AI drug discovery moves beyond outsourced collaboration: pharma companies such as MegaStream are embedding AI into the R&D backbone

Since the development of AI drug discovery began, the industry has gradually returned to rationality and started to calmly examine the real bottlenecks of this track: most current AI technological capabilities are concentrated in the early stages of drug discovery, and commercial value often stops before PCC (preclinical candidate compound) screening. Even if algorithms are used to complete massive target screening, candidate molecules with drug-forming potential remain scarce, and innovative molecules that can smoothly advance to clinical trials and have translational value are even rarer.

Thus, under the iteration of AIDD (artificial intelligence-driven drug discovery) technology, the R&D paradigms of some innovative pharmaceutical companies have gradually shifted: they no longer simply treat AI as a "molecular discovery accelerator," but instead build it into the core driving force for innovative pipeline R&D, thereby launching AI "large-scale infrastructure."

A co-founder and chief scientific officer of an AI pharmaceutical company recently mentioned during the 2026 International Biocomputing Innovation Summit that the traditional computational screening model has long been plagued by off-target effects, and even after completing screening of tens of thousands of targets, the number of candidate molecules that can enter subsequent development remains very limited.

Lei Ming, CEO of MegaStream, has a deep understanding of this. MegaStream was jointly founded by Harbour BioMed and BioMap. Before that, Lei Ming was the youngest chief analyst in the pharmaceutical industry at a foreign investment bank and also served as CFO of a Nasdaq-listed pharmaceutical company. He told Yicai that the capital market's enthusiasm for AI technology and the actual industrial value of AI drug discovery are not completely equivalent. From the perspective of the pharmaceutical industry, AI is indeed expected to improve the efficiency of early-stage drug R&D, but new drug R&D still has the characteristics of long cycles, large investment, and high risk of clinical failure, and early algorithm advantages or experimental results do not necessarily translate into clinical success.

In his view, what truly supports the long-term value of AI pharmaceutical companies is always verifiable core assets—competitive innovative drug pipelines, reproducible experimental validation results, clinical development progress, and commercialization outcomes such as BD collaborations.

The positioning and participation models of AI pharmaceutical companies in the industrial chain have already quietly changed. In the past, the industry generally focused on undertaking R&D outsourcing for innovative pharmaceutical companies; now more and more companies are actively exploring diversified monetization paths: on the one hand, recovering R&D investment through outbound licensing transactions; on the other hand, retaining high-quality projects in their own clinical pipelines, seeking a sustainable self-funding model.

However, candidate drugs obtained through AI screening still need to undergo lengthy human clinical trials. The further into late-stage clinical development, the greater the capital consumption and losses from failure, but the potential commercial value of the drug also rises simultaneously, yet it is difficult for AI pharmaceutical companies to "get a share" in the mid-to-late stages of R&D.

To lengthen the value chain, some AI pharmaceutical companies have begun building integrated automated experimentation and data infrastructure. For example, Recursion Pharmaceuticals in the United States, which has received investment from Nvidia, relies on automated platforms and massive biological datasets to mine drug targets. The company has reached a collaboration with Genentech, a subsidiary of Roche, using 30 million neural cell images and massive gene perturbation data to find neuroscience targets. In August 2026, the two parties officially announced that the neuroscience targets discovered by both parties had formally entered the early development stage of small-molecule drugs.

"We are more optimistic about companies that simultaneously possess core AI capabilities, proprietary data assets, mature drug R&D systems, and clinical translation capabilities. The value of AI ultimately needs to be reflected in whether it can develop more differentiated drugs, improve R&D success rates, or improve R&D efficiency, and cannot be measured solely by model parameters, algorithm metrics, or concept hype," Lei Ming said.

Lei Ming divides AIDD into three stages: in the 1.0 era, AI was more of an independent auxiliary tool embedded in the R&D systems of traditional pharmaceutical companies, achieving single-point efficiency improvements only in some segments, showing fragmented application characteristics; in the 2.0 era, AI-native companies focused on algorithms but easily overlooked the importance of wet experiments and biological understanding; the real turning point appears in 3.0—the AI-native era, whose core hallmark is that AI and biological research are deeply integrated from the very beginning of project initiation and are no longer separated from each other.

Feng Yulan, senior vice president and head of external cooperation at Harbour BioMed, said at the aforementioned summit that AIDD companies currently face two core dilemmas: first, ownership of intellectual property; second, ownership of data. In cooperation between AI companies and pharmaceutical companies, three types of outputs are generated: R&D assets, experimental data, and model capabilities. However, under the traditional project-based cooperation model, models and experimental data are difficult to effectively accumulate, and most cooperation is centered on only a few drug projects, unable to form systematic capabilities, and technical capabilities are difficult to internalize within the pharmaceutical company itself.

Therefore, a more ideal model is to integrate and accumulate models, data, and core technical capabilities into a unified entity. Taking MegaStream as an example, Lei Ming introduced that the two parties upgraded from project-based cooperation to jointly establishing an independent joint venture, with the core goal of coordinating model development, data use, experimental validation, and candidate drug advancement around long-term drug R&D goals, avoiding the disconnect between technology development and drug R&D. On the premise of clarifying data authorization, intellectual property, and compliance boundaries, the technical capabilities of both parties can achieve long-term synergy within a unified R&D system.

More fundamentally, some interviewed industry views hold that under the iteration of AIDD technology, the reason innovative pharmaceutical companies build their own AI drug R&D infrastructure is, on the one hand, to allow the biological data accumulated by pharmaceutical companies over many years to be deposited on their own, so that innovative R&D returns to the underlying causal mechanisms of targets and diseases; on the other hand, to form a complete closed loop of dry-lab computational simulation + wet-lab automated validation, extending the value boundary of AI from early PCC screening to the PCC-to-IND (Investigational New Drug application) stage.

Lei Ming further analyzed that for AI pharmaceutical companies, computing power and data are important foundations for the development of life science large models, but the real challenge lies in how to make models understand complex disease biology mechanisms and translate model predictions into drug molecules that have been experimentally validated and have clinical development potential. This requires specific models to match specific R&D tasks; the needs of pharmaceutical companies also go beyond molecule generation, and require comprehensive evaluation of affinity, specificity, epitope, stability, expression level, and drug-likeness, with verification one by one through continuous experiments.

Regarding the next step for AI drug discovery, Lei Ming proposed that the industry needs to build a "Tesla in the pharmaceutical field." Just as Tesla did not simply add software on top of traditional fuel vehicles, but deeply integrated software and automotive engineering from the product architecture and engineering system levels; the biopharmaceutical industry should also start from the foundation and design and build AI technology and biotechnology as a complete integrated system, rather than simply layering AI tools on the surface of traditional drug R&D processes.