7 people are diagnosed with cancer every minute, what can the AI do?

Author 丨Wu Ying Gang

Editor 丨Zhu Ping

Figure Source 丨 Figure worm

Recently, the National Cancer Center released China's latest cancer report "2016 China Cancer Incidence and Mortality Rate" (the report covers about 380 million people, accounting for 27.6% of China's total population, because the national data collection, collation and verification require a lot of time, so the report is usually delayed by 2 to 5 years), showing that In 2016, there were about 4.064 million new cases of malignant tumors in the mainland, with an average of more than 10,000 people diagnosed with cancer every day, and 7 people diagnosed every minute, of which 828,000 people were diagnosed with lung cancer, and nearly 660,000 patients died, accounting for more than 20% and 27% of all cancers, respectively.

However, in the case of high cancer incidence and mortality, new drug research and development has encountered obstacles, and high research and development costs, long cycles, and low return rates have become the three major mountains that hinder the development of new drugs by pharmaceutical companies.

According to Nature magazine, the cost of new drug development is growing rapidly, with an average cost of about $2.6 billion in 2018. Iqbrow has issued a report that the average development time of new drugs from the beginning of clinical trials to the end of research and development has increased by 26% in the past 10 years, reaching 12.5 years in 2018, but the development success rate has been declining, falling to 11.4% in 2018.

Other data shows that the return on new drug investment is declining, with a return of 10 cents for every $1 invested in research and development a decade ago, and now it has less than 2 cents. In 2017, the return on R&D investment of the world's top 12 pharmaceutical companies was only 3.2%, down 7 percentage points from 10.3% in 2010.

It is worth noting that the rapid development of artificial intelligence (AI) has provided new ideas for new drug research and development, and in January 2020, the world's first AI-designed drug candidate entered phase I clinical trials, and new drug research and development also officially entered the "first year of AI pharmaceuticals". So how much efficiency can AI really improve for new drug development? How much is the cost saving?

The "AI+ drug innovation" market is hot

It is understood that the current application scenarios of AI cover the stages of pre-research, drug discovery, pre-clinical research, clinical trials, etc., such as the early research application scenarios for literature data integration and analysis, the construction of new drug research and development knowledge base, etc., and the drug discovery application scenarios are target discovery, crystal form prediction, compound screening, etc.

Market research firm TechEmergence found that the rate of new drug research and development with artificial intelligence increased by 2%, and Goldman Sachs group also pointed out in the report that artificial intelligence technology is expected to reduce the cost of new drug research and development by $28 billion per year after maturity.

Starting in 2020, the popularity of AI in drug research and development has been rising.

In January 2020, the world's first AI-designed drug candidate, in collaboration with Sumitomo Pharmaceutical of Japan, entered Phase I clinical trials in collaboration with The British AI pharmaceutical company Exscientia.

In March 2020, Eli Lilly partnered with AI pharmaceutical company AbCellera to develop the covid-19 neutralizing antibody bamlanivimab, which screened more than five million immune cells and identified more than 500 unique, fully human antibody sequences, and on November 9 of the same year, the drug was granted an Emergency Use Authorization (EUA) from the U.S. FDA.

According to the data of Zhiyan Consulting, in 2020, the financing amount in the domestic AI+ pharmaceutical field reached 3 billion yuan, an increase of 355% year-on-year, and the popularity continued to rise in 2021, with the financing amount reaching 7.9 billion yuan, an increase of 4.9 billion yuan over the previous year, an increase of 164% year-on-year. According to relevant reports, some insiders have analyzed that the wave of domestic AI pharmaceuticals is mainly dominated by TMT funds, and a large part of the hot money pouring in comes from TMT funds.

The international market is also hot, in 2021, there were a total of 77 financing events in the global Al+ pharmaceutical industry, an increase of 24 cases over the previous year, an increase of 45.28% year-on-year; the cumulative financing amount reached 4.564 billion US dollars, an increase of 2.756 billion US dollars over the previous year, an increase of 152.43% year-on-year; the number of financing events and financing amounts jointly refreshed the financing record of the past years.

In January 2022, Sanofi announced that it will cooperate with AI pharmaceutical Exscientia to develop up to 15 drug candidates in the field of oncology and immunology, with a transaction amount of up to $5.2 billion, and "AI+ drug development" has once again attracted widespread attention.

A revolutionary breakthrough has not yet been brought

So what degree of change and breakthrough has AI brought to the development of new drugs?

According to the relevant research released by the journal "Pharmaceutical Progress", AI can help pharmaceutical companies reduce costs and increase efficiency to a certain extent, but from the perspective of the application practice that has been carried out, AI is not a "panacea" and cannot improve the efficiency of clinical trials overnight, and the current AI technology cannot bring revolutionary breakthroughs to improve the efficiency of new drug research and development.

Sun Xiangyu, a senior analyst in the pharmaceutical industry, also told the 21st Century Business Herald that AI technology has not yet met expectations in the research and development of new drugs, and pharmaceutical companies are willing to pay to see the actual role of AI in research and development.

"Before the artificial intelligence is in the concept stage of emerging technologies, compared with the traditional research and development model, artificial intelligence is indeed a new idea, but at present, artificial intelligence has not yet reached the industry expectations, and there are still few breakthrough cases at the application level." There are more and more AI pharmaceutical research and development enterprises, but pharmaceutical companies are becoming more and more pragmatic, but the concept is good, pharmaceutical companies are not necessarily willing to pay, AI companies must come up with solid things, so that pharmaceutical companies can see that AI is really useful. Sun Xiangyu said.

Researchers such as Jiang Hualiang, academician of the Chinese Academy of Sciences, have analyzed that overall, the prospects of big data and AI technology in the field of new drug research and development are bright, but limited by the complexity of biology and the lack of clinical databases, these technologies are mainly applied in the drug discovery stage, and clinical trials are the "card neck" link of current AI applications.

Sun Xiangyu pointed out to the 21st Century Business Herald reporter that the screening of leading compounds in the drug discovery stage is currently a link with a large impact on artificial intelligence and a more concentrated application scenario, and the traditional screening method mainly relies on manual screening to take time and effort, there is a great cost of funds, if it can be simulated and calculated through artificial intelligence first, and then rely on manual optimization, the efficiency may be higher.

According to related research, for decades, identifying potential drug candidates has been high-throughput screening (HTS) – storing pathogenic proteins with pre-synthesized, drug-like compounds in a chemical library, compounds that show a strong signal will be further characterized and chemically modified, often taking a long time and cost.

AI technology can establish virtual drug screening models, quickly filter "low-quality" compounds, retrieve faster, have wider coverage, and use machine learning technology to select high-potential drug candidates from a large number of compounds.

But Nano Magazine has published an article analysis pointing out that in the drug discovery stage, artificial intelligence has three problems: first, artificial intelligence is trapped within the boundaries of known data - can learn from known data, predict in the field, but does not have the ability to perform well outside the field; second, in drug discovery, artificial intelligence needs data on effective drugs and also needs data that does not work, but the latter is basically absent in the published biological science literature; third, Most life science research results are not repeatable, so any data-based AI prediction is worth considering.

Compared with small molecules, AI in the field of macromolecules is more difficult, it is understood that protein macromolecules different three-dimensional structure to create different functions, but protein structure prediction is very difficult, through the amino acid sequence can predict protein structure and function is the dream of many biological scientists.

It is worth noting that in 2021, Science magazine released the annual scientific breakthrough list, and AlphaFold, an artificial intelligence technology that Predicts Protein Structure by Google's AI team, ranked first. In July 2021, the journal Nature published an article in which AlphaFold2 deciphered 98.5% of human proteins, of which 58% of the predictions of individual amino acid loci reached sufficient confidence, and 36% of the predictions provided detailed atomic characteristics for drug development.

Ma Rui, a partner at Fengrui Capital, once said that AlphaFold completed the process of 0 to 1, and solved the prediction from sequence to structure well, but the prediction from structure to function has not been thoroughly studied, and the ideal situation is actually the opposite - after giving a target, directly design macromolecular sequences, can be combined with the target, or achieve specific biological functions, which are still a certain distance from the application layer.

Nano Magazine published an article analyzing that evolutionary rules constrain the spatial structure of proteins, and in recent decades people have collected a large amount of experimental data, which has created a perfect environment for the application of artificial intelligence technology. But drug discovery, by contrast, is a truly different environment, with existing data on protein and drug binding very sparse and often unreliable.

The difficulty of cultivating compound talents

It is understood that the current AI is also facing the challenge of data, Sun Xiangyu pointed out to the 21st Century Business Herald reporter that in the cooperation, the most core data pharmaceutical companies may not be willing to share it, even if the data is shared, it may not be able to use it well - the chemical drug molecular weight is small, the structure is clear, it is relatively easy to calculate, and the data accumulation is also more, which may be more in line with the AI application idea; but the biological drug molecular weight is larger, the three-dimensional structure is complex, and there are only decades of development history, and the data accumulation is not particularly much.

According to research published in the journal Advances in Pharmacy, the complexity of biology poses huge challenges to data acquisition and AI algorithm design.

Because pharmacy is a blend of chemistry and biology, the two differ greatly at the data level. In general, the chemical data is more stable and controllable, easy to calculate, biological data involves a variety of signals of receptor proteins, it is difficult to calculate quantitatively, and the binding and reaction process of compounds and human targets is very complex, the current theoretical understanding is insufficient, the environmental impact factors are large, and the data stability and reproducibility are poor.

In addition, AI is also subject to high-quality data constraints, and the mainland's pharmaceutical big data has problems such as small amount of data, incomplete data system, inconsistent data standards, and imperfect data sharing mechanisms, such as medical records, follow-up records are currently difficult to standardize and digitize, and clinical data is difficult to use flexibly due to patient privacy.

Liu Pengcheng, associate professor of the International Medical Business School of China Pharmaceutical University, believes that no matter how much data there is and how complex the data is, there will always be ways to solve it with the change of technology, and the biggest problem at present is the lack of compound talents to promote technological development. "Usually colleges and universities with strong pharmaceutical majors, computer majors are not dominant majors, and vice versa. The weak development of interdisciplinary disciplines and the lack of training of compound talents are major problems restricting the development of the industry. ”

According to relevant data, there are about 22,000 high-end researchers in the field of AI in the world, while only about 600 in China, with a large gap in talent demand.

Liu Pengcheng told the 21st Century Business Herald reporter that in fact, there was a long time ago the computer-aided drug design profession, involving simple programming, but "AI + drug research and development" is an interdisciplinary discipline, there must be a disciplinary collision, only the unilateral disciplinary advantages of the school lack of conditions, only in the comprehensive university can achieve a wide range of breakthroughs, but the domestic talent training in the interdisciplinary is lacking, behind which may involve the problem of talent training methods.

"Now most of the pharmacy talents are pure traditional pharmacy background, there are few other professional backgrounds, doctoral studies are undergraduate pharmacy, graduate students read pharmacy, doctoral still read pharmacy, very few students can choose computer science during graduate school, because the knowledge of a discipline is already a lot, if there are two discipline backgrounds, the knowledge system is more complex, the difficulty of training will increase, which is related to the professional training mode of Chinese universities." Liu Pengcheng pointed out.

How does the business model work?

Despite the difficulties, AI pharmaceutical companies still have to survive and develop.

In Sun Xiangyu's view, the business model determines whether AI companies can survive, there are currently three models, the first is that AI companies establish a platform for pharmaceutical companies to use, improve the efficiency of new drug research and development, similar to selling a software to pharmaceutical companies, the second is to sign a cooperation agreement with pharmaceutical companies to develop drugs together, and the third is that AI companies develop new drugs themselves and then sell them to pharmaceutical companies.

At present, these business models are still in the exploratory stage and need to be verified by the market. For the first, the willingness of pharmaceutical companies to pay is actually not very high, and the number of pharmaceutical companies is limited, AI companies have a ceiling for revenue every year, the second and third models risk is that there is no cash flow income in the early stage, and they need to rely on financing support, and AI companies themselves lack pharmaceutical talents and clinical trial capabilities, independent research and development can only rely on outsourcing, and the model is like "small Biotech".

According to Flint Creation Analysis, the business model of "AI + drug development" is not yet clear, and the real output of AI drug research and development is less, and in April 2019, IBM decided to stop developing and selling the drug development tool Watson Artificial Intelligence Suite due to sluggish financial performance. At present, most enterprise development relies on financing, for AI+ drug research and development technology innovation enterprises, whether to do drug research and development or CRO model, is the need to combine their own development to make a suitable choice.

Another analysis shows that the rapid development of artificial intelligence in drug research and development may lead to lag behind regulatory laws and regulations.

However, Sun Xiangyu pointed out that at present, there is no particularly large regulatory obstacle to the application of AI in the development of new drugs, because the review and approval department mainly looks at the final safety and effectiveness of drugs, rather than which research and development method is used, but in the process of cooperation between AI companies and pharmaceutical companies, how to do a good job in the protection of intellectual property rights of both parties may be a problem that needs to be considered.

Liu Pengcheng also said that artificial intelligence is a methodological problem, and the regulatory side focuses on the safety and effectiveness of the results. But if AI is used for the prediction of patient drug regimens, regulation will be very important - if AI is added in advance to discriminatory algorithms, treating the rich and ordinary people differently, then it will be unfair, but AI algorithms are opaque, and it is not known what rules AI uses and how to calculate, so it will be very difficult to supervise.

Algorithm opacity may also be a problem faced by AI pharmaceutical companies in drug research and development, some studies have shown that the output results of artificial intelligence algorithms are difficult to predict and explain, lack of transparency, and developers cannot know the reason and basis for the priority ranking of small molecule drugs obtained by virtual screening, and may not even be able to explain the designers, that is, the "black box phenomenon". However, the accuracy and transparency of artificial intelligence are inversely proportional, the higher the accuracy, the lower the transparency, and to solve the black box problem, you need to weigh between the two.

"In the past, when writing programs, algorithmic rules were formulated through 'if else' conditional statements, etc., and computers ran according to the rules, but AI did not need to write rules, only needed to provide AI with a large amount of data, continuous training, AI can build a set of rules on its own, which is the advantage of AI, but also its problem." Liu Pengcheng said.

Liu Pengcheng further pointed out that although AI has not yet made revolutionary changes in the field of drug research and development, it cannot replace traditional methods, but it will certainly have a major impact in the future, "The development of artificial intelligence is like we watch the train on the platform, we hear the sound from a distance, we can't see the train, but once the train arrives, it will speed in front of you, leaving you far behind, leaving only a gust of wind." ”

This issue is edited by Feng Zhanpeng intern Kailin Li