How far is DrugGPT from ChatGPT? AI pharmaceutical companies: the "stuck neck" is not computing power but data

Source: "Science and Technology Innovation Board Daily"

Reporter: Yu Shiqi, Zhu Jieyan

Image source: Generated by Unbounded AI‌

At the 2023WAIC World Artificial Intelligence Conference that just passed, the upsurge of generative AI almost ran through the entire conference. As the transformative outlet that has received the most attention at present, investors, industry circles, and academia all have a lot of expectations for it, especially how to dig out disruptive opportunities at the application level.

AI+medicine is one of the opportunities to be seen. At this conference, Su Zifeng, chairman and CEO of Advanced Micro Devices (AMD), said in her speech that healthcare is an area where AI can really affect human outcomes and will help doctors make better diagnoses. Accelerate disease prevention research.

Her old rival moved faster. On July 12, Nvidia announced a $50 million investment in AI pharmaceutical company Recursion in the form of private equity. Its founder and CEO Huang Renxun said in the announcement that generative AI is a revolutionary tool in the development of new drugs and new therapies. Recursion is using NVIDIA's related products to conduct pioneering work in the field of biochemistry, accelerating the development of the world's largest generative AI model of biomolecules, thereby advancing the development of biotechnology and accelerating drug discovery for pharmaceutical companies.

AI pharmaceuticals has always been one of the hot spots in China, and a number of leading companies that have reached the forefront of the world in technology have emerged. When the opportunity of the times is coming, how do the front-line AI pharmaceutical companies recognize the current changes? The "Kechuangban Daily" invited He Qi, the co-founder and CEO of TB Medical, Zhang Peiyu, the chief scientific officer of Jingtai Technology, and Shenshi Technology Wang Xiaofo, the head of strategy, and three industry representatives shared the opportunities and challenges in their eyes.

The "stuck" is not computing power but data

Regarding the impact of the wave of generative AI, the common perception of the three entrepreneurs is that it has become "hot".

He Qi, CEO of TBMI Pharmaceuticals, said that the entire pharmaceutical industry is still in a cold winter, but the track of AI pharmaceuticals has begun to pick up. In March of this year, TBM completed the A-round financing of US$35 million. At that time, it received the support of many top institutions, and now many institutions have expressed interest in the business model.

Both Zhang Peiyu, chief scientific officer of Jingtai Technology, and Wang Xiaofo, head of strategy at Shenshi Technology, believe that the impact of generative AI has not yet been directly transmitted to AI pharmaceuticals, but it has already brought positive signals to the industry. Zhang Peiyu mentioned, "GPT's investment hotspots are still around large models, databases, and graphics computing. This is just the beginning. In the future, it will definitely migrate to more subdivided application layers such as medicine and manufacturing. This is the inevitable growth. process.**"

Before ChatGPT exploded out of the circle, AI-enabled new drug research and development has become the consensus of the industry. The research report shows that by enabling drug target discovery and compound screening through machine learning (ML) and deep learning (DL), the success rate of new drug development can be increased from 12% to 14%. %, saving about US$55 billion in compound screening and clinical trial costs worldwide each year. **

But on the other hand, AI pharmaceuticals are also facing bottlenecks. AI pharmaceuticals are currently mainly used in the early stages of drug discovery and lead compound screening. In the clinical trial stage, more people are still required to complete related work. At the same time, AI pharmaceuticals are also limited by the impact of data homogeneity. In a popular sense, AI learning materials are experimental data created by humans, and AI cannot create unpopular target data out of thin air. This also means that the most imaginative capabilities of generative AI are limited.

Therefore, the dilemma facing AI pharmaceutical companies is completely different from the current large model companies. Zhang Peiyu bluntly stated in the interview that computing power and algorithms are not the core barriers that restrict the development of AI pharmaceutical companies. Hundreds of GPUs and the current iterative algorithm are enough to support the needs of an AI pharmaceutical company, the key lies in the data. **

"Whether it is to simulate calculations through the advantages of AI computing power, accelerate the screening and optimization of lead substances, or design new molecular structures based on experience and big data training, a large amount of data is needed as support. For AI pharmaceutical companies, its The core is built on the data production capacity." Zhang Peiyu said.

In He Qi's view, the main reason why AI can play a relatively limited role in the later stage of drug development is the lack of data, especially the data needed in the clinical stage or translational medicine. This poses a great challenge to the training of large models.

The lack of data is not only reflected in the quantity, Wang Xiaofo further analyzed the core of the problem, "The amount of data is not enough now, because the cost of generating data through experiments is very high. What is more troublesome is that the quality cannot be fully guaranteed. For example, the same The experiment, if A does it and B does it, the results may be different. It itself is affected by many off-site factors and associated errors. The quantity and quality of the underlying data cannot be guaranteed, the direct result That is, the performance and results of AI learning will be greatly reduced."

In the opinion of several AI pharmaceutical practitioners, the road from ChatGPT to DrugGPT is tortuous and difficult. What is stuck is not the computing power but the underlying data production capacity. But in the same way, under the wave of generative AI, AI pharmaceuticals has the opportunity to bring qualitative changes to the entire pharmaceutical industry, break the bottleneck of innovation, and solve the fundamental problem of R&D efficiency.

**How far is the future of DrugGPT? **

The first thing to solve is the problem of data production capacity.

Jingtai's idea is "automation + intelligence". Zhang Peiyu believes that the process of drug research and development is a process of continuous trial and error iterations, many of which are traditionally labor-intensive, and can be completely automated to replace human labor to improve efficiency and accuracy. A lot of work they do now is to transform traditional processes into automated processes, and trace high-precision data through automated processes, and feed them back to AI models in real time. This method can collect more, more comprehensive and real data than human experiments, improve human efficiency several times, and empower humans to do more and more successful innovative explorations.

The data generated by automation continuously drives the development and optimization of intelligent algorithms. The higher the efficiency of automation, the more accurate the predictions of intelligent algorithms and the wider the scope of application. At the same time, intelligence is also reflected in the transformation of unstructured information into structured data. According to him, now AI can extract information such as synthetic routes and molecular structures hidden in unstructured documents and patents, convert them into structured data, improve the performance of algorithms, and then output the designed synthetic routes to automated equipment. Enter chemical synthesis test, data production process. In this process, AI can also play a role in scheduling and planning, efficiently calling various tools in parallel, and completing the closed loop from algorithm prediction to experimental verification for different application scenarios.

"This is a development direction worth looking forward to. In the end, AI alone can connect the closed loop of design and production, and automatically complete drug research and development." Zhang Peiyu said.

Shenshi proposed a new scientific research paradigm of AI for Science. Simply put, it is to use AI to learn the scientific laws of the underlying operation of a series of things. Wang Xiaofo said that facing the problem of data scarcity, they introduced AI into the lower-level scientific research field, allowing AI to use its powerful function fitting and data analysis capabilities to learn scientific laws and principles, and obtain usable models to solve practical problems. Scientific research issues, especially assisting scientists to conduct a large number of verifications and trial and error under different assumptions, thus greatly accelerating the process of scientific research and exploration.

At present, we can see the improvement in efficiency. Wang Xiaofo mentioned that many times of high-throughput experiments may need to be done in the drug screening process. Now, we will use the new paradigm of AI for Science to calculate, and then proceed to the calculation. For a small part of verification, we have recently tried to do an order of magnitude less experiments than in the past, and we can get candidate drugs. This is equivalent to obtaining an efficiency improvement of more than 10 times.

The blessing of efficiency has brought about lower-level changes. According to Zhang Peiyu, the automated digital intelligence laboratory designed by Jingtai for biomedicine can not only be used for drug research and development, but also can be further expanded to the direction of chemical engineering and new materials that also require experimental screening. The underlying principles are the same. But the security requirements, validation cycle and project complexity in these areas are significantly lower. This is a huge market that is not weaker than pharmaceuticals. At present, they have reached cooperation with some petrochemical, energy storage materials and other new material research and development companies.

For the future, he has quite optimistic expectations. After crossing the bottleneck of data production, AI pharmaceuticals have the opportunity to cause qualitative changes through quantitative changes. In the future, the entire process of drug development may be guided by AI, making difficult-to-drug targets and new drugs The drug-making mechanism has spawned a new generation of high-quality drugs, creating new drug pipelines and incremental markets. In 20 or 30 years, it can be expected that 90% of the work in the research and development of new drugs can be done more efficiently by AI. While the threshold for innovation is lowered, the ceiling of drug research and development will be raised, with fewer resources , time and risk of failure, so that more drugs come to patients.

At present, He Qi believes that the driving force of AI for drug research and development has reached the second curve. Biotech companies will inevitably need to invest heavily in computing when doing research and development of innovative drugs. Based on this pain point, AI pharmaceutical companies that provide equipment and computing power, as well as expert support, have been recognized by many customers. After laying the foundation for commercialization, companies can explore more AI-enabled drug R&D paths from a longer-term perspective.