Produced by | Tiger Sniff Medical Group
Author | Chen Guangjing, Wang Yipeng
Edit | Wang Yipeng
Header | Visual China
Nvidia's ambitions in the life sciences sector cannot be hidden.
After the rise of generative AI, Nvidia CEO Jensen Huang frequently talked about life sciences in public, declaring that "life science engineering" was coming. Facts have proved that Lao Huang is not playing falsely - at the just-concluded NVIDIA 2024 GPU Technology (GTC) conference, medical health and life sciences became a big hit, and according to industry statistics, at least 90 of the more than 900 events in total were related to the field.
In particular, on the opening day of GTC, Huang officially launched 25 "microservices" related to medical and biopharma, covering areas such as medical imaging, drug discovery and digital health. The goal is to enable "global healthcare organizations to take advantage of the latest advances in generative AI anywhere and on any cloud."
Prior to this, the industry's perception of NVIDIA was more focused on "selling cards". "Computing power, especially GPUs, is an indispensable means of production for training AI models. Song Le, chief technology officer (CTO) of Baitu Biotech, told Tiger Sniff.
And NVIDIA's layout is obviously beyond the "duty" of a hardware supplier. Judging from public data, the global chip giant has reached cooperation with more than 2,500 related companies, including AI companies that provide algorithms, as well as giants in the medical and pharmaceutical fields. For example, we have developed a surgical robot with Johnson & Johnson that can analyze data in real time, and we have developed a "Sky Eye CT" with GE that can "automatically" take images of patients.
In the past year, Nvidia has included at least 10 leading companies in the AI pharmaceutical field through investment and other means.
This is not the first time that a tech giant has entered the life sciences field, but the timing of 2024 has a special meaning with Nvidia itself.
"In the Internet/cloud computing industry, it may be the first time in the industry that Nvidia has mentioned the layout of the medical field to a strategic position. Xie Changyu, a professor at the School of Pharmacy of Zhejiang University, told Tiger Sniff.
In the medical field, algorithms are not believed
Insights into NVIDIA's healthcare landscape must be traced back to the development of generative AI itself. However, generative AI, which seems to have infinite stories to tell at both the technical and business levels, can be regarded as a rock in the field of healthcare.
Judging from the industry's response, AI is still in its infancy in the medical and health field, and many companies, especially pharmaceutical companies, are still waiting and seeing. The relevant person in charge of a well-known domestic pharmaceutical company has publicly stated that the company's work to introduce large models is still in the cost accounting stage.
This hysteresis is not a problem for any one company. According to the data, more than $250 billion is spent on new drug research and development in the world every year, of which only more than $1 billion is invested in AI pharmaceuticals, and it is expected to not exceed $3 billion by 2026.
According to a survey by IDC, only 14% of healthcare and life sciences-related companies "have already made significant investments in generative AI and have a spending plan for GenAI-enhanced software and consulting services through training in the next 18 months," well below the global level (34%) and only half that of manufacturing.
In contrast to the alienation of AI in healthcare companies, technology vendors have maintained a great deal of enthusiasm. According to the data, there are more than 700 AI+ biomedical companies in the world alone, and in the field of medical and health care, giants such as Google and IBM have long been deployed. Especially after ChatGPT became popular, nearly 50 medical models have emerged in China alone, covering almost all categories, from ToB's doctor assistants to ToC's personal health butlers.
Medical health is like a calm "iceberg beauty", "there are many suitors", but they are extremely pragmatic. She made it clear that the algorithm is just a young man's "development potential", and the data can represent his "background and net worth".
And that's where most technology vendors get their sore spots.
Some industry investors revealed that because it is difficult to obtain good data, about 80% of the medical models have not entered the second stage (that is, the feeding of professional data is enhanced in a certain field), and the vast majority of the 20% who have entered the second stage have not carried out high-level fine-tuning for different scenarios and tasks.
When some startups evaluate their internal medical models, they also find that their capabilities are only at the level of "assistants" and cannot be called "companions". As a result, AI cannot be independent in the short term, and its scope of work is limited to guidance, assisting in reading films, and writing medical records.
In the biopharma space, AI has performed well in the fields of protein structure prediction, protein generation, and even antibody drug generation, but these are still in the preclinical stage. In the most time-consuming and expensive clinical trial stage, which accounts for more than 70% of the total R&D cost of new drugs, only simple tasks such as recruiting patients and making records can be done. There is no substantial improvement in pain points such as the low success rate of new drug research and development.
Take Biomap as an example. Their long-term goal is to use AI to simulate living systems, such as the human immune system, and ideally better predict the entry of antibody drugs into the human body. But it is not easy to land in reality. "To achieve this goal, we need to break through a series of AI modeling problems at multiple scales. ”
Song Le, CTO of Baitu Bioscience, told Tiger Sniff that because there is a relatively large amount of data in the protein field, which has reached a scale of more than one billion, the fastest progress in this field can not only generate proteins, but also generate functional proteins according to diversified design goals, such as antibody drugs with good druggability, enzymes with relatively high catalytic efficiency, etc.
However, in simulated life systems engineering, it not only involves the generation of a single protein molecule, but also involves a large number of protein interactions, intracellular and intercellular interactions, etc., but this aspect of the problem is more complex, relatively speaking, data is in a scarce state, which will require continuous innovation and breakthroughs in AI models and experimental data acquisition methods.
"It's conceivable that protein data will grow exponentially, and the number of cases for generating designs will grow rapidly. But that's only part of the story, and to simulate the immune system, you need to have the same amount of data at other levels, and the innovation and iteration of AI models that go with it grow just as fast. Song Le said.
So with data, can AI be unimpeded in the healthcare field? Not really.
For example, there are relatively more open data in the medical field, and there are more ways to obtain it, and the progress of AI + medical care in this field is also faster. At the beginning of this year, Google launched a medical dialogue AI - AMIE (Articulate Medical Intelligence Explorer). This app is another "AI doctor" after Med-PaLM and Med-PaLM2, and even passed the Turing test, and its performance can be described as amazing.
Although in terms of specific performance, this can do better than health care physicians in the field of diagnosis in the field of cardiovascular diseases, but it still cannot be easily used on real patients, Harvard Medical School experts bluntly said in an interview that medical treatment is by no means as simple as collecting information, "it is about human relationships."
Behind it, ethical, regulatory, institutional, and scientific research itself are barriers that are difficult to break through. It can be said that the frequently mentioned research data is only an entry threshold, and AI+ medical health and life sciences are not essentially to solve a technical problem, but a comprehensive social problem.
Xie Changyu told Tiger Sniff that stronger hardware and algorithms must be helpful in accelerating the development of the industry, but it does not mean that with 1,000 more GPUs today, I can tell you what kind of progress AIDD (artificial intelligence drug discovery and design) will have tomorrow.
Artificial intelligence is also being used to improve the accuracy, safety, and more of surgical robots.
The picture shows a small audience watching a single-port laparoscopic surgical robot peeling quail eggs.
From: Visual China
Nvidia version of "curve to save the country"
Therefore, like all technical services in the past, it is absolutely impossible to simply develop and deliver technical solutions to Party A by Party B independently. Experts in the medical and pharmaceutical industries must participate in the evolution of technical solutions, which is the final conclusion reached by AI companies after 60 or 70 years of repeated failures in the medical and pharmaceutical fields.
In this partnership, pharmaceutical and medical device companies are not only the payers, but also the producers of data, which not only solves the problem of who pays, but also provides enough data for product iteration. Today's most mature AI+ medical imaging is a typical case.
Giants such as GE and Siemens have artificially built data mining machines by tying AI to large-scale equipment. As long as the CT machine and MRI machine work normally every day, they can continuously provide data nourishment for AI. In the same way, in the field of pharmaceuticals, no one is more confident in collecting data than pharmaceutical companies that fund R&D, and AI companies have a chance to break the game if they can participate in the workflow of pharmaceutical companies.
Everyone understands this principle, but it is actually difficult to do, and the reason is very simple: pharmaceutical companies do not agree.
For pharmaceutical companies, this data is too expensive. Large pharmaceutical companies with a large amount of R&D data have invested billions of dollars in R&D every year, and companies with a long tradition of R&D such as Merck and Roche have long exceeded $10 billion in related investment, and there is a trend of increasing year by year. These data, which rely on heavy money, not only have the opportunity to achieve popular items, but also widen the distance from competitors in the future competition.
In today's increasingly competitive pharmaceutical market, no company is willing to easily contribute data. In this case, if you want to get the clinical data within the enterprise and hospital, you can only tailor the model for them, so that they can feel the power of the large model in the "safe zone".
However, the cost of this operation is also extremely high. Some industry insiders revealed to Tiger Sniff that a similar situation occurred in the process of cloud computing promotion, and eventually many companies had to stop related businesses because of serious losses. The cost of customizing large models will only be higher, and he believes that the agreement amount will most likely have to reach the $1 billion level to break even.
Nvidia's strategy for this is to "save the country by curve".
NVIDIA is a good hand in building an ecosystem, and it laid a competitive barrier in the chip industry with CUDA (compute Unified Device Architecture) more than 10 years ago (2006).
To put it simply, CUDA is a general-purpose computing architecture based on the combination of hardware and software designed by GPUs. There are two main advantages, one is that users can directly control the chip in combination with the GPU, and the other is that the CUDA architecture provides free development tools for software manufacturers to facilitate software development. The former greatly lowers the threshold for GPU use, while the latter is user-friendly and quietly digs into the "moat" - as there are more and more parts in the "toolkit", its substitutability is also greatly reduced.
The resulting CUDA ecosystem has deeply bound NVIDIA to AI, and NVIDIA has gradually thrown off its competitors and achieved a counterattack.
When entering the healthcare and life sciences sector, NVIDIA has the momentum to replicate the successful experience of the CUDA ecosystem.
According to NVIDIA's public information, the 25 microservices they launched this time are actually easier to operate versions of the previous "professional version" - these microservice suites include optimized NVIDIA NIM™ AI models and workflows, and provide industry-standard application programming interfaces (APIs) for creating and deploying cloud-native applications.
In other words, hospitals and pharmaceutical companies can directly meet their needs in areas such as medical imaging, natural language and speech recognition, as well as digital biology generation, prediction and simulation functions, at the click of a "button" according to their needs.
For AI companies, NVIDIA's brand endorsement and industry influence have brought more opportunities. "It's a diversion platform. Song Le told Tiger Sniff. In 2023, Baitu Biotech joined NVIDIA's "NVIDIA Startup Acceleration Program" to recruit startups, and at this year's GTC conference, Song Le also shared his experience in doing AI + biomedicine in the past three years on behalf of Baitu Biotech as an ecological partner.
On the other hand, NVIDIA has also lowered the threshold for the use of large models, making it easier for traditional hospitals and pharmaceutical companies to use the "toolkits" they provide, such as various AI large models. With the gradual enrichment of the toolkit and the habits and dependence of industrial partners, a new and irreplaceable ecology has been formed.
Do you mind if software companies are directly involved in business processes and share data? OK, I'll sell you the "dumb" tools, and you can do it yourself. Nvidia is not involved in the high-cost thing of customizing the model, no matter how the pharmaceutical companies and cloud computing companies toss, as long as it is based on my infrastructure, this business model is established.
Nvidia's abacus is very loud, but it should be noted that CUDA used to face the software industry, which is completely different from the medical industry, and if you just copy CUDA's play, it will not even be able to open up the medical market.
Compared with the Internet industry, medical and pharmaceutical are very traditional and closed, with their own unique processes. For example, in China, convincing a hospital to purchase a certain product requires not only a complex process, but also the right channels. At the grassroots level, where AI is considered to be in need of the most, many hospitals' procurement channels are in the hands of individuals or small agents, and if they can't be found, no matter how good the product is, they won't be able to do it. Therefore, many AI+ medical companies have not been able to make a profit in the past 10 years.
Moreover, the determination of technology giants to "capture" the positions of hospitals and pharmaceutical companies has always been firm, resulting in this market not yet developing and has become a red ocean.
AWS, Tencent Cloud, Baidu Intelligent Cloud, etc., are all "staking their ground" in this track, in order to win as many partners as possible, and even the relevant person in charge of a domestic technology company publicly and bluntly said that "the right to build AI scene models should be returned to scientists", which is almost a confession to the other party, which can meet all the customized needs of the other party at any cost.
In addition, the "toolkit" provided by NVIDIA is still at the level of the "public version", and hospitals and pharmaceutical companies still need to cooperate with AI companies if they need a "professional version" of large models. This also makes it difficult to form a "moat" similar to the CUDA ecosystem.
For a long time, the domestic business for large Party A has been torn between the buyer's strong demand for customization and the seller's "ability to lose money". In contrast, NVIDIA's approach, while lightweight, struggles to meet the needs of healthcare organizations and physicians, who are often more accustomed to active door-to-door marketing and lack the incentive to proactively find a "toolkit" for training data.
Technology companies that entered the medical industry earlier helped grassroots hospitals use remote technology to improve the level of diagnosis.
From: Visual China
There may be an opportunity to reconcile the cost battles
Nvidia's opportunities are more trendy - the industry's main behavior towards AI is wait-and-see, but it also acknowledges in its attitude that AI is no longer a dispensable thing, and it will fundamentally reshape the life sciences industry.
Since the 80s of the last century, the main battlefield of new drug research and development has shifted from small molecule chemical drugs to the field of biological drugs with more complex structures, and the total drug screening space can reach 10 to the 60th power.
"AI is gradually becoming an indispensable basic experimental equipment in biomedical research and development. Song Le explained to Tiger Sniff. He believes that in the future, the demand for AI and computing resources in the research and development of biological drugs will be increasing.
Now, the application of AI technology has shortened the time to explore preclinical compounds in new drug development from 3/4 of the original 4 years to 13.7 months, and even compressed to one month or more than 20 days. According to the McKinsey Global Institute (MGI), generative AI is expected to bring hundreds of billions of dollars in economic value to the medical and pharmaceutical industries every year.
From: McKinsey
Such a statistic is certainly impressive, but it is not a fatal attraction for an industry that generates more than a trillion dollars in global sales every year. As mentioned above, the biggest expense of new drug research and development is in the clinical trial stage, and what AI can do in this field is still very limited. At the same time, there is not a single AI-designed drug on the market, and it cannot be proven that AI can improve the success rate of new drug development.
Therefore, in the process of cooperating with AI companies, pharmaceutical companies invest more tentatively in small amounts, and are more inclined to cooperate on details similar to codon optimization. This makes it difficult for many AI pharmaceutical companies to find an outlet for their commercialization ambitions.
But what does this have to do with NVIDIA? At least for five years, Nvidia will still sell infrastructure, not drug development solutions, and it is much more reliable to let pharmaceutical companies pay for the attitude and build infrastructure first than directly throwing money at customizing large models.
Moreover, in China, the demand for AI from medical institutions and pharmaceutical companies goes far beyond the research and development of new drugs.
Since 2018, the reform of cutting off the interest chain of "medicine for medical care" such as the procurement of drugs and consumables, as well as the new medical reform policies such as national medical insurance negotiation and medical insurance payment method reform, have jointly announced the end of the era of profiteering in the entire drug research and development, production, circulation and use links.
In the crazy price reduction wave after round, any member of the entire industrial chain must make careful calculations and find ways to reduce costs and increase efficiency. In addition, after the escalation of the pharmaceutical anti-corruption storm, the circulation enterprises responsible for selling drugs must also find a compliant marketing method as soon as possible.
For all of these issues, AI is a lifesaver. And NVIDIA is the most eye-catching company on this "straw".
People who are changing and want to change the world are all in the Tiger Sniff APP