Why 'Monopoly' Global Pharmaceutical Companies Laugh and Cry Over Artificial Intelligence [Reading Science]

AI and Deep Learning Techniques Dramatically Reduce Drug Development Costs, Time, and Workforce
High-Cost, High-Profit Pharmaceutical Industry Paradigm Shaken
Multinational Giant Capital Monopoly → Opportunities for AI Startups to Enter
Challenges Remain in Interdisciplinary Integration, Protein Structure Prediction, and Data Accumulation

Reference photo. Not related to the article.

[Asia Economy Reporter Kim Bong-su] Artificial intelligence (AI), the most representative technology of the 4th Industrial Revolution, is changing the paradigm of new drug development. It has become a world where small-scale AI startups can lead the pharmaceutical industry, which was a typical high-risk, high-reward field with low success rates despite enormous investments of capital and time.

◇Traditional New Drug Development Takes an Average of 15 Years

According to a paper jointly published last December in the Journal of Korea Institute of Information and Communication Engineering (JKIICE) by Professors Jeong Myeong-hee and Kwon Won-hyun of Anyang University, new drug development largely involves stages such as discovering target proteins related to diseases, identifying hit and lead compounds, evaluating synthesis feasibility, efficacy, and toxicity (Scoring), and finally selecting the optimal candidate compounds before conducting preclinical and clinical trials. This rigorous verification process ensures that a specific drug is effective above a certain threshold for the disease and free from side effects or toxicity.

Accordingly, traditional new drug development typically takes about 15 years on average, with a very low success rate. According to the Korea Health Industry Development Institute, only one out of approximately 5,000 to 10,000 compounds ultimately succeeds. Specifically, it takes an average of 5 years just to select 10 to 250 compounds that enter preclinical trials. Then, narrowing down to about 10 compounds for clinical trials takes an additional average of 2 years, and conducting Phase 1, 2, and 3 trials to find one meaningful compound takes another 6 years on average. After that, it takes about 2 more years on average to obtain new drug marketing approval from the U.S. Food and Drug Administration (FDA). In the U.S., which sets the global standard, the total process takes 14 to 16 years and requires development costs of 2 to 3 trillion won. U.S. pharmaceutical companies have invested about 520 trillion won in new drug development over the past 15 years, which is five times the aviation industry and 2.5 times the computer industry. This is mainly why large multinational pharmaceutical companies based in the U.S. can form monopolistic structures.

◇AI Dramatically Reduces Costs and Time

AI using big data is revolutionizing the new drug development process by significantly reducing costs and time. Previously, discovering new drug candidates required enormous time and cost for information search and drug design. After selecting the target disease, researchers had to manually filter through 400 to 500 related papers to identify candidate compounds. Analyzing vast data such as patent information, compound structures and efficacy-related big data, medical data, and clinical data was essential. It is a complex task that requires simultaneous optimization of various factors including maximizing compound activity and efficacy while minimizing toxicity and side effects.

AI can greatly improve speed and efficiency through automation, reduce randomness and errors, and intelligently search and recognize patterns in vast data. For example, it can search over one million papers and one billion compounds at once. What would take dozens of researchers up to five years can be completed in a single day. The same applies to clinical trial stages. AI can calculate the binding ability between compound structures and in vivo proteins to suggest new drug candidates first. It also makes it easier to find clinical patient groups highly related to the disease being studied based on hospital medical records. By predicting genomic variations and drug interactions, AI can significantly reduce trial and error in clinical trial design and personalized drug development stages. Oh Du-byeong, head of the New Drug Division at the National Research Foundation of Korea, explained, "Currently, AI is most widely used to search for drug candidate compounds," adding, "A representative case is the rapid identification of candidate compounds through AI-based drug repositioning technology to discover COVID-19 treatments."

In other words, there are about 20,000 drugs currently on the market. Verifying each one’s effectiveness for a specific disease through clinical trials required substantial financial resources and time. However, AI was able to find candidate drugs effective against COVID-19 among these 20,000 drugs in just a few days. Particularly, the prediction of protein 3D structures, considered the biggest challenge in new drug development, has recently entered an innovative phase due to rapid AI advancements. A representative example is Google DeepMind’s protein structure prediction AI AlphaFold2, which released a database at the end of July predicting the structures of over 200 million proteins existing on Earth. Oh said, "If protein structures can be predicted, the drug’s effects can be anticipated and designed in advance," adding, "It can become an innovative tool for new drug development."

Public Artificial Intelligence Drug Development Platform KAIDD Established by Korea.

◇Challenges in Data Collection and Interdisciplinary Integration

The main bottleneck in new drug development has been understanding protein functions and structures. Proteins, along with DNA and RNA, are core substances of life phenomena. Thousands of amino acids fold in unique ways to form three-dimensional structures, and if these can be predicted and understood, it can bring innovation to understanding disease causes, symptoms, and drug development. AlphaFold2, trained with deep learning technology, has demonstrated results predicting protein structures with over 90% accuracy. It takes protein sequences and multiple sequence alignment (MSA) features as input, passes them through convolutional neural networks (CNN), and outputs distributions of amino acid distances and chemical bond angles. Using this, it creates an energy function that well represents protein structures and optimizes it repeatedly by gradient descent until the energy is minimized, thereby predicting the given protein’s 3D structure.

Interdisciplinary research integrating chemistry and biology is also urgently needed. Professors Jeong and Kwon pointed out in their paper, "While understanding of chemical systems has greatly advanced and quantitative success has been achieved, qualitative information and data reflecting biological relationships are necessary to lead to new drug discovery," adding, "For AI to change the landscape of new drug development and approval, more cooperation and integration among broad specialized fields such as chemistry, biology, toxicology, and pharmacokinetics are required."

To improve AI’s accuracy and safety, securing sufficient qualitative and quantitative data is also necessary. Oh said, "Although there are some cases of clinical success using AI-based new drug development in the U.S. and other countries, it is still in the verification and nascent stages," adding, "Accumulating abundant high-quality data is important to reach a paradigm shift."