AI Drug Discovery: An Engineering Revolution from 'Alchemy' to 'Drugmaking'

目录

“We’re not doing ‘alchemy’; we’re ‘making drugs.’”—That line from Dr. Liu Qingsong, founder of Weiyuan Synthesis, instantly energized this technologist in me. A 300-million-yuan funding round, an AI biocomputing platform, a high-throughput experimental closed loop… behind these keywords lies a profound transformation in biopharmaceuticals, from a “handicraft workshop” to an “intelligent factory.”

Background: why did AI drug discovery suddenly “heat up”?

As a veteran who’s slogged through more than a decade in the internet industry, I’ve seen too many tech hype cycles. From cloud computing to big data to AI, every wave claims it will “disrupt” some traditional industry. But honestly, many of the so-called disruptions eventually turn into “empowerment” or “optimization,” far from a true revolution.

But AI drug discovery is different. This time, technology might be gnawing on the hardest bone of all—life sciences.

First, a few hard truths:

1. The “double-ten rule” curse: Developing a new drug takes an average of 10 years and 1 billion dollars, yet the success rate is under 10%. That efficiency is slower than internet-era product iteration by more than one order of magnitude.
2. “Needle in a haystack” screening: Behind a single target may lie millions or even hundreds of millions of candidate molecules. The traditional experimental method is like using a fishing net to catch one specific needle in the Pacific Ocean.
3. Extremely high cost of failure: 90% of candidate drugs fail in the clinical stage, meaning huge amounts of prior time and money go down the drain.

So when a company like Weiyuan Synthesis charges into this field armed with AI, automated lab robots, and high-throughput screening, its target isn’t simple efficiency gains—it wants to fundamentally reconstruct the scientific paradigm of drug discovery, shifting from experience-based trial-and-error to data- and algorithm-based prediction and design.

This is no longer “optimizing the process,” but “rewriting the rules.” Anyone who’s built enterprise systems knows that a change at this level often means enormous opportunity, accompanied by enormous technical challenges.

Technical breakdown: what is the “three-layer architecture” of an AI biocomputing platform?

Reading a press release, it’s easy to get excited, but we technologists need to peel it open and see how it’s actually built. Based on public information and my understanding of such systems, the technical core of a mature AI-driven drug R&D platform can usually be abstracted into a three-layer architecture: the data layer, the algorithm layer, and the execution layer.

graph TD
    A[AI biocomputing platform core architecture] --> B[Data layer: multimodal biological data]
    A --> C[Algorithm layer: prediction and generation models]
    A --> D[Execution layer: automated experimental closed loop]
    
    B --> B1[Genome/proteome data]
    B --> B2[Compound libraries and molecular structures]
    B --> B3[Historical high-throughput experiment data]
    
    C --> C1[Property prediction models<br>（ADMET etc.）]
    C --> C2[Molecular generation models<br>（GAN/diffusion models）]
    C --> C3[Protein structure prediction<br>（AlphaFold-like）]
    C --> C4[Target-disease association analysis]
    
    D --> D1[Automated synthesis robots]
    D --> D2[High-throughput screening platforms]
    D --> D3[Automated experimental data collection]
    
    B3 -.-> C1
    C1 -.-> C2
    C2 -.-> D1
    D3 -.-> B3

Layer one: the data layer—from “data wasteland” to a “high-quality fuel pool”
This is the cornerstone of every AI application, but in biology it’s especially hard. The difficulty isn’t that there’s little data, but that the data is “dirty,” “scattered,” and “biased.”

• “Dirty”: Biological experiment data is noisy, and reproducibility is a challenge.
• “Scattered”: Data is strewn across literature, patents, and companies’ internal databases, in wildly varied formats.
• “Biased”: Public data is concentrated on certain popular targets, creating a serious selection bias.

The “molecule entity libraries ranging from tens of thousands to hundreds of millions” and “high-throughput experiment data” that Weiyuan Synthesis mentions are exactly about consciously building its own high-quality, structured, closed-loop private data assets. This is far more valuable than simply calling public databases. From an engineering standpoint, this is equivalent to self-building a precisely annotated, continuously growing training dataset—the moat for model performance.

Layer two: the algorithm layer—from “prediction” to “creation”
This is the most dazzling part, and the story the media loves most. But we need to see through to the essence:

1. Property prediction models (ADMET, etc.): This is a relatively mature area. Using models like graph neural networks (GNN) and Transformers, it learns the relationship between molecular structure (graph data) and its absorption, distribution, metabolism, excretion, and toxicity properties. In essence, it’s a complex, high-dimensional regression/classification problem. The key lies in feature engineering and how to fuse multi-source data.
2. Molecular generation models: This is the current focus. From the early VAE and GAN to today’s hot diffusion models (similar to DALL-E, Stable Diffusion), the goal is to de novo design molecules with specific properties. Its technical essence is performing controllable sampling in a continuous or discrete representation of chemical space. For example, using a conditional diffusion model, with “inhibitory activity against target X > IC50 10nM” as the condition, to generate molecular structures that meet the criteria.
3. Protein structure prediction and design: AlphaFold2 has already shown AI’s astonishing ability in protein structure prediction. The next step is “protein design”—designing entirely new proteins for a specific function (like binding a certain molecule). This is equivalent to upgrading from “recognizing images” to “creating things.”

Pseudocode illustrating a simplified molecular generation loop:

# Simplified AI-driven molecular design loop
while not found_qualified_molecule:
    # 1. Generate candidate molecules
    candidate_molecules = diffusion_model.generate(
        condition="inhibit target A & Lipinski's rule of five"
    )
    
    # 2. Virtual screening (fast, low cost)
    for mol in candidate_molecules:
        admet_pred = admet_model.predict(mol)
        activity_pred = activity_model.predict(mol, target="A")
        if passes_virtual_screen(admet_pred, activity_pred):
            promising_mols.append(mol)
    
    # 3. Physical experimental validation (high cost, high value)
    for mol in promising_mols[:top_k]: # select the Top K for synthesis and testing
        # Executed by the automated synthesis platform
        synthesized, experimental_data = automated_lab.synthesize_and_test(mol)
        if experimental_data.confirms_prediction():
            hit_molecule = mol
            break
    
    # 4. Feedback learning: feed experimental data back to the model
    model.finetune_with_experimental_data(experimental_data)

Layer three: the execution layer—the bridge from “bits” to “atoms”
This is the key that distinguishes AI drug discovery from internet AI. An algorithm generating a perfect molecule on a computer is meaningless. You have to synthesize it and validate its effect. This is the “wet lab” capability.

• Automated synthesis robots: Chemical synthesis steps are encoded and executed by robotic arms, enabling 7x24 nonstop work and greatly improving synthesis throughput and reproducibility.
• High-throughput screening platforms: Processing thousands of experiments in parallel, rapidly acquiring biological activity data.
• The key closed loop: Experimental data is collected automatically and flows back to the data layer to optimize the algorithmic models. This “design-synthesize-test-learn” (DEL) loop is an amplifier of platform capability.

What Weiyuan Synthesis emphasizes with “full-stack” and “closed loop” refers exactly to the seamless connection of these three layers of capability. Miss any one layer, and it becomes “armchair strategizing” or “blind men feeling an elephant.”

My cool-headed take: the “three icebergs” beneath the boom

Capital is hot, technology is dazzling, but as a veteran who’s been through the ups and downs of technology cycles, I think it’s necessary to pour a few “buckets of cold water” and do some cool-headed thinking. Ahead of AI drug discovery lie at least three “icebergs”:

Iceberg one: the “quality” and “quantity” dilemma of data
The complexity of biological systems far exceeds images and text. Many current AI models are trained on limited, biased data, and their “extrapolation” ability is questionable. A generation model that performs well on known drug-like molecules may be unable to explore truly novel chemical space. More thorny still: what often determines a drug’s success or failure is low-frequency, long-tail side-effect data (like hepatotoxicity), which is extremely hard to obtain in the preclinical stage. Without high-quality, high-coverage negative data, the “ceiling” of AI prediction will be very obvious. This isn’t like a recommendation system, where click-through data can be obtained in real time at scale.

Iceberg two: the “interpretability” black box of algorithms
Drug regulatory agencies (like the FDA and NMPA) won’t easily approve a drug designed by a “black-box model.” “Why does this molecule work? What’s the side-effect mechanism?” AI may not give a scientifically logical explanation. This is not only a regulatory barrier, but also limits scientists from distilling and transferring knowledge from AI discoveries. The current trend is to develop “explainable AI” (XAI) and causal inference models, but that road is still long. From an engineering standpoint, we may need to design “human-machine collaborative” hybrid systems, letting AI handle massive search and pattern discovery while scientists make key decisions and validate mechanisms.

Iceberg three: the “valley of death” of commercialization
From “discovering a candidate molecule” to “successfully launching a new drug,” there lies a long road and enormous capital consumption through clinical phases I, II, and III. AI can significantly boost efficiency in the preclinical stage, but whether it can reduce the clinical failure rate remains unknown. Many AI drug-discovery companies may ultimately fall into an awkward position: becoming an efficient “candidate molecule supplier” rather than a true “pharmaceutical company.” Is its business model selling software, selling services, or bearing the risk of R&D itself? This requires enormous strategic resolve and capital endurance. 300 million yuan is a lot, but for the money-burning pharmaceutical industry, it may just be the “ticket to enter.”

So my view is: AI drug discovery is at a critical stage of transition from the “technology validation period” to the “value realization period.” It has already proven its power in the “discovery” link, but hasn’t yet fully traversed the whole process from “discovery” to “drug.” The next stage of competition will be a comprehensive contest of full-stack capability, closed-loop data quality, and clinical advancement capability.

What this implies for product builders: the “three-layer thinking” of hard-tech products

Setting aside the pharma industry, the practice of a company like Weiyuan Synthesis offers valuable lessons to everyone working on complex systems and hard-tech products (including my own experience building enterprise RPA and AI systems). I call it the “three-layer thinking of hard-tech products”:

1. Data-layer thinking: don’t just build algorithm models—build a “data flywheel”
Internet products emphasize the “user growth flywheel,” while hard-tech products must build a “data flywheel.” Every time your product is used (an experiment is run), it should generate high-quality, structured data that feeds back into the core model, making the product smarter and thus attracting more use—a positive cycle. When designing a product, you must figure out the data collection, annotation, and feedback loop clearly. Just like when we built RPA: every automation flow’s run logs and exception data were precious fuel for optimizing the scheduling algorithm and exception-handling models.

2. Algorithm-layer thinking: combine “80% automation” with “20% expert intervention”
Full automation is often unrealistic in complex domains. Excellent product design isn’t about fully replacing people, but using AI to handle the repetitive, time-consuming “dirty, hard work” (like screening out 99% of invalid molecules), freeing up human wisdom and energy to focus on the links that most need creativity and judgment (like analyzing the mechanism of top candidate molecules, or designing experiments). Product interfaces and interactions need to be deeply optimized for this human-machine collaboration. For example, providing clear visual analysis, highlighting key conflicting data, and guiding experts to make efficient decisions.

3. Execution-layer thinking: value “end-to-end” delivery and reliability
For To B, To Science hard-tech products, what customers buy isn’t a dazzling algorithm demo, but a reliable solution that produces actual results. This means you must go deep into the customer’s business processes and solve every link from “input” to “output.” In AI drug discovery, that’s the full delivery “from target to active molecule.” In enterprise systems, that’s full-lifecycle management “from requirement to deployment and ops.” Reliability, repeatability, and ease of integration are often more important than a few percentage points of pure algorithmic accuracy.

Conclusion

Every marathon runner knows that the hardest part isn’t the excitement at the start, but the “hitting the wall” phase around kilometer 30. This AI drug-discovery marathon may now be on the eve of the “wall,” transitioning from the excitement of technical breakthroughs toward commercialization and clinical validation.

Weiyuan Synthesis securing a huge funding round and launching an open platform is a strong signal: the industry’s leading players are shifting from single-point technical breakthroughs to building ecosystem and platform capabilities. An open collaboration platform aims to pool more data, algorithms, and needs, accelerating that all-important “data flywheel” and “application closed loop.”

For us technologists, the charm of this field is that it requires us to step out of the purely virtual world of code, to understand and connect the real physical world and the laws of life. It challenges not only our algorithmic ability, but our systems-engineering ability, our cross-disciplinary understanding, and our patience in the face of extreme uncertainty.

Perhaps our generation of technologists has the chance to witness how the computational power of the silicon-based world can systematically crack the health code of carbon-based life. This road is destined to be long and hard, but just like a long run, every solid step forward brings us a little closer to the finish line.

This engineering revolution from “alchemy” to “drugmaking” has only just begun.

中英对照

文章作者：阿文

文章链接： https://www.awen.me/post/5086b843.html

版权声明：本博客所有文章除特别声明外，均采用 CC BY-NC-SA 4.0 许可协议。转载请注明来自 阿文的博客！