Dialogue with Carbon Silicon Wisdom CEO Deng Yafeng: Why did I give up my position as an Internet executive to work in "AI pharmaceutical manufacturing"?
Latest update time：2022-09-30
Reads：
"
In this wave of AI pharmaceutical boom, almost all founders have technical backgrounds. AI pharmaceuticals is an opportunity for people like Deng Yafeng.
"
Author | Li Yuchen
Editor | Wang Chuan
For Deng Yafeng, was resigning from the position of dean of the 360 Artificial Intelligence Research Institute and devoting himself to AI pharmaceutical manufacturing a decision he would not regret?
In 2012, the AlexNet deep convolutional neural network defeated all parties in the ImageNet classification competition, and the deep learning revolution began.
A year later, Deng Yafeng joined Baidu Deep Learning Research Institute and became a senior scientist.
In August 2016, Deng Yafeng left Baidu to join GreenTon and became CTO six months later. His professional role changed from a technical expert to the manager of a technical team of 100 people.
In 2020, Deng Yafeng joined 360 Group.
His past three career stages have been four years in each, but he has always switched tracks and roles around AI.
By chance, Deng Yafeng met Hou Tingjun, the company's current chief scientist, who is a professor at Zhejiang University School of Pharmacy and has more than 20 years of experience in the field of computer-aided drug design.
At that time, Deng Yafeng was already the vice president of 360 Group, dean of the Artificial Intelligence Research Institute and general manager of the search business unit, managing a product technology team of about 400 people.
But after the two had in-depth exchanges, Deng Yafeng decided to enter the AI pharmaceutical industry.
At the 2022 Natural Language Processing Summit, Deng Yafeng made a public appearance as the founder and CEO of Carbon Silicon Intelligence, and recently completed an angel round of financing of 50 million yuan, jointly led by Lenovo Ventures and Lenovo Star.
For Deng Yafeng, everything starts over again.
01
Tear off the old labels
Deng Yafeng gave three reasons for making his own AI pharmaceuticals.
First of all, the feasibility of technology “migration”.
If you count carefully, Deng Yafeng has been working in the field of CV (computer vision) for nearly 20 years, which is a label on him.
During his three years at Baidu Deep Learning Institute, Deng Yafeng participated in proposing the first end-to-end one-stage object detection framework based on deep learning, DenseBox. Together with his team, he achieved the highest accuracy in the LFW evaluation, with a score of 99.77%, which is close to the limit.
He serves as CTO at Gelin Shenzhen, and is mainly responsible for the research and development of algorithms and software products such as face recognition, human body re-identification, and vehicle recognition.
“However, I don’t want to just do a CV because I don’t want to set limits on my life.”
So after joining 360, in addition to CV, Deng Yafeng began to work on NLP, multimodal representation learning, robotics and other fields.
Under the leap of AI technology characterized by deep learning, the migration and integration of different research directions have gradually become a trend.
"Around 2002, license plate recognition and face recognition were two different groups of people, because their technology stacks were very different. If you wanted to change careers, you had to spend two years learning. However, 20 years later, with the help of models such as CNN/Transformer, cross-border work can be achieved relatively easily within the CV field and between CV and other fields."
Like other fields where AI is being applied, drug development requires domain knowledge, which not only involves pharmacy, but also includes physics, chemistry, biology, medicine, etc. This is a very interdisciplinary field.
For Deng Yafeng, it is impossible to learn all the biomedical knowledge in a short period of time.
However, in his view, with certain domain knowledge, abstracting the requirements of specific scenarios into AI modeling problems will test the team's AI algorithm modeling capabilities and software and hardware product development capabilities.
"Our team already has a lot of students with pharmaceutical backgrounds, so we have spent a lot of time recruiting people recently, focusing more on AI algorithm talents. We hope to find people who have a deep understanding of deep learning, are interested in life sciences, and have a sense of mission. . These algorithm students need to learn some domain knowledge at the beginning, and then they can work together in the team to solve the modeling problems of each sub-task of new drug discovery. It only takes a month or two to start the project.”
Secondly, it is necessary to plan your life.
During the three and a half years from 2016 to 2020 at Megvii, Deng Yafeng was a technical manager, but after he sorted out the internal team and business, he became tired of this role.
Deng Yafeng said frankly that he left Gelingshentong because he didn’t want to keep doing the same things.
"This year Green Shen Tong was also launched. I still gave up a lot when I left. If I had stayed with Green Shen Tong, it would be much better from a financial point of view, but I still want to do something that makes me more passionate. matter."
After arriving at 360, Deng Yafeng has become a practical business leader. Whether in the Artificial Intelligence Research Institute or the Search Division, he must be responsible for the budget and consider the input-output ratio of the project.
Therefore, Deng Yafeng's decision to join 360 was a result of his intention to transform himself from a technical manager to a business leader.
The last reason is that it is the opportunity given to him by the times.
This wave of AI entrepreneurship essentially belongs to technicians. The reckless entrepreneurial atmosphere of the Internet and mobile Internet determines that entrepreneurs themselves do not need to be proficient in technology, have insight into human nature, engage in marketing fission, and spend money to recruit new people, in order to successfully build a business project.
However, this extensive entrepreneurial path has long been unsustainable in the current entrepreneurial environment.
Therefore, we can find a trend:
CEOs in the AI industry, especially those in the early stages of their business, are mostly from a technical background.
Of course, after a company matures, for commercialization considerations, technical managers may give way to business talents with expertise in sales and channels, or to professional managers. However, the thinking and organizational framework of technical managers have left a deep imprint on the company.
An obvious phenomenon is that whether it is Xu Jinbo, the first person to "predict protein structure through deep learning", or the founders of a series of AI pharmaceutical companies such as Huashen Intelligent Medicine Peng Jian, Bio-Geometry Tang Jian, Suikun Intelligent Zeng Jianyang, etc. They all have technical backgrounds and have successively received tens of millions of dollars in financing, which shows that capital recognizes this track.
AI pharmaceuticals is an opportunity for people like Deng Yafeng.
02
AI pharmaceuticals, can they be avoided?
Image
A lesson from the past?
Entrepreneurship in the AI industry originated 10 years ago. It was not until AlphaGo in 2016 that the entrepreneurial scene of medical AI officially opened.
Early medical AI is marked by medical imaging AI, but frankly speaking, the threshold for starting a business in medical imaging AI is not high: based on open source databases and algorithms, you can achieve good "laboratory" effects and easily obtain it. Financing.
The same situation also happens to AI pharmaceuticals.
In 2021, there were 77 financing events in the global AI+pharmaceutical industry, with a cumulative financing amount of US$4.564 billion. The number of financing events and the amount of financing jointly broke the financing records of previous years.
Compared with 2020, the growth rate of financing amount in 2021 reached 152%.
Deng Yafeng believes that from a technical perspective, medical imaging AI is a vertical application field of CV. There are existing image detection and segmentation algorithm models that can be borrowed, and the technical barriers are not high.
From a business perspective, medical imaging services have never found a reasonable extra charge model from consumers. At the same time, it is also difficult to deal with the ethical relationship with doctors, that is, they ultimately require the doctor’s signature and cannot truly replace the doctor.
Therefore,
medical imaging algorithms ultimately become an appendage of medical devices and are unable to create very high commercial value.
In comparison, AI pharmaceutical manufacturing is in a different situation from medical imaging AI.
In terms of the pain points of AI pharmaceuticals, the current research and development efficiency of new drugs is very low and the failure rate is very high. Deng Yafeng revealed a set of figures: "Currently manual experiments are very inefficient. A doctor can only synthesize about 100 compounds in 5 years, and the efficiency is not high."
In every aspect of new drug discovery, there is room for AI algorithms to play. AI is the key to solving pain points in the field of new drug research and development. Whoever masters the ability to develop new drugs based on AI platforms will stand out. The pharmaceutical market is a 10 trillion market with a very large room for imagination.
On the other hand, the technical threshold in the field of AI pharmaceuticals is very high. There are no mature frameworks and models that can be used directly like those in the fields of computer vision and natural language understanding. Existing algorithms need to be continuously polished to generate value, which requires a team with very strong capabilities. algorithm modeling capabilities.
AI technology in the field of new drug discovery is divided into two levels:
The first level is modeling from a microscopic and bottom-level perspective,
such as the interaction between molecules and targets or the prediction of molecular properties, and modeling of underlying physical and chemical laws based on Transformer or graph neural networks;
The second level is to model the data correlation between compounds, proteins, genes, and diseases from a macro level
.
Technologies such as multi-modal pre-training and knowledge graphs will be used here.
"In the field of life sciences, original models and methods must be developed to truly solve a specific problem in the research and development process before target customers will pay. Moreover, the target customers of AI pharmaceuticals are different from hospitals, and there are fewer policy and ethical constraints."
Like other "AI+" tracks, the core of AI pharmaceuticals lies in pharmaceuticals, which requires talents with a strong pharmaceutical background. This brings us to Professor Hou Tingjun, another co-founder of Carbon Silicon Wisdom.
Professor Hou Tingjun graduated from Peking University with both bachelor's and master's degrees. As a leading scientific and technological innovation talent in the national "Ten Thousand Talents Plan", he has won the Elsevier China Highly Cited Scholar, the WuXi AppTec Life Chemistry Research Award, and the Royal Society of Chemistry's "Top 1%" Cited Chinese author, more than 400 SCI academic papers, 30 ESI highly cited and extended ESI highly cited papers, total citations (google) > 20,000 times, H factor 70, 19 software copyrights, 43 invention patents, and has Authorized 23 items.
As a distinguished professor at the School of Pharmacy of Zhejiang University, Hou Tingjun has more than 20 years of experience in drug design methodology and applied research. In the global scholar academic influence ranking just released in 2022, he ranked third on the list of leading figures in the domestic pharmaceutical discipline.
The research team he leads is also one of the best teams in the field of artificial intelligence-assisted drug design in China.
The division of labor between the two is: Deng Yafeng, as chairman and CEO, is responsible for the company's strategic planning, operations management, and research and development of artificial intelligence software and hardware products; Hou Tingjun, as chief scientist, focuses on the company's research and development and layout in the field of pharmacy, as well as exploration of cutting-edge directions.
In September 2022, Hou Tingjun’s team, Zhejiang University’s Xie Changyu’s team, Wuhan University’s Chen Xi’s team, Central South University’s Cao Dongsheng’s team and the Carbon-Silicon Intelligence team jointly published a paper in the Journal of Medicinal Chemistry.
One problem in drug discovery is
how to effectively find new molecules with desired properties, such as biological activity, druggability, and safety. This has always been an urgent problem to be solved in drug discovery.
A major difficulty is that the estimated number of molecules in drug-like chemical space is between 10 to the 30th power and 10 to the 60th power. How to intelligently generate or identify useful molecular structures from such a huge collection has been a long-standing obstacle to de novo drug design.
Genetic algorithm (GA)-based molecule generation methods do not need to simulate the distribution of the training data set (because they do not require training at all), and therefore they exhibit higher exploration capabilities.
Therefore, the above-mentioned five teams jointly proposed two molecule generation algorithms, ChemistGA. In the single-target (DRD2) and multi-target (GSK3β and JNK3) molecule generation tasks of the model, compared with the existing traditional GA and DL molecule generation models, ChemistGA not only retains the advantages of the traditional GA molecule generation algorithm, but also greatly improves The rate at which a generated molecule with the desired properties can be synthesized and the efficiency with which it is generated.
Such cases also show that carbon-silicon wisdom has officially started on the original research path of AI pharmaceuticals.
03
Do service or
Make
pipeline route selection
It is not enough to have AI pharmaceutical models and algorithm strategies alone, a complete process must also be formed.
Drug design requires an EDA tool like the one used in chip design to help drug design experts see the overall picture of drug design and form a design platform for closed-loop iteration of data and models.
Recently, many AI pharmaceutical teams have successively proposed similar concepts, such as Zhiyu Biotech, Tianli Intelligence, etc.
According to Deng Yafeng, Carbon Silicon Wisdom has established DrugFlow, an industry-leading one-stop new drug discovery platform with complete independent intellectual property rights, including
modules such
as
target evaluation, virtual screening, lead compound optimization, and drugability prediction
, which can help pharmaceutical experts find potential drug molecules more efficiently and conveniently.
This is similar to EDA software in the field of chip design, which helps experts make better decisions and judgments. The modules such as drugability prediction, molecule generation optimization, and AI scoring are unique functions that other software in the industry do not have.
In addition, based on the DrugFlow platform, the internal pharmaceutical chemistry and computing experts of Carbon Silicon Intelligence have summarized the best drug design practices based on AI computing and provided external drug molecule design services.
Leo Han Lianyi, who has worked at the National Center for Biotechnology Information for many years and is currently a partner at Quark Capital, once told Leifeng.com's "Medical AI Gold Digging": "Before talking about the prospects of AI, pharmacists and medical practitioners will first pay attention to the business logic of AI pharmaceutical companies, where their core competitiveness lies, and what kind of track they will ultimately affect. Whether it is crystal form prediction or small molecule drug screening, how AI can find service relationships for drug development is the key."
Deng Yafeng said that
Carbon Silicon Intelligence does not position itself as a Biotech company
, but wants to be an industry AI infrastructure and service enabler by building a drug design platform based on AI and physical computing models, data-driven, and wet and dry experimental closed loops. Design services ultimately allow customers to decide whether to use software and hardware to design products or directly use carbon silicon smart drug design services.
In previous reports by Leifeng.com, according to an article published in Drug Discovery Today magazine in June 2020, 21 leading multinational pharmaceutical companies published a total of 398 papers related to "AI drug research and development" from 2014 to 2019. At the same time, 73 internal AI R&D projects, 61 cooperation projects with external AI companies, and 11 investments/acquisitions of start-up AI companies were launched.
This involves a core question:
Since AI technology is very important to pharmaceutical companies, can't pharmaceutical companies achieve it by building their own teams?
A similar thing is that the security company Hikvision has set up an AI division in addition to the camera business to do edge-end algorithms; or equipment manufacturers such as Philips and Siemens do AI algorithms such as image reconstruction and enhancement in the scanning process. In this case, the AI pharmaceutical team will face considerable challenges.
“I think it’s difficult for pharmaceutical companies to build their own AI teams. AI pharmaceutical manufacturing is different from AI security. Security is essentially a hardware business, and the challenge is the construction of the supply chain and after-sales system.”
Deng Yafeng gave his own point of view: "Traditional pharmaceutical experts or internal IT engineers have different understandings of AI. AI in the field of life sciences cannot be made with open source software. It has a high level of The threshold requires a top team to really do well. It is not easy for mature companies to build such a team. The core reason is that top AI talents will not choose a particularly traditional platform, and the cultural atmosphere of the two teams is very different. Big. Even if an AI team is established, it will be difficult to evaluate the value of AI under huge investment, and the team will face huge pressure.”
In addition,
the thoughts and perspectives of key decision-makers within the enterprise are still inevitably focused on the "value chain" of traditional businesses
, and the input-output ratio of new businesses will often be subject to internal disputes.
Therefore, Deng Yafeng believes that for pharmaceutical companies, relying on the AI capabilities of external teams is the most cost-effective option for them and is also an indispensable supplement.
04
Conclusion
It only took Deng Yafeng three months from his resignation on June 18 to officially operating Carbon Silicon Intelligence. Currently, the company has established a team of nearly 60 people.
The first priority of Carbon Silicon Intelligence is to build a design platform that truly improves the efficiency of drug discovery based on AI and physical computing models as well as software processes and hardware automation. Whether you are selling software and hardware products, providing drug design services, or jointly manufacturing drugs with strategic partners, you can talk about value realization based on efficiency.
In the past two years, many AI pharmaceutical companies such as Schrodinger, Exscientia, and Relay have gone public, from finding preclinical candidate compounds to the ensuing pharmaceutical company collaborations; from continuous large-scale financing to winning orders worth tens of billions.
The speed at which AI new drug research and development companies are conquering cities and territories has refreshed the perceptions of various pharmaceutical companies, investors, and people in the AI circle.
"There is no doubt about the value of AI to the field of new drug research and development, but whether it can grow into a large platform worth hundreds of billions remains to be verified. The key is whether you believe that the inherent model of pharmaceuticals will be changed by AI. As long as you create value, There will be ways to get rewarded.”
//
Recent Hot Articles
京公网安备 11010802033920号