When the topic of 'AI-driven drug discovery' arises, many envision a future where artificial intelligence significantly shortens the new drug development cycle and cuts costs. Indeed, in the capital market, companies involved in this field have drawn considerable investor attention in recent years.

However, the stark reality is that no drug developed solely through AI has been successfully launched to date. AI-driven drug discovery faces several insurmountable bottlenecks that cannot be easily resolved through mere capital infusion.

This report aims to dissect these bottlenecks in simple terms for everyone's reference.

1. First Hurdle: Data Scarcity - Messy and Insufficient Data

It is a well-known fact that AI relies heavily on data to learn from human experiences and make informed decisions.

In the context of AI-driven drug discovery, the necessary data includes precise information such as drug molecular structures, protein interaction mechanisms, and clinical trial results. The challenge lies in the fact that this data is often insufficient or inaccurate.

1.1 Data Shortages Due to Data Silos

Core data in the pharmaceutical industry is scattered across various sources, including hospital medical records, pharmaceutical company experimental data, and regulatory approval data. Much of this data remains isolated, as organizations, institutions, and units guard it as confidential information.

From an industry standpoint, this fragmentation prevents AI from comprehensively learning from the diverse data types within the pharmaceutical sector.

1.2 'Eating the Wrong Food' Due to Messy Data

In different laboratories, the instruments, reagents, and data recording methods can vary significantly.

This diversity is akin to having recipes written in different languages, some in Chinese, others in Spanish, and some using slang or dialects.

For AI, this poses a significant challenge as it struggles to interpret the data uniformly.

1.3 Data Incompleteness Due to Success Bias

Both academia and industry have a tendency to publish only successful experimental results, often sidelining failed data.

For instance, records indicating that a certain molecule is ineffective or a target has no impact are rarely made public.

This is comparable to only showing AI successful recipes without revealing which ingredient combinations lead to failure. Consequently, AI continuously recommends its learned 'successful formulas,' potentially leading it down the wrong path.

2. Second Hurdle: The Algorithmic 'Black Box' - Difficult to Explain and Hard to Pass Regulatory Approval

If data is the fuel for AI, then the algorithm is its brain. However, this brain has a significant flaw: no one can clearly explain how it thinks or reaches its conclusions. This is the infamous 'black box' problem.

In the pharmaceutical field, the 'black box' can be detrimental. To bring a drug to market, it is crucial to clearly explain to regulatory authorities why a particular molecule can treat a disease and how its safety is ensured.

Yet, when AI provides drug molecules or targets, it fails to explain specifically why it arrived at that conclusion. It merely states that the model calculated it that way, without offering a scientific and precise explanation of potential risks.

For example, AI might prioritize recommending a molecule with hidden toxicity due to a bias in the data, but humans cannot detect this because of the 'black box' issue.

This poses a significant problem for regulatory authorities. It's akin to grading an exam where the teacher cannot just look at the final answer but must also see the steps taken to arrive at it.

This 'lack of steps' issue prevents AI-driven drug discovery from passing regulatory approval.

Additionally, algorithms are prone to becoming outdated.

In drug development, factors such as patient population characteristics, disease definitions, and treatment guidelines may frequently change. These changes can cause AI models' predictive results to become increasingly inaccurate, a phenomenon known as 'data drift.'

For instance, if AI is trained on older patient data, when a mutated disease strain emerges, AI may struggle to provide accurate or up-to-date treatment plans, posing a significant risk to drug development.

3. Third Hurdle: Clinical Translation Falls Short - Large Gap Between Experimental Data and Clinical Beds

Even if AI identifies promising drug candidates in the laboratory, it does not guarantee success in reaching patients.

This is because clinical trials are essential.

Traditional clinical trials are inherently high-stakes elimination zones. While AI can enhance early-stage experimental success rates, the clinical stage presents a different set of challenges.

On one hand, AI struggles to accurately simulate the complex human body environment.

Data from cellular and animal experiments in the laboratory differs significantly from real human conditions.

Even if AI can calculate the binding mode of drug molecules and target proteins, it cannot confirm how the drug will be absorbed and metabolized in the human body or whether it will interact with other organs.

On the other hand, AI's auxiliary role in clinical trials is also constrained.

For example, patient recruitment.

AI needs to find suitable patients through precise profiling, but due to data non-interoperability, it is challenging to integrate patient information across hospitals, resulting in low recruitment efficiency.

Another example is data monitoring during clinical trials, which requires real-time processing of massive amounts of data and timely adjustment of protocols. This places extremely high demands on AI's computational power and real-time response capabilities, which current technology struggles to fully meet.

4. Fourth Hurdle: High Computational Costs - Unaffordable for Small and Medium-Sized Enterprises

AI-driven drug discovery is a costly endeavor. Beyond research and development investment, computational costs pose a significant burden. Processing vast amounts of molecular data and simulating complex biological reactions require immense computational power. For large pharmaceutical companies, this represents a substantial expense; for small and medium-sized enterprises (SMEs), it is an even higher barrier to entry.

This leads to the emergence of a 'Matthew effect': only a few large enterprises with financial resources and capabilities can sustain continuous investment. SMEs either struggle to maintain operations through financing or confine themselves to niche areas.

Innovation often stems from diverse exploration, but the computational cost barrier prevents many small and medium-sized teams from realizing their visions, indirectly slowing the progress of the entire industry.

5. Conclusion: AI-Driven Drug Discovery Is a Long Marathon

The potential of AI-driven drug discovery is undeniable. It holds promise for addressing unmet medical needs such as tumors, autoimmune diseases, and metabolic disorders, bringing new hope to the medical field. However, to fully realize this potential, several major hurdles—including but not limited to those mentioned above—must be overcome:

Some experts predict that 2030 may be a critical turning point for AI-driven drug discovery. However, drug development has always required patience and rigor. While AI may accelerate progress, it cannot skip the necessary 'hurdle-clearing' steps.

END