A New Paradigm for AI Drug Discovery: From Incremental Change to Qualitative Leap | BlueRun Ventures Biotech Salon Highlights

Recently, BlueRun Ventures hosted an online Bio2X salon series on biology themed "AI's Paradigm Revolution in Drug Discovery." Dr. Xi Bie, Vice President at BlueRun Ventures, spoke with Dr. Yue Song, Chief AI Scientist at BioMap, and Dr. Yue Peng, Co-founder and CEO of ReviR Therapeutics, discussing the challenges and opportunities in AI-driven pharmaceutical research.

Key takeaways:

• In traditional pharma, AI and project teams operated in silos; the new paradigm requires direct integration between the two.
• Data volumes today far exceed what was previously available, making this the moment for AI to truly deliver value. Batch data is becoming more accessible within organizations.
• Unifying data across pharma companies remains difficult, but progress is being made. More critical still is identifying data that actually matters.
• Biotech company value must be demonstrated through data. Large pharma firms are beginning to share relevant datasets and invite innovative biotechs to participate — an entry point and gradual convergence for both sides.
• The defining feature of CADD is physics-based modeling, while AI starts from molecular-level protein and nucleic acid sequences for computation and prediction. A trend is emerging toward fusing these two approaches.

Below is an edited transcript of the conversation. Stay tuned for more BlueRun Ventures Bio2X salons, and follow the participating companies:

Q: What are the old and new paradigms of AI in drug discovery?

Peng Yue: There's broad consensus that AI represents a major trend in pharmaceutical development. Yet for traditional pharma companies, embracing AI has been a top-down process. Constrained by project-team organizational structures, AI applications haven't been tightly integrated with core drug development workflows. Meanwhile, many innovative AI pharma startups and the MNCs they serve also haven't achieved close integration. At this stage, everyone is still exploring in various directions.

Song Yue: From practice, in academia, researchers from different backgrounds have different interests. When AI algorithm-driven discoveries need to be communicated to biology labs, the feedback loop is very long, and it's rare to iteratively tackle a single problem together through downstream experiments. In fact, only by truly integrating AI technology with laboratory validation — running continuous, rapid iteration — can we genuinely impact target discovery and drug design. What attracted me most to BioMap was its ability to integrate both aspects well and concentrate resources toward a single goal.

Q: What can AI company collaborations with pharma achieve? From a corporate perspective, must all stages be kept in-house to generate meaningful data?

Song Yue: From an AI application standpoint, using AI for target discovery and drug design means searching for specific signals in an extremely high-dimensional space. The required data volume typically scales exponentially with dimensionality. Using machine-driven active learning, generating data through multiple rounds of small batches, is more practical than acquiring massive datasets all at once. Additionally, AI-predicted targets or optimized drug molecules often carry uncertainty due to data limitations and model imperfections, requiring wet-lab experiments for data collection and validation. Feeding this data back into AI models improves predictive confidence. This is why companies pursue dry-wet closed loops — only through sustained AI utilization and model iteration can effective target information ultimately be found.

Peng Yue: In pharma, target discovery isn't simply about finding a strong signal in some genetic or big dataset. It's more about target validation — using extensive data to prove a target is viable. If an AI company stops at providing data analysis without generating further data for target validation, the target's value remains limited. Only validated targets justify pharma investment.

From CADD onward, computational-assisted drug discovery has existed for years. Was the failure to deliver expected value due to algorithmic limitations or organizational structures in pharma?

Song Yue: AI models aren't universal solutions. Typically one model optimizes for a single metric — drug toxicity, solubility, etc. For instance, I might predict small molecule-RNA binding affinity and screen a large pool down to a smaller set. But each small molecule has multiple druggability properties beyond binding affinity that the AI didn't predict. Predicting all properties would require multiple models, each with validation steps. Building so many models demands correspondingly more data, and potentially more complex data types.

Peng Yue: The problems between dry and wet experiments that existed in the past persist today. But new opportunities have emerged. One is data — volumes are now vastly larger, with better methods for generating high-quality data, making AI adoption imperative. AI platform software has also emerged, saving considerable time. But ultimately, AI's impact on drug discovery hinges on dry-wet integration.

Q: How should we evaluate the role of active learning and similar algorithms in AI drug discovery?

Previously, Exscientia partnered with Sumitomo Dainippon Pharma to develop DSP-1181, a long-acting 5-HT1A receptor agonist that entered Phase I clinical trials in Japan for obsessive-compulsive disorder (OCD). Exscientia claimed this candidate was the world's first fully AI-designed drug to reach clinical trials, using active learning AI algorithms in the discovery process.

Song Yue: AI models using active learning technology fill gaps in model knowledge and continuously update, achieving better predictive performance with less data. When evaluating affinity between drug molecules and target proteins, AI's generalization capabilities can build models from candidate compounds in a library — advancing beyond traditional screening methods.

Peng Yue: AI companies often boast about speed improvements, but this isn't pharma's primary concern. What pharma cares about most is whether good drugs can be found — whether AI can solve problems that traditional methods cannot.

Q: How can small companies cold-start? Is public data useful? And how can they source proprietary data?

Song Yue: Public data is increasingly available but also complex in format. It's more suitable for model pre-training. Solving specific problems still requires substantial specific data — an advantage and moat for large pharma or companies with wet-lab capabilities.

Peng Yue: Finding truly novel insights from public data alone is somewhat difficult. But if you can identify a very unique angle for framing and analyzing questions, there's still significant value.

Q: What does data management and sharing look like inside pharma companies? Can software and AI help utilize data more effectively?

Peng Yue: The harder challenge in pharma is actually data management. There's encouragement for cross-project data sharing now, but it's a long road. Especially after data is organized, which portions can actually be utilized? Additionally, some US companies acquire and sell clinical data from hospitals, but in my experience this data has limited value — it wasn't collected for specific projects, so while volumes appear large, extracting meaningful information proves difficult.

Song Yue: Of course, with high-quality data like PDB, we can solve problems like protein folding well, though that's a somewhat special case.

Q: What are the opportunities and challenges when capable new AI companies partner with large pharma?

Peng Yue: For pharma companies, data matters most. From a certain perspective, biotech company value must also be demonstrated through data.

Song Yue: Some large pharma partnerships with AI drug discovery companies begin by releasing datasets with clear benchmarking objectives. The pharma company wants to see if your model improves performance on these datasets before internally advocating for more specific, precise data collaboration. This is a common testing pattern pharma uses with AI innovators, and a way both sides gradually move closer.

Q: If building a wet-lab platform matched to AI computing power, how should one think about investment costs and lab scale?

Peng Yue: For cell-based experiments, data generation isn't terribly expensive — this may not be a limiting factor. For drug screening, it depends on the method for generating training data. If it's proof-based combined with some AI computation, that might be a good approach. Costs vary by method.

Song Yue: Designing AI models to experiment continuously with existing data to see AI improvement also requires human and time costs. Alternatively, one can generate a new batch of data. Generally, more data improves AI models, but data collection is very expensive, and the resulting improvement is often uncertain. Where the optimal balance lies depends on the specific problem.

Q: How does AI differ from traditional CADD? What are its advantages and specific characteristics?

Song Yue: CADD actually spans a broad category. Its most important hallmark is physics-based models — energy functions describing molecular interactions, minimizing or optimizing these functions to design or discover drugs, or even using dynamic information from these energy functions to observe drug effects. AI differs in that it bypasses this intermediate physics process, predicting directly from data. So industry now sees a trend toward fusing CADD and AIDD approaches — introducing more physics knowledge into AI models to improve their design.

Q: What are potential applications of AI in large-molecule drug development?

Song Yue: AI technology has even greater prospects in large-molecule drug applications. Large molecules are more complex, with vastly larger design spaces than small molecules. When designing large molecules, one rarely creates entirely new ones — often working from existing biological macromolecule scaffolds to design specific protein sequence segments. This requires integrating existing knowledge, understanding that structure determines function.

Peng Yue: From pharma experience, antibody discovery itself hasn't generated much AI interest, but optimization after discovery is labor-intensive. From this perspective, it's quite interesting to see if AI can guide this process.

Further Reading

BlueRun April Report | Growing and Resilient Amid Uncertainty

What Does a Career Defined on Your Own Terms Look Like? | BlueRun Talent Philosophy

Year of the Tiger New Beginnings | BlueRun Bio2X Winter Camp Successfully Concludes

BlueRun Ventures was founded in Silicon Valley in 1998. BlueRun Ventures China was established in 2005 as a venture capital firm focused on early-stage startups.

Currently, BlueRun Ventures China manages multiple USD and RMB dual-currency funds with over RMB 10 billion in assets under management. It focuses on Pre-A and Series A investments across hard tech and innovative interaction, enterprise technology, new consumer, and healthcare. It has invested in over 150 startups including Li Auto, Waterdrop, QingCloud, Guazi.com, Qudian, Songguo Mobility, Ganji.com, Energy Monster, Yuntion Semiconductor, Machenike, CloudSaints, Anxin NetShield, and BioMap.

BlueRun Ventures has been ranked #1 in Zero2IPO's "Top 30 Early-Stage Investment Institutions in China," #1 in ChinaVenture's "Top 30 Best Early-Stage Venture Capital Institutions in China," and Top 10 globally among VC fund managers with consistently high returns by Preqin.

Additionally, BlueRun Ventures has received consecutive recognition from Forbes China, 36Kr, Cyzone, Caixin Media, CBNweekly, Jiemian, and other media as "China's Best Early-Stage Firm of the Year," "China's Top Venture Capital Firm," "Most Entrepreneurur-Friendly Early-Stage Firm of the Year," and "Most Influential Early-Stage Firm of the Year."