiber image characteristics for drug screening, and its underlying logic differs significantly from many of its peers.

Recurison's feature lies in the high-throughput parallel execution of multiple experiments through a closed-loop of dry and wet lab experiments. First, human cells are made sick in various ways in the laboratory, and these diseased cells are photographed. Then, a machine learning program is used to learn the differences between these diseased cells and healthy cells. Finally, various drugs are applied to the diseased cells, and the machine learning program determines whether the cells return to a healthy state, thereby evaluating the effectiveness of the drugs.

In Recurison's AI-driven drug discovery process, foundational research at the cellular level plays a crucial role. Behind this lies a logic that starts from the essence of complex biological phenomena to identify targets and develop drugs. As traditional AI drug discovery models trained on pharmaceutical R&D data are showing signs of fatigue, extending the AI drug discovery pipeline is emerging as a new approach.

In the summer of 2022, less than two years after basking in the spotlight of the capital market, AI drug discovery encountered its first cooling phase. In addition to the widespread external downturn, highly anticipated superstar products entered clinical trials with great fanfare but quickly stumbled, slamming the brakes on the development of AI drug discovery.

In July 2022, Sumitomo Pharma announced the discontinuation of DSP-1181 development as Phase I clinical trials failed to meet expected standards. Subsequently, DSP-1181 disappeared from both Exscientia and Sumitomo Pharma’s official websites. With this, the attempt to develop the world's first AI-designed drug molecule ended in failure.

As early as 2014, Exscientia's technology for automatically generating compounds and its knowledge-based artificial intelligence predictive models were highly favored by Sumitomo Pharma, leading to an immediate collaboration. Sumitomo Pharma became one of the earliest pharmaceutical companies globally to partner with an AI company. In the following years, Sumitomo Pharma and Exscientia worked together intensively, eventually selecting to develop monoamine G protein-coupled receptor (GPCR) drugs for the treatment of mental illnesses.

In the collaboration, Sumitomo Pharma's chemistry team synthesizes compounds proposed by Exscientia, while the pharmacology team evaluates these compounds. Both companies share activity data and work together to further refine the drug. Based on Exscientia’s AI algorithm model, the two parties tested and synthesized up to 350 compounds in less than a year. DSP-1181 was the 350th compound synthesized since the project began. At that time, the industry average for completing this task exceeded five years.

In addition, both parties also synthesized analogs during the project process. Sumitomo Pharma's chemists simultaneously synthesized intermediate compounds proposed by Exscientia, and designed and synthesized some compounds with assumed pharmacological data, which were then input into Exscientia’s predictive models. These included compounds that provided crucial structure-activity relationships for optimizing compound structures, further accelerating the drug discovery cycle and enabling the company to discover DSP-1181 in a short period of time.

At the beginning of 2020, Exscientia proudly announced that DSP-1181, co-developed with Sumitomo Dainippon Pharma, had entered Phase I clinical trials. When DSP-1181 began its clinical trial, Sumitomo Dainippon Pharma was very excited and couldn't help but praise the innovative methods adopted by Exscientia, which they believed would make a significant contribution to the development of drugs for the central nervous system.

Regarding the ultimate failure of DSP-1181, some researchers pointed out that the root cause lies in the lack of innovation in the drug molecule itself.

Todd Wills from the American Chemical Abstracts Service (CAS) conducted a detailed analysis of DSP-1181 and found that the receptor targeted by DSP-1181 is a classic target of significant importance for antipsychotic drugs. In other words, the development of DSP-1181 did not deviate from its original target. After systematically studying the patent for DSP-1181, Wills discovered that the DSP-1181 molecule is very similar to haloperidol, a typical antipsychotic drug approved by the FDA in 1967. In this sense, Exscientia likely optimized a molecular scaffold that had been discovered long ago.

The failure of DSP-1181 cast a shadow over the shining moment of AI-driven drug discovery, but it also brought a critical turning point to the industry. Since then, when people talk about AI in pharmaceuticals, they gradually emphasize pioneering research in laboratories, in addition to algorithms and data.

Having gone through the confusion of the early technology and data accumulation stages, for today's AI pharmaceuticals, building a clinical trial pipeline is not particularly surprising. According to statistics from VCBeat, new drug pipelines developed by China-based AI pharmaceutical companies such as Accutar Biotechnology, Rigel Medicine, Insilico Medicine, and Red Cloud Biotech have successively entered the clinical trial stage. At the end of June, Insilico Medicine was the first in the world to complete the dosing of the first patient in the Phase II clinical trial of the AI-designed drug INS018_055.

The real challenge lies in how to advance clinical trials, as many AI drugs are stuck in Phase I clinical trials. According to statistics from VCBeat, among 80 AI drug pipelines globally approved for clinical trials, only 29 have progressed to Phase II clinical trials, and no AI drug pipeline has entered later stages.

After a decade of sprinting with eyes closed, AI drug discovery is starting to lose steam. In addition to DSP-1181, which failed in Phase I clinical trials, not long ago, Benevolent AI, another leading British AI drug discovery company, announced that a candidate drug for treating atopic dermatitis did not meet the secondary efficacy endpoint in Phase II clinical trials. Meanwhile, Insilico Medicine, known for its bold moves in AI drug development, has been extremely cautious when discussing Phase II clinical trials.

Despite several ups and downs, there is still no clear definition of AI pharmaceuticals within the industry. The attempt to use artificial intelligence (AI) technologies such as machine learning, deep learning, natural language processing, and knowledge graphs for drug chemical molecule analysis, target discovery, compound screening, and even clinical trial research in new drug development is referred to as AI pharmaceuticals.

In many cases, AI pharmaceuticals are regarded as the ultimate solution to improve the efficiency of new drug research and development. However, AI technology, when divorced from strict pharmaceutical logic, achieves breakthroughs in core aspects of new drug R&D in a fragmented manner.

Specifically, in the previous stage of exploration, AI pharmaceuticals were used to complete two extremely tedious yet crucial tasks: discovering new targets and screening compounds.

On one hand, people hope to rely on the powerful computing and analytical capabilities of AI-driven drug discovery to fully explore the potential of difficult-to-drug targets, bypassing the highly competitive red ocean. Statistics show that in the human proteome, difficult-to-drug targets account for more than 75%, and more than half of human diseases currently have no effective treatment available in clinical practice. However, for validated targets such as PD-1, GLP-1, etc., hundreds of pharmaceutical companies often rush into development within a short period.

To date, AI pharmaceuticals have been used to replace many aspects of conventional new drug development. For example, target confirmation, a key step in drug development and one of the most complex processes. Currently, most targets used in new drug research are proteins. In AI-based target discovery, researchers first extract raw features from protein sequences, structures, and functions, then use machine learning methods to build accurate and stable protein models, and finally apply these models to infer, predict, and classify target functions. This has become an important method in AI target research.

In addition to structural data, multi-omics data such as genomics, proteomics, and metabolomics can be extracted from patient samples and vast amounts of biomedical data. Deep learning can be used to analyze the differences between non-disease and disease states, which can also help identify proteins that impact diseases.

On the other hand, AI technology may simplify drug screening and synthesis, reducing costs. For the compounds screened out, it is often necessary to evaluate conditions such as solubility, activity/selectivity, toxicity, metabolism, pharmacokinetics/pharmacodynamics, and synthetic feasibility. This involves a time-consuming and labor-intensive process of repeated experiments, driving up preclinical research costs. Such highly repetitive tasks involving extensive calculations are exactly what computer programs excel at.

In this process, AI technology is used to achieve molecular generation, which allows machine learning methods to generate new small molecules. Specifically, AI can learn the rules of compound molecular structures and drug-like properties by studying a vast number of compounds or drug molecules. Based on these rules, AI can then generate many compounds that have never existed in nature as candidate drug molecules, effectively building a molecular library of a certain scale with high quality.

In addition, AI technology is also used to complete chemical reaction design and compound screening. At present, one of the chemistry fields where AI is making progress is modeling and predicting chemical reactions and synthetic routes. Based on AI technology, molecular structures are mapped into forms that can be processed by machine learning algorithms. Multiple synthetic routes are formed according to the structures of known compounds, and the optimal synthetic route is recommended. Conversely, given the reactants, deep learning and transfer learning can predict the outcomes of chemical reactions. AI technology can even be used to explore new chemical reactions. In compound screening, AI technology is used to model the relationship between the chemical structure and biological activity of compounds, predicting their mechanisms of action.

It can be said that AI pharmaceuticals have performed exceptionally well at every independent node. However, this excellence struggles to extend beyond computer software. Aside from clinical trials that fail to progress, AI pharmaceuticals are widely criticized within pharmaceutical companies, a phenomenon that has become public knowledge. In interviews with VCBeat, complaints from AI pharmaceutical engineers about low molecular activity and long production cycles, as well as critiques from medicinal chemistry experts regarding the difficulty of operating technical platforms, have almost become an unavoidable fate for many AI pharmaceutical enterprises.

Looking back, the gap between AI drug discovery and pharmaceutical companies can be attributed to one key reason: the former prioritizes efficiency, validating its value by compressing development time, while the latter emphasizes quality, requiring repeated verification to select promising candidates before proceeding. In a sense, AI drug discovery follows a straight line, pushing forward with all its might, whereas the process of new drug development is more like a closed loop, allowing for revisions and restarts.

The practical implementation of AI in pharmaceuticals may require halting attempts at single-point breakthroughs and instead integrating into a closed-loop approach for new drug development.

"An increasing number of pharmaceutical companies are building automated laboratories," an investor told VCBeat. "Introducing AI technology in areas such as drug discovery and chemical synthesis has almost become a standard feature for innovative drug companies." Some practitioners have even indicated that if the function of automated smart laboratories in improving new drug R&D efficiency is verified, it will trigger a new wave of infrastructure development among large pharmaceutical enterprises.

VCBeat's analysis of publicly available data reveals that over the past two years, AI pharmaceutical companies have been investing heavily in building automated laboratories. Leading AI drug discovery firms such as Exscientia, Relay Therapeutics, Insitro, BenevolentAI, XtalPi, and Insilico Medicine have successively established dry-wet closed-loop laboratory environments. Meanwhile, multinational pharmaceutical giants like Pfizer, AstraZeneca, and Eli Lilly have also been actively funding automated laboratories for AI-driven drug research and development.

For example, at AstraZeneca iLab in Gothenburg, Sweden, AstraZeneca is exploring the construction of a fully automated pharmaceutical chemistry laboratory, seamlessly integrating the Design, Make, Test, Analyze (DMTA) closed loop for new drug research and development with the technical platform of Molecular AI, an AI-driven drug discovery company. In this process, AI technology primarily handles the design and analysis stages of the DMTA loop, utilizing AI and machine learning to assist chemists in making better decisions more quickly, enabling effective interaction between chemists and computers, thereby accelerating the exploration of chemical space and the design of potential new drug molecules.

For example, Pfizer has partnered with XtalPi to accelerate drug discovery using an "AI prediction + experimental validation" approach. The latter has established an automated laboratory in Shanghai.

"The development of drugs is a multi-dimensional synchronous optimization process," a practitioner told VCBeat. The scale of data in new drug research and development is extremely large, with complex types and structures. Building a closed-loop between dry and wet labs can more efficiently complete the design and validation process."

On the one hand, pharmaceutical companies have developed more systematic data management methods. Traditional drug research and development is primarily experimental science-based. In past new drug development processes, the recording, governance, and storage of data were all centered around experiments, requiring dynamic adjustments based on experimental needs. In other words, data was merely a byproduct of experiments. However, as a method within the realm of virtual science, computational science, and data science, the importance of data in AI is self-evident. This necessitates that pharmaceutical companies strictly standardize the format, standards, quality, and quantity of data in drug research and development.

On the other hand, the algorithm models of AI pharmaceutical companies have also been optimized in a targeted manner, rather than simply being invoked. The deep integration of AI with the core businesses of the traditional pharmaceutical industry emphasizes profound industry understanding and higher technical accuracy. In addition to mining new knowledge from a vast amount of existing papers and experimental data, it is also necessary to have the ability to fully extract and refine real-time experimental data. Based on data feedback, models are optimized, and algorithms are iterated.

"Beyond algorithm models and data, AI pharmaceuticals are increasingly focusing on issues at the biological level," another industry practitioner pointed out. Indeed, relying solely on experiments can only validate formed hypotheses, whereas AI pharmaceuticals face a much more complex system with many unknown questions remaining. In recent years, phenotype-based drug discovery methods have started to gain attention, which directly use biological systems for new drug screening.

How intricate the problems in life sciences are! More fundamental than creating a patented molecule is the logic of understanding biological mechanisms, which can solve the ultimate challenge of AI drug discovery. New changes in the industry may represent a positive shift in the operational model of AI-driven pharmaceuticals—moving away from relatively fragmented independent development based on data from pharmaceutical company labs, clinical data, and ideal biological models, and tracing back upstream to use mathematical approaches to deconstruct disease mechanisms from a biological perspective, ultimately working backward to discover drugs.

"And this process will undoubtedly involve more extensive data analysis and computation, which is also an important reason why companies like NVIDIA, with control over computing power, are deeply involved. 'Low-dimensional models cannot explain high-dimensional problems; only by building tools that understand extremely complex systems can we answer the complicated questions in life sciences,' said Dr. Zhao Yu, Deputy Director of the Turing Darwin Laboratory and Co-founder of Zhe Yuan Technology."

For AI-driven drug discovery, the single-point breakthrough operation model has been, in a sense, proven invalid, but the industry’s growth curve continues to trend upward.