Soochow: AI drug discovery moves from proof of concept to commercialization; focus on AI platform companies over the medium to long term.
Currently, there is still no drug fully discovered or designed by AI that has received full regulatory approval globally. Once such a drug is approved, the valuation center of the entire sector is expected to shift upward.
Soochow released a research report stating: Short term: position in the upstream "shovel sellers." The massive number of candidate molecules generated by AI must all go through processes such as protein expression and preclinical wet-lab validation. Suppliers in upstream segments such as recombinant proteins, reagents, and culture media are expected to gradually realize orders and performance. Medium- to long-term focus: 1) AI platform companies. This may become one of the directions with the greatest elasticity in the future, mainly tracking BD transactions and clinical data of platform companies. 2) The approval and launch of AI drugs is expected to bring industry value re-rating. Currently, there is still no drug globally that has been entirely discovered or designed by AI and has received full regulatory approval. Once this happens, the valuation center of the entire sector is expected to move upward.
Soochow's main views are as follows:
AI drug discovery breaks through the constraints of the traditional pharmaceutical industry's "double-ten law"
In the traditional pharmaceutical industry, the average cost from initiating R&D to marketing for an innovative drug exceeds US$1 billion, and the R&D time exceeds 10 years. AI technology, by integrating multi-dimensional data such as literature, genomics, proteomics, and clinical medicine, empowers areas including target discovery, molecule design, and clinical trial optimization. According to a BCG 2024 report, the success rate of AI-generated drug molecules in Phase I clinical trials reached 80%-90%, far higher than the historical industry average of 40%-65%. AI is expected to help significantly increase the overall success rate of new drug R&D.
The AI drug discovery technology stack has completed a fundamental reconstruction from single algorithm tools to a multimodal technology matrix
There are currently two main types of mainstream models: 1) Diffusion models: currently the main models used in molecule design. They reconstruct samples by gradually adding noise and then learning to reverse it. This route has solved the key limitations of early generative methods such as unrealistic bond lengths and unstable conformations, and is one of the directions with the most intensive industrial investment at present. 2) Transformer: a general foundation from protein structure to clinical design. With the self-attention mechanism at its core, it efficiently processes long sequences in parallel. Diffusion models and Transformers possess the compounding effect of pretrained large models plus generalization capability, and are one of the few routes that can accumulate into platform-type barriers; while CNN/GNN/RNN, as mature tools, find it difficult to form differentiated competition.
Technological evolution is jointly driven by three underlying logics
1) Data availability: The fundamental reason why the AlphaFold series achieved breakthroughs first is that protein structure prediction is a data-rich field. In contrast, real efficacy data of drugs in the human body (such as clinical trial results) is very scarce, and annotation is highly subjective (efficacy evaluation standards vary). 2) Maturity of computing paradigms: Pure statistical learning faces two bottlenecks: first, out-of-distribution generalization; second, the causal confusion problem. The integration of physical mechanisms and AI large models alleviates these two problems. 3) Industrial validation closed-loop capability: The biggest early doubt about AI drug discovery was the disconnect between dry experiments (computational prediction) and wet experiments (laboratory validation). Under the traditional model, AI models rely on public datasets for training, with uneven data quality and no ability to iteratively optimize; under the closed-loop model, the data generated in each round of experiments (whether successful or failed) becomes training material for the model, and model performance continuously improves as the number of experiments increases.
Market space and competitive landscape
According to Global Market Insights, the global AI drug discovery market was about US$3.1 billion in 2025 and is expected to reach US$43.9 billion by 2035, with a CAGR of 30.5%. Global AI drug discovery companies can be divided into three categories: 1) AI technology giants, such as Google, Anthropic, and others, building large models + computing power + global ecosystems, deeply bound with MNCs; 2) leading AI Biotechs, whose core advantages lie in clinical validation data + proprietary AI platforms + depth of BD cooperation; 3) MNCs' self-developed AI platforms, whose core advantages lie in proprietary data accumulation + clinical resources + global commercialization capability.
Risk warnings: R&D and trial progress falling short of expectations; commercialization falling short of expectations; intensifying industry competition.
Related Articles

HK Stock Market Move | AI hardware chain hit hard by Nvidia personnel changes; Victory Giant Technology (02476) falls over 9%, Zhongji Innolight (03308) drops nearly 8%.

Jensen Huang return rumors heat up, NVIDIA Corporation (NVDA.US) futures contracts rise over 3%

HK Stock Market Move | SHEIN-W (00625) rises over 5%; recently announced an additional 500 million investment to expand its Poland logistics hub.
HK Stock Market Move | AI hardware chain hit hard by Nvidia personnel changes; Victory Giant Technology (02476) falls over 9%, Zhongji Innolight (03308) drops nearly 8%.

Jensen Huang return rumors heat up, NVIDIA Corporation (NVDA.US) futures contracts rise over 3%

HK Stock Market Move | SHEIN-W (00625) rises over 5%; recently announced an additional 500 million investment to expand its Poland logistics hub.






