Recently, computer technology leader NVIDIA disclosed its third quarter report, with its performance taking off with the booming generative AI. Its net profit in the third quarter increased by 588%, and its data center business revenue, including AI services, reached a new high, with a year-on-year increase of 2.8 times, accounting for 80% of the total revenue.
Among them, NVIDIA's AI tentacles have not missed the field of medical pharmaceuticals. On the same day as the release of the third quarter report, Nvidia and Genentech, a subsidiary of multinational pharmaceutical company Roche Pharmaceuticals, jointly announced that they have reached a strategic AI research cooperation to accelerate drug discovery and development.
Specifically, the two companies will collaborate to accelerate and optimize Genentech's proprietary machine learning (ML) algorithms and models on NVIDIA's DGX Cloud. DGX Cloud is an AI supercomputing service platform launched by Nvidia this year, which includes the generative AI application NVIDIA BioNeMo for drug discovery.
Meanwhile, in the collaboration, Nvidia will also gain in-depth insights into the challenges related to AI in drug discovery and development. Its plan is to improve BioNeMo and other platforms to further meet the requirements of models used in the biotechnology industry.
In fact, this is not Nvidia's first foray into AI pharmaceuticals. As early as 2021, it collaborated with AstraZeneca to develop the drug development model MegaMolBART, which is planned for reaction prediction, molecular optimization, and molecular generation. Its "circle of friends" also includes a group of AI pharmaceutical companies, including American companies Schrodinger, Recursion, and Chinese companies Insilico Medicine.
From the perspective of AI pharmaceutical industry chain, Nvidia, AI companies, and pharmaceutical companies are in different positions. The upstream of this industry chain are companies that provide hardware such as GPU chips and quantum computers, as well as software such as databases and cloud computing platforms, including Nvidia, Google, etc. The midstream provides AI drug algorithm development, including AI drug research and development enterprises and IT enterprises, such as Schrodinger and Recursion mentioned above, as well as domestic Yingsi Intelligent and Jingtai Technology. Downstream are pharmaceutical companies with a demand for AI pharmaceuticals, including Pharma (a large pharmaceutical company) and Biotech (a biotechnology company).
And what are the specific tasks and challenges faced by participants from various positions in AI pharmaceuticals?
As mentioned earlier, NVIDIA provides an AI supercomputing service platform, which can be simply understood as helping downstream companies' models "accelerate" and make faster and more efficient models.
Dr. Guo Jinjiang, head of the Data Science Department at the Global Health Drug Development Center (GHDDI), introduced to Interface News that previously, Nvidia's accelerated scenarios included image recognition, processing, video surveillance, and more. In contrast, the paradigm of two-dimensional image data is unified and basically composed of pixels. But in the field of life sciences, the types and formats of data will become heterogeneous. For example, proteins are based on amino acids, compounds consider atoms and chemical bonds, and genomes and transcriptomes are based on DNA and RNA sequences, respectively. The challenge lies in how to meet the accelerated processing needs of various data in the field of life sciences, such as the need specifically for protein 3D structures.
As for downstream companies, the current situation of AI pharmaceuticals can be seen from the aforementioned Gene Tech ML algorithm and model. According to public information, the accelerated and optimized Gene Tech "lab in a loop" is an iterative framework used to generate and explore molecular designs with predictive properties.
Simply understood, this iterative framework consists of two parts, namely experimental data and generative models. This year's hot ChatGPT has made generative AI well-known to the public. Guo Jinjiang told Interface News that generative AI has also been one of the hot topics in the AI pharmaceutical field in recent years. Compared with the most mainstream "virtual screening" in the field before, its biggest feature is that it can provide new solutions in an ideal situation in one step.
The most important topic in the pharmaceutical field is to develop and design drugs based on disease-related target information, so that drugs can activate or inhibit target proteins, thereby exerting their effects. It's like first figuring out the shape of the keyhole, then finding a key that can unlock it. Initially, pharmaceutical professionals typically screened out molecules from known compound libraries that could potentially bind to target proteins, and then conducted wet experiments to determine whether the lock could be unlocked by continuously removing the potentially correct key from the keychain and personally placing it in the lock eye.
With the accumulation of data and technological development, this process has become virtualized in the AI era, which has led to the emergence of "virtual filtering". Guo Jinjiang introduced that there are two methods of "virtual screening" at present. One is to determine the affinity between a compound and a target based on physical information such as molecular mechanics and quantum mechanics, such as the spatial positions and different types of interactions between molecules and atoms, in order to determine the activity of the compound.
The second is data-driven, highly dependent virtual filtering of high-quality data. The process of AI judgment is an invisible black box, which means that there is no need to know what each key looks like or how to stay tightly in the lock hole. Instead, a large amount of data from previously active and inactive compounds is learned through machine learning to directly determine the activity of candidate molecules.
In short, under virtual screening, pharmaceutical people no longer have to manually insert keyholes. But the disruptive aspect of generative AI is that this set of alternative keys doesn't even have to exist. According to the instructions of the pharmaceutical manufacturer, that is, the description of unlocking requirements and keyholes, generative AI can directly provide the correct model drawing of the key.
Guo Jinjiang told Interface News that the most attractive aspect for pharmaceutical people is that the answers given by generative AI are not only correct and can unlock, but also often a brand new key, which is not any known or already synthetic compound in the compound library. In the current situation where the low hanging fruits are gradually being harvested, the input-output ratio of drug research and development is decreasing, and competition is becoming increasingly fierce, such potential "First in class" molecules undoubtedly have infinite charm.
In addition, the time required for generating AI to provide results is also shorter. For example, Guo Jinjiang used a library screening of 4 million compounds and a protein with a size of 60000 atoms as an example. Using physics based "virtual screening" would take about a week. The ideal generative AI is to provide answers in one step.
However, the current AI model is still unable to achieve such an ideal level, relying on experimental data to iteratively become the other half of this framework.
Guo Jinjiang explained that the model drawings of the correct keys are easy to obtain, but the actual keys made may not necessarily work. The most significant issue is composability. Whether the synthesized compound is stable and exhibits good activity in the protein or cell layer still needs to be verified and optimized in real experiments.
At present, the usual approach is for pharmaceutical manufacturers to first obtain tens of thousands of compound molecules through generative AI, and then use virtual screening methods to screen them. At the same time, experienced drug development experts will also modify molecules, conduct wet experiment measurements, and return the experimental results to the AI model to help them relearn, thus becoming more and more suitable for research and development tasks.
In fact, the reason behind this is still the complexity of the human body itself and its interactions with drugs. At present, the scientific and pharmaceutical communities may only be observing a tree without seeing a forest, which is also why AI pharmaceuticals need collaboration from upstream and downstream parties in the industry chain to promote model progress.
However, some of the paradoxes are that on the one hand, traditional large pharmaceutical companies tend to be more conservative in their mindset, with a small portion attempting in the AI pharmaceutical field, but more still doing so in traditional ways. Pharmacists generally prefer to rely on their personal experience, such as conducting experiments in the early stages and optimizing and modifying molecules based on experience. Only complex systemic problems that are difficult to solve beyond experience can they hope to find AI solutions. At the same time, there is currently no AI developed drug that has been truly approved for market globally, which means that the path of AI pharmaceuticals has not been truly paved. This is also one of the reasons why pharmaceutical companies are in a wait-and-see state and find it difficult to all in.
In addition, data has always been a dilemma that AI Pharmaceuticals has not broken through. After all, data is not only the nourishment of AI life, but also the core asset of pharmaceutical companies. In this collaboration between Nvidia and Genentech, it was also pointed out that Genentech has the right to decide whether to share its proprietary data, and Nvidia cannot directly access such data without authorization from Genentech.