Original source: arterial network
Image source: Generated by Unbounded AI
On July 12, Nvidia announced a $50 million investment in Recursion to accelerate breakthrough fundamental models in the field of artificial intelligence drug discovery. This move has aroused widespread concern in the industry, and the stock price of related targets in the secondary market has skyrocketed.
In fact, Nvidia is a little hesitant to deploy AI pharmaceuticals. As early as 2018, Nvidia launched the Clara platform specifically for medical scenarios. Subsequently, Clara gradually expanded its boundaries from imaging AI research tools and began to get involved in genomics. The Clara platform has quickly become an efficient tool in the development of new drugs. It can be used in drug design, through different AI to generate molecules, to complete tasks such as protein generation, molecular generation and docking, and even predict the three-dimensional interaction between proteins and molecules, so as to optimize How the drug works in the body.
By March 2023, NVIDIA has cooperated with more than 100 companies around the world, including new drug research and development, on the Clara model. But the $50 million invested in Recursion is Nvidia’s first direct investment in global AI pharmaceuticals. Founded in 2013, this established AI pharmaceutical company mainly uses the fiber image features of cells for drug screening. The underlying logic is quite different from that of other peers.
The feature of Recurison is that multiple experiments can be parallelized with high throughput through closed-loop dry and wet experiments. First, human cells are made sick in various ways in the laboratory, and these sick cells are photographed. Then, let the machine learning program learn the difference between these diseased cells and healthy cells. Finally, various drugs are applied to the diseased cells, and the machine learning program is used to judge whether the cells return to a healthy state, so as to judge the effect of the drugs.
In Recurison’s AI pharmaceutical process, basic research at the cell level is a key link. Behind this is a logic of finding targets and developing drugs based on the essence of complex life phenomena. At the moment when the traditional AI pharmaceutical model trained with drug research and development data is a little tired, extending the chain of AI pharmaceuticals is becoming a new way of thinking.
Disappeared DSP-1181, and new AI drug that can’t run
The summer of 2022 has just arrived, and after less than two years of rushing under the spotlight of the capital market, AI pharmaceuticals ushered in the first cooling down. In addition to the wide-ranging cold external environment, high-profile superstar products entered the clinical trial stage, but quickly encountered Waterloo, stepping on the brakes on the development of AI pharmaceuticals.
In July 2022, Sumitomo Pharmaceutical announced that it would stop the development of DSP-1181 because the Phase I clinical trial did not meet the expected standards. Immediately, DSP1181 both disappeared from the official websites of Exscientia and Sumitomo Pharmaceuticals. Since then, attempts to develop the world’s first AI-designed drug molecule have failed.
As early as 2014, Sumitomo Pharmaceuticals favored Exscientia’s automatic compound generation technology and knowledge-based artificial intelligence prediction model, and the two parties immediately reached a cooperation. Sumitomo Pharmaceuticals became one of the first pharmaceutical companies in the world to cooperate with AI companies. In the following years, Sumitomo Pharmaceuticals and Exscientia worked together to finally select a monoamine G protein-coupled receptor (GPCR) drug for the treatment of mental illness.
In the collaboration, the chemical team of Sumitomo Pharmaceuticals synthesizes the compounds proposed by Exscientia, the pharmacology team evaluates these compounds, and the two companies share activity data together to continue to improve the drug. Based on Exscientia’s AI algorithm model, the two parties tested and synthesized as many as 350 compounds in less than a year, and DSP-1181 is the 350th compound synthesized since the project started. At that time, the average time to complete this work in the industry was more than 5 years.
In addition, the two parties are also synthesizing analogues during the course of the project. Chemists at Sumitomo Pharmaceuticals simultaneously synthesized intermediates of compounds proposed by Exscientia, and also designed and synthesized some compounds with putative pharmacological data, and fed these data into Exscientia’s predictive models. These include compounds that provide important structure-activity relationships for the optimization of compound structures, which further accelerates the drug discovery cycle and allowed the company to discover DSP-1181 in a short period of time.
At the beginning of 2020, Exscientia announced with a high profile that DSP-1181, developed in cooperation with Japan’s Sumitomo Pharmaceutical, entered phase I clinical trials. At the beginning of the clinical trial of DSP-1181, Sumitomo Pharmaceuticals was very excited and couldn’t help but praise the innovative approach adopted by Exscientia will make a great contribution to central nervous system drugs.
Regarding the failure of DSP-1181, some researchers pointed out that the root cause is that the drug molecule itself is not innovative enough.
Todd Wills of the American Chemical Abstracts Service (CAS) conducted a detailed analysis of DSP-1181 and found that the receptor that DSP-1181 acts on is a very important classic target of antipsychotic drugs. In other words, the development of DSP-1181 did not deviate from the original target. After systematic research on the patent system of DSP-1181, Wills found that the DSP-1181 molecule was very similar to haloperidol, a typical antipsychotic drug approved by the FDA in 1967. In this sense, Exscientia is likely to optimize on a long-discovered molecular framework.
The failure of DSP-1181 cast a shadow over the bright moment of AI pharmaceuticals, but also brought a key turning point to the industry. Since then, when people talk about AI pharmaceuticals, in addition to algorithms and data, they also gradually focus on innovative research in the laboratory.
After going through the confusion of the early technology and data accumulation stages, it is not uncommon for today’s AI pharmaceuticals to build a clinical trial pipeline. According to the statistics of the Bureau of Smart Medicines, the new drug pipelines developed by domestic AI pharmaceutical companies such as Iceland Stone Bio, Ruige Pharmaceuticals, Yingsi Intelligent, and Hongyun Bio have entered the clinical trial stage. At the end of June, Insilicon Intelligence was the first in the world to complete the administration of the first patient of the AI drug INS018_055 in the Phase II clinical trial.
The real difficulty is how to advance clinical trials, as many AI drugs are stuck in Phase I clinical trials. According to statistics from the Bureau of Smart Drugs, among the 80 approved clinical AI drug pipelines in the world, only 29 R&D pipelines have advanced to phase II of clinical trials, and no AI drug pipeline has entered a later stage.
After running blindfolded for 10 years, AI Pharmaceuticals began to be a little unable to run. In addition to DSP-1181, which fell into the phase I clinical trial, not long ago, Benevolent AI, another leading British AI pharmaceutical company, also announced that a candidate drug for the treatment of atopic dermatitis failed to reach the target level in the phase II clinical trial. Secondary efficacy endpoints. Insilicon, which is aggressively making new AI drugs, is extremely cautious when it comes to phase II clinical trials.
Fighting Single Point Breakthrough
Although there have been several ups and downs, for AI pharmaceuticals, there is still no clear definition in the industry. People’s attempts to use artificial intelligence (AI) technologies such as machine learning, deep learning, natural language processing, and knowledge graphs to carry out medicinal chemical molecular analysis, target discovery, compound screening, and even clinical trial research and other new drug research and development related work, that is, AI pharmaceuticals.
On many occasions, AI pharmaceuticals are regarded as the ultimate solution to improve the efficiency of new drug research and development. However, the AI technology, which is separated from the strict pharmaceutical logic, breaks through the core link of new drug research and development at a single point in a way that is separated from each other.
Specifically, in the previous stage of exploration, AI pharmaceuticals were used to complete the two extremely tedious but extremely important tasks of discovering new targets and screening compounds.
On the one hand, people hope to rely on the powerful computing and analysis capabilities of AI pharmaceuticals to discover the potential of fully exploiting difficult-to-drug targets and bypass the homogeneous competition in the Red Sea. According to statistics, in the human proteome, difficult-to-drug targets account for more than 75%, and more than half of human diseases are clinically untreatable. For targets that have been verified to be effective, such as PD-1, GLP-1, etc., hundreds of pharmaceutical companies often rush to develop them in a short period of time.
So far, AI pharmaceuticals have been used to replace many links in conventional new drug development. For example, target identification, which is a critical step in drug development and one of the most complex steps. At this stage, most of the targets used in the development of new drugs are proteins. In AI-based target discovery, researchers first extract the original features from the sequence, structure, and function of the protein, then use machine learning methods to construct an accurate and stable protein model, and finally use this model to identify the target function. Inference, prediction and classification. This has become an important means of AI target research.
In addition to structural data, multiple omics data such as genomics, proteomics, and metabolomics are extracted from patient samples and massive biomedical data, and deep learning is used to analyze the differences between non-disease and disease states. It can also be used to discover proteins that have an impact on disease.
On the other hand, AI technology may simplify drug screening, synthesis, and reduce costs. For the screened compounds, dimension conditions such as solubility, activity/selectivity, toxicity, metabolism, pharmacokinetics/efficacy, and synthesis are often required. This will involve repeated experimental processes, which is time-consuming and laborious, and will increase the cost of preclinical research. And this kind of highly repetitive, calculation-intensive work is exactly what computer programs are good at.
In this process, AI technology is used to achieve molecular generation, that is, to use machine learning methods to generate new small molecules. Specifically, AI can obtain the laws of the molecular structure and druggability of compounds through the study of a large number of compounds or drug molecules, and then generate many compounds that have never existed in nature as candidate drug molecules according to these laws, effectively constructing drugs with certain properties. Large-scale and high-quality molecular libraries.
In addition, AI technology is also used to complete chemical reaction design and compound screening. One of the areas of chemistry where AI is currently making progress is the modeling and prediction of chemical reactions and synthetic routes. Based on AI technology, the molecular structure is mapped into a form that can be processed by machine learning algorithms, and multiple synthetic routes are formed based on the structures of known compounds, and the best synthetic route is recommended. In turn, deep learning and transfer learning can predict chemical reaction outcomes given reactants. AI techniques 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 and predict the mechanism of action of compounds.
It can be said that on every independent node, AI Pharmaceuticals has done very well. But this kind of excellence is difficult to extend beyond computer software. In addition to clinical trials that cannot run, AI pharmaceuticals have been criticized within pharmaceutical companies, which is already a public phenomenon. In the interview with Arterial.com, being complained by AI pharmaceutical engineers about the low molecular activity and long production cycle, and being disliked by medicinal chemistry experts for the difficult operation of the technology platform has almost become a fate that many AI pharmaceutical companies cannot escape.
Looking back, the gap between AI pharmaceuticals and pharmaceutical companies cannot be ignored because the former pursues efficiency and verifies its own value by compressing development time, while the latter emphasizes quality and requires repeated demonstrations to select good ones. Goal, move on. In a sense, AI pharmaceuticals are walking in a straight line, striving to move forward, while the process of new drug research and development is more like a closed loop, which can be overthrown and restarted.
The actual implementation of AI pharmaceuticals may need to stop trying to make breakthroughs at a single point, and instead integrate into the closed-loop thinking of new drug research and development.
Return to the true rules of making medicine
“The hotter and more pharmaceutical companies are building automated laboratories,” an investor told Arterial.com, “the introduction of AI technology in drug discovery, chemical synthesis and other links has almost become a standard configuration for innovative pharmaceutical companies.” The author said that if the function of the automated intelligent laboratory to improve the efficiency of new drug research and development is verified, it will trigger a new wave of infrastructure construction for large pharmaceutical companies.
Arterial.com sorted out the public data and found that in the past two years, AI pharmaceutical companies have invested in the construction of automated laboratories. The laboratory environment, and multinational pharmaceutical companies such as Pfizer, AstraZeneca, and Eli Lilly have also paid for the automated laboratory of drug research and development based on AI technology.
For example, at the AstraZeneca iLab in Gothenburg, Sweden, AstraZeneca is exploring the construction of a fully automated medicinal chemistry laboratory, seamlessly integrating the closed-loop design, manufacturing, testing, and analysis (DMTA) of new drug development with the technology platform of Molecular AI, an AI new drug research and development enterprise. . Among them, AI technology mainly completes the design and analysis links in the DMTA closed loop, uses AI and machine learning to help chemists make better decisions faster, realizes effective interaction between chemists and computers, and thus accelerates the exploration of chemical space and the design of potential new drug molecules.
For another example, Pfizer cooperated with Jingtai Technology to accelerate the development of new drugs by using the “AI prediction + experimental verification” method. The latter established an automated laboratory in Shanghai.
“Drug development is a multi-dimensional simultaneous optimization process,” some practitioners told Arterial Network. The data for new drug development is huge, with complex types and structures. Building a closed-loop dry and wet laboratory can complete the design more efficiently. , verified illusory.
On the one hand, pharmaceutical companies have formed a more systematic data management method. Traditional drug research and development is based on experimental science. In the past research and development of new drugs, the recording, management and storage of data were all centered on experiments, which needed to be dynamically adjusted according to experimental needs. In other words, data is just a by-product of experimentation. As AI is a method within the category of virtual science, computing science and data science, the importance of data is self-evident. This requires pharmaceutical companies to strictly regulate the format, standard, quality, and quantity of data in drug research and development.
On the other hand, the algorithm model of AI pharmaceutical companies can also be optimized in a targeted manner, rather than simply called. AI is deeply integrated with the core business of the traditional pharmaceutical industry, emphasizing deep industry understanding and higher technical accuracy. In addition to mining new knowledge from a large number of existing papers and experimental data, it is also necessary to have the ability to fully explore and refine real-time experimental data, and optimize models and iterative algorithms based on data feedback.
“In addition to algorithm models and data, AI pharmaceuticals are increasingly concerned with biological issues.” Another practitioner pointed out. It is true that purely relying on the experiment itself can only verify the formed hypothesis, but what AI pharmaceuticals is facing is a more complex system, and many problems are still unknown. In recent years, phenotype-based drug discovery methods have begun to attract attention, that is, the direct use of biological systems for new drug screening.
How complex the problems of the life sciences are! The underlying logic of being a patented molecule is that the understanding of biological mechanisms can solve the ultimate problem of AI pharmaceuticals. The new changes in the industry may represent a positive change in the operating mode of AI pharmaceuticals, from relatively fragmented independent development based on pharmaceutical company laboratory data, clinical data, and ideal biological models, backtracking upstream, and using mathematical methods Try to deconstruct the disease mechanism from a biological perspective, and start to find drugs with the end in mind.
And this process will undoubtedly involve larger data analysis and calculations, which is also an important reason why companies with computing power such as Nvidia are deeply involved in it. “Low-dimensional models cannot be used to explain high-dimensional problems. Only by establishing tools for understanding extremely complex systems can complex problems in life sciences be answered.” Dr. Zhao Yu, deputy director of Turing Darwin Laboratory and co-founder of Zheyuan Technology, said .
For AI pharmaceuticals, the single-point breakthrough operation mode has been falsified in a certain sense, but the growth curve of the industry is always upward.
