As AI-driven drug development moves beyond the initial stage of technical feasibility verification, technology and capital in the biopharmaceutical field are rapidly converging on this track, but this has also exacerbated the resource differentiation among different research directions: compared to the more popular drug development for oncology and autoimmune diseases due to AI support, there are very few AI-driven drug development projects launched in the already relatively quiet field of infectious disease drug development such as tuberculosis, malaria, and Nipah virus.

　　However, while AI-driven antiviral drug development is a niche area, it is a necessity. Currently, innovative drugs in the public health field...Drug development faces numerous challenges, including R&D investment far exceeding the "double ten cycle" (meaning that the average R&D cycle for an innovative drug from initiation to final market launch exceeds 10 years and the R&D cost exceeds $1 billion), greater uncertainty in commercial returns, scarcity of targets, easy mutation and drug resistance of viruses, scarcity and poor predictability of animal models, and high challenges in drug toxicity and safety. Therefore, according to industry insiders interviewed, the empowering effect of AI large models on this track may be more significant.

　　How AI generates "promising" new drug molecules

　　"From the perspective of the AI ​​drug development platform itself, it can be used to design AI macromolecular models for infectious disease drugs, as well as in the research and development of oncology, neurodegenerative diseases, and rare diseases. The key lies in the priority of investment," said Guo Jinjiang, head of the Data Science Department at the Global Health Drug Discovery Center (GHDDI), in an exclusive interview with CBN.

GHDDI is an independently operated, non-profit innovative drug research and development institution 　　jointly established by the Beijing Municipal Government, Tsinghua University, and the Gates Foundation . Recently, GHDDI officially released "AI Kongming," an AI platform independently developed in China that integrates generative AI molecular design models, high-precision virtual screening, and multi-task ADMET evaluation models, pioneering intelligent design across the entire drug development process.

　　Guo Jinjiang stated that the platform will focus on new drug development in key global health areas such as malaria, tuberculosis, and viruses. However, in terms of platform capabilities, "AI Kongming" has completed system verification for dozens of real R&D pipelines, which include both infectious diseases and common and rare non-infectious diseases. The hit rate and optimization efficiency of the designed candidate molecules have been improved by several times to dozens of times compared to traditional processes.

　　He gave two examples: the first is tuberculosis, an infectious disease that infects approximately 350 million people nationwide and causes about 800,000 new cases each year. "AI Kongming" generated and verified a series of novel chemical structures targeting multiple key proteins. Experimental results showed that the hit rate of biologically active compounds was approximately 38%. Based on this, optimized lead compounds exhibited high cellular and enzymatic activity, demonstrating the potential value of this AI model in guiding the synthesis of new drug molecules.

　　The second case is amyotrophic lateral sclerosis (ALS), also known as Lou Gehrig's disease, a non-infectious and rare disease. Guo Jinjiang stated that "AI Kongming" designed 10 novel compound structures targeting SARM1, a key target protein of the disease. SARM1 acts as a "switch" responsible for the disconnection and breakage of nerve fibers in the human body, and its disruption directly affects the progression of ALS.

　　"In fact, whether it's invasive damage to the human body caused by foreign organisms or dysfunction of the body's own functions and structures, from the perspective of the underlying mechanisms of disease, it can all be attributed to abnormalities in protein structure and their interactions. The protein structures of different species have certain similarities in functionality, structure, and quantity, which provides a common biological basis for cross-species and cross-disease mechanism research and intervention strategies." According to him, the protein database pre-trained by "AI Kongming" currently reaches the millions and has a "natural advantage" in accumulating protein structure data of tuberculosis, malaria, and other viral pathogens.

　　According to Guo Jinjiang, although the "AI Kongming" platform still lags behind some purely commercial AI drug development platforms in terms of data and computing power, it is sufficient to support scientific research on the development of new drugs for major infectious diseases. "Currently, some large multinational pharmaceutical companies (MNCs) and global research teams are using the 'AI Kongming' platform or have expressed their willingness to cooperate."

　　MNCs and leading global research teams do not lack self-built AI platforms. However, they have two main expectations for third-party AI-driven drug development platforms in the field of antiviral research: First, infectious diseases are sporadic, and viral characteristics are often complex, resulting in a limited number of viruses available for research, potential drug targets, and available research data. This necessitates collaboration among governments, businesses, and research institutions worldwide, which raises concerns about data security.Sharing and intellectual propertyThere are also issues related to protection. At this point, a third-party platform independent of the pharmaceutical company's self-built AI platform is needed. This third-party platform can support both local and private cloud deployment to ensure the compliant, controllable, and secure transfer of core scientific research data. Secondly, they often need to rely on third-party AI pharmaceutical platforms to achieve multiple goals, from target structure analysis, molecule generation, activity prediction to drugability assessment, and ultimately generate drug molecules with "thousand-mile horse" potential.

　　The latter expectation is realized through three modules: Repeating Crossbow, Astrology, and Eight Trigrams Formation, which correspond to molecular generation, virtual screening, and drugability assessment, respectively.

　　“In the molecular generation stage, we use a ‘repeated crossbow’ approach to selectively generate compounds with potential biological activity and novel structures, based on the characteristics of a validated target for a specific disease. This results in a smaller number of lead compounds, a higher hit rate, and a lower duplication rate, which greatly improves the efficiency of subsequent screening and validation,” said Guo Jinjiang.

　　In their past experience with AI-driven drug development, Guo Jinjiang and his team discovered that some virtual or commercial molecule production platforms can generate 9 to 15 billion candidate molecules. This approach is essentially closer to the indiscriminate enumeration and expansion of chemical space. In this scenario, even with high-performance computing, a rough molecule screening process might only take tens of seconds. However, considering the need to multiply this by billions or even tens of billions of molecules, the overall cost of virtual screening is enormous.

　　Drug safety is also a major challenge in the development of new drugs for infectious diseases. In the past, many candidate drug compounds have been eliminated in the drugability evaluation or have failed in subsequent animal trials and human clinical trials.

　　Taking antimalarial drugs as an example, Guo Jinjiang said that some candidate drug compounds may have good biological activity against malaria parasites, but after entering the human body, they will cause biotoxicity and metabolic problems, especially when the target group of such drugs is women and children.

　　“If we rely solely on expert experience to modify molecules to improve the biotoxicity of drug compounds, there are many directions, and the process is very long, potentially taking 2-3 years. This is where the intervention of AI comes in,” he said. He explained that through the “Eight Formations” module of the AI ​​Kongming, the center’s researchers conducted toxicity optimization on lead compounds for antimalarial drugs. Within about six months, they obtained three lead compounds that performed well in safety tests. By May of last year, all related work had been completed, significantly accelerating the project’s progress toward the next milestone.

　　The Next Step in AI-Driven Antiviral Drug Development

　　The efficiency of AI-assisted development of new drugs for infectious diseases can be further improved.

　　Guo Jinjiang believes that when discussing AI-driven drug development, it's not just about "AI for drugs"—generating "high-potential molecules" through AI platforms—but also about "AI for scientists"—using AI platforms to create "talent scouts." This means systematically integrating the knowledge of experts in biology, medicinal chemistry, and pharmacokinetics into AI platforms. This allows them to collaboratively utilize various tools for molecule generation, property prediction, and evaluation, automatically generating and analyzing massive amounts of data. The results are then "organized" and "translated" into information and decision-making suggestions that experts can quickly understand and use, thus supporting researchers in timely optimization of AI-generated, rationally designed, or existing drug candidate molecules.

　　"Especially for some small research teams in the field of global infectious disease research and development, their number of experts and coverage areas are limited. They especially need technology platforms to empower them with 'AI for Scientists' capabilities. Otherwise, if they lack effective integration, interpretation and decision support mechanisms when faced with AI generating more valuable data in a shorter time, it will not only be difficult to maximize the improvement of research and development efficiency, but may even backfire," said Guo Jinjiang.

　　To date, AI-driven drug development in the field of infectious diseases, similar to other drug development areas, has primarily focused on the preclinical stage. According to a report by the Boston Consulting Group, AI-generated drug molecules have a success rate of 80%–90% in Phase I clinical trials, far exceeding the traditional 50%; however, the overall success rate for new drug development remains in the range of 9%–18%.

　　In the preclinical stage, the efficiency of AI-enabled processes may also be limited. "For example, an AI platform can complete the toxicity assessment of a candidate drug molecule in one or two hours, but subsequent biological testing of the candidate drug is required, such as zebrafish or mouse experiments, which may take up to four or five months," said Guo Jinjiang.

　　This means that AI accelerating new drug development does not equate to capital necessarily being able to "exit faster" or necessarily reducing the probability of research failure.

　　However, Guo Jinjiang holds a positive view on the future of AI-driven drug development. He cited two reasons:

　　First, drugs designed by mature AI drug development platforms can have certain "foresight" advantages in terms of drug toxicity, efficacy, and selection of indication populations, thereby reducing the risk of failure in subsequent preclinical and clinical stages.

　　Secondly, as many countries promote the reduction of animal testing, coupled with the increasing cost of test animals and the limitations of animal models in predicting the efficacy and toxicity of some antiviral drugs in humans, AI is expected to further reduce the reliance on animal models in the preclinical and non-clinical testing stages of new drugs, forming a new model for evaluating drug efficacy and safety based on computational prediction.