Why Implementing AI for Science Feels More Rewarding – Insights from Prof. Hong Liang

HyperAI interviewed Prof. Hong Liang, a distinguished professor at Shanghai Jiao Tong University, to explore the development of AI in protein engineering and the broader challenges and pathways for AI for Science (AI4S) to achieve practical impact.

Decisive Pivot: Earning an AI Degree on Bilibili

Prof. Hong notes that AI for Science has surged in the past two years, moving from small‑scale tool experiments to a key driver that can surpass human experts in many research tasks. However, only a few AI solutions have truly left the laboratory.

He cautions against “the emperor’s new clothes” mentality and stresses the need to think about deployment. He recounts how his team successfully applied a general‑purpose AI model for protein engineering to over 20 companies, seeing designed molecules produced in 5,000‑liter fermenters, which gave him a profound sense of achievement.

His group has built the proprietary “Pro” series large models for protein engineering, delivering the world’s first and second large‑model‑designed protein products that have been industrialized.

AI4S: AI Must Respect Science, Science Must Learn AI

Prof. Hong emphasizes that AI for Science should start by defining scientific or engineering problems before proposing AI solutions. He cites DeepMind’s multidisciplinary team as a model, where experts from traditional science and data/computer science collaborate to close the loop from problem definition to AI method.

His own team mirrors this approach, recruiting both computer‑science students and protein‑engineering researchers. CS members often lead feasibility discussions for AI‑driven optimizations, while domain scientists articulate concrete scientific challenges for AI to address.

He advises researchers to self‑study AI, especially team leaders, because transitioning to AI4S resembles a strategic corporate transformation that requires clear project proposals and understanding of new technologies.

While AI can accelerate many stages of drug discovery, Prof. Hong points out that the closed‑loop time remains long, with in‑vitro predictions often poorly correlated with animal studies and clinical outcomes.

Consequently, his team focuses on enzyme function, where molecular experiments provide rapid, verifiable results that can quickly benefit industries such as food, cosmetics, textiles, and biomedicine.

He warns against indiscriminately applying large models to domains lacking standardized, structured input‑output data, urging researchers to choose problems where AI can meaningfully reduce experimental cost and leverage deep scientific understanding.

Conclusion

Throughout the interview, Prof. Hong repeatedly mentions the role of luck, foresight, execution, and willingness to experiment in achieving AI4S breakthroughs. He hopes that more scientists will boldly explore AI‑driven research directions, turning innovative ideas into tangible, industry‑ready outcomes.