AI drug discovery stands at the crossroads of bubble and validation: domestic financing in the first nine months of 2026 reached 15.765 billion yuan, exceeding the total of the previous three years

Main text

Some investors also said that they have been building positions this year in AI4LS (AI-enabled life sciences) related targets. AI4LS is moving from single-point prediction to a new stage of multi-task collaboration; concepts such as virtual cells and digital life are bringing drug efficacy prediction into a completely new paradigm, and fields such as AI protein design, AI synthetic peptides, and AI nucleic acid drugs are also worth watching. But the entire track is too hot, and the current atmosphere feels as if it is recreating the scene in 2021 when the valuations of some biotech companies rose two or three times in a single year.

Heat in the track does not equal maturity. Data dilemmas, the ceiling on understanding biological causality, and validation difficulties at the clinical end remain exams that all participants must face directly.

Tech giants enter

The anti-Moore's law in the new drug field has limited the continuous decline in new drug output. The reasons behind this are: the easy-to-mine dividends of basic scientific research have been exhausted, and finding molecules superior to existing drugs has become more difficult; regulators are cautious, and pipeline R&D investment is enormous; at the same time, popular targets are crowded together, and the supply of potentially non-consensus targets is insufficient, further suppressing new drug output.

In the second half of last year, Huawei's pharmaceutical team was established.

In June this year, Anthropic, the US AI large-model giant, officially released the AI scientific research workbench Claude Science. It is not an entirely new foundation model, but a multi-agent system built on Claude Opus 4.8, integrating more than 60 scientific databases, preconfiguring toolchains in fields such as genomics, proteomics, and cheminformatics, and emphasizing that results are traceable and reproducible.

In September this year, Anew Labs, an AI drug discovery company under ByteDance, was pushed to the capital market as an independent company.

Fan Jie, president of Huawei's pharmaceutical team, told Yicai reporters that the positioning of Huawei's pharmaceutical team is as an AI enabler for the pharmaceutical industry, rather than competing with others to make drugs. Huawei comes from the computer and communications field, not biology or chemistry, so it focuses on leveraging its own strengths.

The entry of tech giants has intensified industry competition.

In his view, investors will have very different evaluation criteria for different AI drug discovery companies. For companies such as Insilico Medicine that have already entered a mature stage, the evaluation criteria are often how many innovative targets they have, how many projects have entered clinical trials, and whether the business model can be implemented. For startup AI drug discovery companies, what matters is still the team, technology, data, algorithms, and so on.

"They are still in the sunrise period where they only talk about technology, and no one demands that they implement it," Ren Feng said. But in the end they will also have to enter the second half and will be evaluated on their ability to commercialize.

CVSource data from CVINFO shows that in the first nine months of 2026, 57 investments were completed in the domestic AI drug discovery track, 44 companies received financing, and the disclosed total financing amount reached 15.765 billion yuan, exceeding the combined total of 2023, 2024, and 2025. But while the financing amount surged, the number of investment deals and invested companies did not grow in sync. The four largest investment amounts totaled about 9.5 billion yuan, accounting for 60% of the total financing in the entire track.

"The current AI drug discovery track has passed the early concept stage, and pipeline data are the most intuitive confirmation: industry pipelines grew from 4 in 2017 to about 179, of which 9 have entered phase III clinical trials," Wei Yaguang, a partner at Guofang Venture Capital, said at the conference. Resources and capital are clearly concentrating toward the top, and many targets have significant valuation premiums. According to calculations by the company's internal model, the valuation premium range of targets it has reviewed is roughly 40% to 180%.

"At the pharmaceutical level, such large computing power cannot be used. The hardest thing right now is data: data silos, uneven data distribution, poor quality, and too much dirty data," Wei Yaguang said.

Pharmaceutical companies' experimental data carry intellectual property and they are unwilling to share it externally. Public datasets are merely a drop in the ocean relative to the vast biological search space. Regarding data, Ren Feng said their approach is to use open-source data to help them build models, while using feedback and optimization from internal experimental data to continuously optimize and feed back into the models during the experimentation process.

Liu Wei, co-founder and CEO of BioMap, told Yicai reporters that they insist on building the underlying "operating system" for life sciences and positioning themselves as an industry enabler. They do not rely on pipelines for monetization; instead, they build their own laboratories and create a closed loop integrating dry and wet lab work, using the continuous production of high-quality biological data to bridge the gap between the investment cycle of large models and the pace of new drug R&D.

Liu Wei recalled that early AI drug companies mostly focused on small molecules: chemical space was relatively controllable and synthesis processes were mature, with AI playing more of an efficiency tool role, and the R&D chain of pharmaceutical companies did not undergo fundamental change. But de novo design of entirely new sequences for thousands of amino acids is a completely different track, helping to develop FIC (first-in-class) drugs, but the accompanying immunogenicity and safety risks are higher, forcing the industry to move safety evaluation much earlier, and the logic of pharmaceutical companies' R&D, decision-making, and regulatory communication faces being reshaped.

Wei Yaguang believes that the next main battlefield for AI drug discovery is mining causality at the biological level, which is the core of whether the industry can break through the ceiling. This year may become the first year of China's virtual cells, and the industry is exploring: can AI identify and learn biological causal laws beyond humanity's existing experience, breaking through the bottleneck of biological cognition? If this layer cannot achieve a breakthrough, the industry's historical lessons must be faced squarely: valuation bubbles objectively exist, and optimistic market expectations cannot continue forever.

“Virtual cells are the next sea of stars.” Liu Wei believes that the role of virtual cells in target discovery, drug function research, safety evaluation, and early clinical prediction has only just begun to emerge; current AI accuracy on cellular problems is still not high, and only when cell experiment technologies enter a golden period in the coming years can virtual cells move from usable to accurate.

“Avoid the narrative line”

China’s innovation capability and scale have become the second largest in global pharmaceutical innovation, but a gap remains compared with the United States.

At the conference, Song Ruilin, chief expert of the China Pharmaceutical Innovation and Research Development Association, provided a set of data: as of July 2024, China had 1,775 First-in-class R&D pipelines, accounting for 19% of the global share, while the United States had 4,511, accounting for 48%. Original innovation still has a long way to go. “AI-enabled pharmaceutical innovation is accelerating, and pharmaceutical innovation is precisely the industry that most needs AI and also has mutual demand with AI.”

As for the current stage of AI drug discovery, Fan Jie used the Gartner Hype Cycle to make a judgment: the industry is still in the “peak of inflated expectations,” where expectations are rising rapidly, and has not yet entered a stage of full maturity. But “overheated” does not mean the technology has no value; rather, market expectations may run ahead of actual implementation progress.

At present, some capabilities such as protein structure prediction, molecular dynamics simulation, and molecule generation and screening have gradually been applied in R&D. AI is also used in R&D document processing, intelligence analysis, patent analysis, and some pharmacokinetic and toxicological predictions. The problem is that these capabilities remain scattered across different stages.

“Many points are already mature, but the whole chain has not yet been connected.” Fan Jie said that from target discovery and validation to candidate molecule generation, wet-lab screening, and then to preclinical research and clinical development, drug R&D involves a series of interconnected judgments. Only by connecting model predictions, experimental validation, and result feedback can AI continuously iterate on real data.

Ren Feng said that bubbles in the current AI drug discovery industry inevitably exist. Any emerging track, especially a field like AI where technology iterates and expands at high speed, will certainly be accompanied by bubbles during development. The key is what companies rely on after the bubble recedes to stand firm and support their value. “For us, if the AI bubble recedes, what can truly support the company’s market value is only real product pipelines and Phase II and Phase III clinical projects. These are the hard-core assets.”

“Next year will be a crucial observation window for the AI drug discovery industry and a year of reconstructing the industry narrative. Data from some drugs’ Phase III pipelines will be revealed, and the market will reprice AI drug discovery: whether valuations hold, rise, or correct. When we screen AI drug discovery targets, we focus on four core points: whether they have proprietary data, whether they can achieve a closed dry-wet loop, whether they can produce priceable clinical evidence, and whether the business model chooses the path of self-developed pipelines. Finally, valuation must fall within the range the company’s option model can bear. We still want to earn the money of validation and avoid the narrative line,” Wei Yaguang said.

(Source: Yicai) Middle of the main text Embedded advertisement