This year, cross-border deals in China’s innovative drug sector have accelerated, with artificial intelligence-driven firms emerging as key players. Syneron, Helixon, and XtalPi illustrate the trend, together closing business development (BD) transactions worth several billion USD between March and August.
Syneron and Helixon, founded just four to five years ago, distinguished themselves in multinational pharmaceutical firms’ competitive BD processes by using AI to reshape biologics R&D and improve success rates. The deals appear to have revived investor confidence, enabling other firms in the sector to secure new funding after years of stagnation.
The foundations of biopharmaceutical R&D are shifting from high-volume screening and intuition-based methods toward rational design and de novo creation. In the past decade, AI tools played a supporting role, helping to streamline parts of early discovery. Now, the ability to design drugs from the ground up marks a more fundamental change. The capacity to generate proteins not found in nature and address previously undruggable targets could open paths to treatments for diseases that have long resisted progress.
After reviewing protein design data released by Chai Discovery, backed by OpenAI, and by ByteDance, Wang Chengzhi, founder of Zhiyuan Shenlan, concluded that a qualitative leap is near. His company specializes in biomolecular design and manufacturing, and he has more than 20 years of experience in life sciences, including a stint as chief scientist at MegaRobo. He said:
“I wouldn’t have made such a judgment last year. But the advances in AI this year make me believe that disruptive events in biopharma may arrive sooner than anyone is prepared for.”
FreesFund partner Ma Rui shared a similar view:
“The momentum now is driven by innovative drugs multiplied by AI. While the boundaries of AI drug discovery are still unclear, there could be momentous changes in the next one to three years.”
In the previous cycle, FreesFund profited by leading XtalPi’s Series A round and backing Metis Pharmaceuticals at seed stage. Exiting XtalPi generated cash returns worth dozens of times its initial investment.
So how is generative AI rewriting the logic of drug R&D? What changes might it bring to biopharma, and how are current players approaching disruption?
From needle-in-a-haystack to precision design
AI’s role in drug discovery has not been free of overpromises. Between 2018–2021, a surge of AI-driven small molecule projects attracted hundreds of billions of RMB, but many struggled to deliver. Deep learning and virtual screening sped up discovery, yet the molecules identified often failed to outperform existing drugs or proved difficult to synthesize. This exposed the limits of early models.
The current wave builds on mainstream AI breakthroughs spilling into drug discovery, enzyme engineering, and related fields:
- AlphaFold2 (AF2) demonstrated that transformer-based architectures could solve the protein-folding problem that had eluded biologists for decades. Humanity identified about 200,000 protein structures over 60 years. Yet within two years of AF2, AI predicted more than 200 million with high accuracy, covering most known proteins.
- David Baker’s team then applied diffusion models, adapted from image generation, to biology. By iteratively “denoising” data, success rates in designing novel proteins have been increased by an order of magnitude.
- AlphaFold3 (AF3) advanced beyond predicting individual proteins to modeling interactions among proteins, nucleic acids, and small molecules. This all-atom approach improved generalization even when data was limited.
This year, multiple groups released next-generation models: Chai Discovery’s Chai-2, Evolutionary Scale’s ESM3, and ByteDance’s Protenix. All aim to generate new functional molecules from scratch.
“Chai-2’s latest results show that for specific targets, its antibody candidates have a much higher hit rate than traditional methods. Where once you needed to screen millions to billions of molecules to find a few positives, now you might find hits among just a few dozen sequences. That was unimaginable before,” Wang told 36Kr.
Traditional antibody development required immunizing animals and screening for high-affinity antibodies, a process that took about three years and USD 5 million. Chai-2 and similar models can now compress this to hours of computation and two weeks of wet-lab validation.
Ma of FreesFund believes this could redefine antibody development, replacing methods like hybridoma, phage display, and animal immunization with de novo computational design. “If breakthroughs extend to small molecules, then almost every drug modality could be AI-enabled,” he said.
In the future, when designing an antibody drug, researchers may turn first to AI models that generate and rank sequences, then synthesize and test only the most promising candidates. This shift could reverberate throughout the value chain.
With many straightforward targets already addressed, AI could revive stalled research areas by designing new leads. It could also help replace drugs with problematic side effects with safer alternatives.
Wang expects AI to sharply shorten preclinical timelines, particularly in oncology, autoimmune disorders, and metabolic diseases. Chronic conditions may see the earliest breakthroughs, with drugs similar in impact to semaglutide emerging more frequently. Conversely, screening platforms that rely heavily on animal testing may lose commercial value as AI tools advance.
In this division of labor, AI-powered biotech firms will likely lead early discovery, while pharmaceutical firms focus on clinical trials, regulatory approvals, and commercialization. Rewards will be shared through BD licensing and co-development deals.
With limited capital, where to focus first?
Each technological wave reshuffles industries, and investors are racing to back the most promising players. Three main archetypes are emerging:
- Tech giants with capital and compute power, including Google DeepMind, Meta, Xaira, and ByteDance, are building foundational bio-related models, setting standards, and supporting open-source ecosystems.
- Startups led by AI and computational biology experts are pushing new algorithms. Once refined, these models can serve pharma partners or support in-house pipelines. Examples include BioMap, Helixon, Insilico Medicine, MoleculeMind, and BioGeometry.
- Traditional drug developers are adopting AI tools without building foundational models. They rely on open-source AI and strong lab capabilities to accelerate programs where they already have expertise in targets and indications.
Ma said the key metric for evaluating players is their ability to understand and evolve models. While fine-tuned open-source models may achieve strong results, solving real-world discovery problems often requires breakthroughs at the algorithmic level.
Another expert noted that few firms can afford the cost of training foundational models. Compared with language models, biological data is more expensive to generate. One China-based AI firm, backed by a major tech company, reportedly spent tens of millions of RMB synthesizing and testing thousands of AI-generated antibodies, but still lacked the data volume to reach efficiency thresholds.
Wang agreed, saying that firms able to rapidly generate high-quality experimental data will have the strongest models:
“In the past, automation and high-throughput experiments were about improving screening efficiency. In the AI era, they are also about producing structured, iterative datasets that directly feed model training and optimization.”
Algorithms and data remain the twin pillars of AI drug discovery. Commercialization, however, converges on one goal: producing molecules with real therapeutic value and winning buyer recognition through BD deals.
“As things stand, every new drug company will use AI to some degree. For AI-driven pharma firms, building models alone isn’t enough. They must deliver viable molecules to earn value,” Ma said.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Hai Ruojing for 36Kr.