In this series, we have been reviewing the drastic changes in the research and development environment for drug creation since around 2010. In the previous issue, it was stated that “pharmaceutical companies are increasingly exploring business models that not only sell ‘products’ like drugs but also provide ‘solutions’ for health.”1)
Pharmaceutical companies have already begun their efforts for self-reformation to adapt to such an environment. Meanwhile, many ICT companies are trying to enter the next-generation healthcare market. Instances of collaboration between the two are increasing. In these efforts, there is growing anticipation for Artificial Intelligence (AI) and Machine Learning (ML). On top of this, the Internet is bringing another major wave of change.
This essay aims to explore the impact of the new waves of the Internet and Artificial Intelligence on drug development and next-generation healthcare, and how to respond to these changes.
New Waves Related to ICT
Let us begin with a brief explanation of the so-called “Second Internet Revolution” and Artificial Intelligence. The “Internet” refers to a distributed wide-area network based on packet communication, initially built in the late 1960s in the United States for national defense purposes. It was originally a closed environment, limited to government-related research institutions and universities. However, during the Clinton administration (around 1994), the Internet was opened to the public2), leading to explosive growth. In a short time, vibrant and innovative new companies and services sprang up, marking the “First Revolution.”3)
Subsequently, mobile phones evolved into smartphones, essentially “palm-sized computers with communication capabilities.” This, combined with the widespread adoption of high-capacity communication lines, caused an exponential expansion of Internet services for consumers, with a corresponding increase in users. The spread of easy-to-use tablet PCs and affordable cloud services accessible via the Internet also contributed to this growth. Moreover, advancements in devices that surpass human sensory organs—such as audio, speech, and video sensors—combined with declining costs, have given rise to new Internet-based businesses and dramatically increased user populations. This ongoing transformation represents the “Second Internet Revolution.”
However, change continues to accelerate, and another massive wave is overlapping the Second Revolution. One symbol of this shift is the Internet of Things (IoT), which combines the digital information world with the physical manufacturing world. Another symbol is Artificial Intelligence (AI). The convergence of these three massive waves marks the “Third Internet Revolution,” which is expected to gain momentum over the next 10 to 15 years. This revolution is predicted to bring significant transformations to manufacturing and social services. A key application symbolizing this change is autonomous vehicles. For instance, the market capitalization of Tesla Motors surpassed that of GM this April, despite Tesla’s annual sales being only 100,000 vehicles. Another area of anticipated transformation is healthcare, exemplified by advancements in cancer treatment development and dementia prevention.
Artificial Intelligence and Pattern Recognition
Let us explore what “Artificial Intelligence (AI),” a major technological trend riding the new wave of the Internet, actually is. Today, AI has become a fascinating, ambitious, and future-oriented topic widely covered by mass media, touching areas such as drug development, health and medical care, caregiving service revolutions, work revolutions, and industrial transformations.4)
In simple terms, the essence of the current AI revolution lies in the digitization across various fields, the advancements in algorithms to utilize the data generated, and the improvements in computational environments that execute these algorithms.5) Consequently, its impact on health, medicine, and drug development is broad, and its future potential is highly anticipated. However, its true nature is often difficult for outsiders to grasp. One reason for this is companies promoting mere automation as “using AI” to ride the trend, and the media’s role in perpetuating this narrative. To clarify, let’s briefly introduce the history of AI in the healthcare domain.
Figure 1: Starting point of data handling by computers: mapping observed values of subjects into points (vectors) in multidimensional (linear) space. This is a 3-dimensional example, representing the basis of data science.
Research now called AI began in the 1950s when practical computers became available for scientific research. Shortly thereafter, philosophical debates emerged about what machines (computers) could do compared to humans.
In the 1960s, research into pattern recognition began, focusing on applications such as medical diagnosis, character recognition, and satellite image analysis. This involved collecting data derived from observations of target subjects and creating logic for decision-making (Figure 1).
Naturally, various multivariate analysis techniques based on statistics and probability theory were tested. These included Principal Component Analysis, Perceptron, Linear Discriminant Analysis, and methods like Dead Zone Maximization (DZM), now known as Support Vector Machines (SVM), as well as more general clustering techniques. These techniques are still in use today. I, too, was involved in this research while at Professor Satoru Watanabe’s lab at the University of Hawaii.
By the early 1970s, young researchers in computing and medicine in the U.S. began developing computer systems to formalize human experiential knowledge into logical forms that machines could execute. This was part of AI and was called Expert Systems—a term coined by C.A. Kulikowski, who had moved from the Watanabe lab to Rutgers University. After returning to Japan in 1971 and joining Hitachi’s research lab, my team developed a differential diagnosis system that expressed expert knowledge of heart disease specialists in rule-based formats. This system treated knowledge as data and used it directly for judgment, marking the early stage of knowledge systems (Figure 2).
Figure 2: Equivalent representations of diagnostic logic. When data is insufficient, human experiential judgments are logically represented. These methods form the basic model of expert systems.
A major challenge in developing medical diagnostic systems is the lack of sufficient data. My approach involved utilizing the experiential knowledge of clinicians to first create a framework for diagnostic logic. This framework was then implemented in experimental systems in clinical settings to collect real-world data and refine judgment rules6). This approach could also be applied to systems diagnosing cells and tissues based on pathologists’ experiential judgment rules.
The work of young researchers in the U.S. gained attention through Japan’s Fifth Generation Computer Systems project by the Ministry of International Trade and Industry (MITI, now METI), becoming widely recognized. As a result, Expert Systems became a highlight of AI in the early 1980s, with high expectations for applications in medical diagnostics. However, by the late 1980s, interest shifted to neural networks inspired by biological neural circuits. Notable applications included DECtalk, which read printed English aloud, and image recognition.
By the 1980s, AI had attracted significant attention and investment. However, many projects failed to achieve their ambitious goals, leading to a “winter” period for AI researchers. In recent years, however, AI has gained unprecedented attention even in general media, marking what could be called the “third AI boom.”7) The emergence of IBM’s Watson, known for intelligent responses, and deep learning as a pattern recognition technique, triggered this resurgence.
IBM’s Watson
The high-performance computer experimentally developed by IBM, known as Deep Blue (named after IBM’s corporate color), defeated the reigning world chess champion in May 1997. The company then invested heavily in developing a computer called Watson, which, in 2011, defeated human champions on the television quiz show *Jeopardy!* This system possesses advanced cognitive functions, including natural language processing for interacting with humans, storage of domain-specific documents, encyclopedic knowledge and data, generation of candidate answers to questions, estimation of the likelihood of multiple answers, and integrated reasoning capabilities.8)
Today, parts of Watson’s functionality appear to be accessible to external users, but initially, researchers criticized the lack of disclosure on key aspects. Nonetheless, mass media often portray computers that excel in clearly defined tasks—such as chess, Go, or quiz games with explicit rules and win-loss criteria—as possessing intelligence comparable to a master strategist or physician. IBM, perhaps recognizing this media tendency, crafted announcements that gave the impression Watson was a versatile and intelligent general-purpose response system. However, IBM’s current strategy appears to focus on collaborating with clients to develop AI systems tailored to their specific needs, rather than selling prebuilt AI systems. This approach seems successful in attracting well-funded customers.
Deep Learning
Deep learning is a computational model for pattern recognition that builds upon the neural networks popularized in the 1980s (Figure 3).
Figure 3: Basic model of deep learning. Intermediate layers may vary, as do the number of input lines and connection strengths. This model aligns with representations in Figures 1 and 2.
This model was introduced in 2006 by Geoffrey Hinton and colleagues at the University of Toronto9). Its effectiveness in applied experiments was recognized, leading to its use in numerous applications, such as image analysis10). Unlike IBM Watson, which is a large-scale and expensive computational system, Hinton’s deep learning model is an algorithm that general researchers can experiment with within their environments and budgets. As a result, deep learning has been widely tested by various research groups, leading to the emergence of numerous new models inspired by its principles.
A notable milestone came in 2012, when researchers from Google presented at a conference their deep learning model aimed at “unsupervised learning” of images11). They claimed to have successfully enabled a computer to “recognize the concept of a cat” by analyzing numerous photos. Subsequently, Google’s AlphaGo system defeated the world’s top Go players, and in Japan, systems threatening human champions in shogi have been developed. Today, deep learning models, broadly defined, have virtually become synonymous with artificial intelligence.
The Impact of Artificial Intelligence and Machine Learning on Drug Development
Currently, new AI technologies such as Watson and deep learning are attracting attention from pharmaceutical companies, hospitals, medical research institutes, and life insurance companies. ICT companies are also attempting to enter this field.
Deep learning models, in particular, have made headlines for their impact on drug development. This was highlighted in a comprehensive report on Hinton’s deep learning10), referring to the work of Ma et al., who participated in the QSAR competition (Kaggle competition) sponsored by Merck in 2012, later detailed in their publication12).
In pharmaceutical research and development, a wealth of computational techniques, databases, and knowledge bases related to chemistry and biomedicine has already been established13-15). Since the 1970s, popular techniques have included molecular calculations fundamental to chemistry, molecular graphics capturing molecules in 3D, QSAR (Quantitative Structure Activity Relation) techniques for predicting relationships between compound structures and their biological activity, and expert systems for synthetic route exploration leveraging chemists’ heuristics. These techniques primarily fall within the domain of Chem(o)informatics. While machine learning has already been applied in this field16), recent years have seen increasing use of deep learning17).
Since the 1980s, advancements in molecular biology have revolutionized medicine, leading to the accumulation of knowledge about genes, genomes, and pathways. Consequently, drug development has shifted to start with elucidating the molecular mechanisms of diseases and identifying the targets (primarily proteins) within the body on which drugs act. In this digitized biomedical field, bioinformatics has become an indispensable tool. The acceleration of this trend occurred after the successful completion of the Human Genome Project at the beginning of this century, with deep learning techniques increasingly applied in this area as well18).
The ultimate goal of leveraging ICT in pharmaceutical R&D is to enhance productivity. This includes the introduction of robotics and data handling systems like LIMS (Laboratory Information Management System) in laboratories, along with Chem-Bio Informatics techniques, which have been familiar for nearly 40 years. Now, these areas are poised for further innovation through IoT and new AI-based computational techniques.
However, having access to advanced AI techniques does not guarantee the development of effective drugs. Researchers must possess deep domain-specific expertise, experience, and insight to fully utilize these techniques. This necessity is well known in QSAR studies, docking studies involving precise molecular calculations, and other domains. In this sense, future AI will remain a tool to assist experts (Figure 4).
Figure 4: Scientific thinking combines universal theories and techniques with specific theories and techniques.
On the other hand, using systems like IBM Watson (Cognitive Computing) to accumulate vast amounts of literature and databases could significantly improve the efficiency of various tasks in the drug development pipeline by helping researchers gather and organize information and knowledge. However, this would require a major overhaul of the foundational work environment, potentially necessitating a review of the entire operations of a research institute or company. Such a transformation would involve substantial costs, and if the system is not designed for continuous development, it could quickly become outdated.
Next-Generation Healthcare and Artificial Intelligence
The new wave of artificial intelligence is spreading beyond drug development to encompass all of healthcare19). Last year, Otsuka Pharmaceutical and IBM announced the establishment of a joint venture. This venture aims to develop a solution called “MENTAT” by leveraging Otsuka’s accumulated knowledge in the central nervous system field and IBM’s technology from developing Watson. The first implementation of this collaboration is taking place at Okazama Hospital.
The core of this project lies in the automatic organization of vast medical records, particularly integrating and analyzing descriptive data such as symptoms and medical histories, which are difficult to quantify. The goal is to create a database that can be effectively utilized by healthcare professionals. Additionally, Dai-ichi Life Insurance has announced a study in collaboration with Fujita Health University to analyze medical records of chronic disease patients, such as those with diabetes, using IBM’s Watson to benefit healthcare and insurance services.
These initiatives aim to improve productivity for healthcare service providers. On the other hand, from the perspective of healthcare service recipients, the primary goal of ICT utilization is to provide better services at a lower burden. Specifically, it means enabling better treatment options for patients under limited resources at reduced costs.
There are two challenges to leveraging ICT in this way. The first is formulating the goals as a mathematical model and solving them. The second is collecting and implementing real-world data to apply the findings. These challenges are interconnected, as the data to be collected depends on the formulation. ICT holds the key to addressing both challenges.
The discipline involved in this formulation is control theory20). Richard Bellman, known for developing dynamic programming, famously stated, “All medicine is control theory!” The difference between control theory and pattern recognition or AI lies in the former’s intrinsic connection to human values—seeking optimal or near-optimal solutions, which makes it impossible to execute entirely mechanically. This is fundamentally tied to the essence of medicine. Optimal treatment depends on individual values, linking the concept of optimal care to personalized medicine.
These concepts are most realistically applied in pharmacological treatment. For example, PGx (Pharmacogenomics) examines patient responses to drugs through genome analysis to avoid harmful reactions and use only effective medications. However, the challenge extends beyond selecting the right drugs to ensuring proper administration. Recent studies show that individual responses to drugs vary based on factors such as environment, diet, gut microbiota, and circadian rhythms. Thus, if pharmaceutical companies truly aim to provide solutions, they must consider the circumstances in which drugs are used and strive to offer background knowledge for their proper use.
Ultimately, the responsibility for proper drug use lies with physicians, although it is only part of their duties. Physicians generally face more complex situations requiring judgment and decision-making. The Mayo Clinic in the U.S. is developing Decision Support Systems (DSS) to help healthcare providers make better decisions21). DSS was a topic of expert system research in the 1980s, and now the field is being revisited on a larger scale.
Another notable area of research is cancer care, where the U.S. is taking the lead. The American Society of Clinical Oncology (ASCO) is working as a society to collect cancer treatment records and explore better care through a project called CancerLinQ22). Initially, IBM’s Watson was considered for the system’s foundation, but ultimately, SAP’s HANA platform was chosen.
A key consideration in such projects is the rapid expansion and evolution of foundational knowledge, which is expected given advances in cancer research. Determining what constitutes appropriate care must also evolve rapidly. This is the concept behind the Rapid Learning Health Care System, which has become a guiding principle for next-generation healthcare (Figure 5). The Mayo Clinic’s system embodies this concept.
Figure 5: Development of research infrastructure for generating knowledge from data in healthcare.
Challenges in Artificial Intelligence Research in Healthcare
In Japan, IBM’s Watson was experimentally introduced at the University of Tokyo’s Institute of Medical Science (Hospital) in 2015, where it reportedly provided useful insights to specialists. At the end of last year, the National Cancer Center, the National Institute of Advanced Industrial Science and Technology (AIST) Artificial Intelligence Research Center, and Preferred Networks Inc. announced the development of an AI-powered cancer care system. Similarly, Kyoto University Hospital and Fujitsu began research on AI applications in cancer care, and the Cancer Institute of the Japanese Foundation for Cancer Research, along with FRONTEO Inc., announced the development of AI-based “cancer precision medicine.”
These ambitious initiatives in Japan to apply AI to cancer care are highly promising. However, significant challenges remain. Collecting clinical data is inherently difficult, and ethical, social, and legal restrictions on data handling are becoming increasingly stringent. Additionally, the infrastructure for medical terminology thesauruses in Japanese is considered insufficient. Systems also require built-in mechanisms for continuous updates and rapid learning, but securing ongoing funding for such efforts is expected to be quite challenging.
Belatedly, the Ministry of Health, Labour and Welfare established the “Council for Promoting AI Utilization in Health and Medical Fields” in January of this year to support AI adoption. However, substantial foundational efforts will be necessary to overcome the unknown challenges that lie ahead.
Conclusion
In this discussion, we have looked at the past and present of artificial intelligence, focusing on its applications in drug development and next-generation healthcare. The highlighted advancements, Watson and deep learning, represent just a fraction of the broader developments in this field. However, the astonishing reality of the new wave of AI lies not in novel theories or models but in advancements in sensors and computational systems. What remains a concern is the lack of opportunities and training for young talent in ICT, especially in D2K (data to knowledge) science. Furthermore, creating workplace environments where these individuals can collaborate as equal partners with experts in drug development and healthcare is absolutely essential but remains underappreciated. Another urgent issue is improving the (comparatively inadequate) computational environments for AI research to maintain international competitiveness.
Looking slightly further into the future, it seems that models like deep learning are increasingly being aligned with the circuits of the human brain and nervous system. This alignment is fostering more collaborative research between wet-lab researchers, clinicians, and computational researchers23). Such teams are particularly expected to yield results in studies on cognitive impairments. In an even more distant future, connections with quantum information and quantum computing are anticipated24),25). This integration will lead to a world where natural sciences, information science, and engineering are seamlessly intertwined (Figure 6). Now is the time to bolster educational support for nurturing young talent who will pave the way for such a future.
While the discussion continues to broaden, the first step in addressing the challenges of the Second Internet Revolution and the new wave of artificial intelligence is to increase future-oriented dialogue among experts from a wide range of fields.
Figure 6: AI is the study of aligning thought processes with computation. Its foundation lies in mathematical or computational techniques shared across various academic disciplines. The theories and computational techniques shown in the figure are universally applicable.
References
Tsuguchika Kaminuma., Transforming Drug Development: Towards 2020, *Sōyaku no Hiroba*, 3:3-7, 2016.
*The Internet Unleashed*, Sams Publishing, 1994.
Tsuguchika Kaminuma., *The Third Opening: The Impact of the Internet*, Kinokuniya, 1994.
Yamagiwa, D., *Artificial Intelligence and Industry/Society: How to Survive the Fourth Industrial Revolution*, Japan Economic Research Institute, 2015.
Kushida, K., *The Algorithm Revolution from Silicon Valley: A Shocking Perspective*, Asahi Shimbun Publishing, 2016.
Tsuguchika Kaminuma., translated by Kurashina, S., *Computer-Based Medical Consultations*, Bunkodo, 1981. (Original: E.H. Shortliffe, *Computer-Based Medical Consultations: MYCIN*, Elsevier, 1976; see translator’s notes)
Amari, T., *Brain, Mind, and Artificial Intelligence*, Kodansha, 2016.
Ferrucci, D., Watson: Beyond Jeopardy!, *Artificial Intelligence*, 199–200:93–105, 2013.
Hinton, G. E., Osindero, S., and Teh, Y.-W., A Fast Learning Algorithm for Deep Belief Nets, *Neural Computation*, 18:1527–1554, 2006.
LeCun, Y., Bengio, Y., and Hinton, G., Deep Learning, *Nature*, 521:436–444, 2015.
Le, Q. V., et al., Building High-Level Features Using Large Scale Unsupervised Learning, *Proceedings of the 29th International Conference on Machine Learning*, Edinburgh, UK, 2012.
Ma, J., et al., Deep Neural Nets as a Method for Quantitative Structure-Activity Relationships, *J. Chem. Inf. Model.*, 55:263–274, 2015.
Csermelya, P., et al., Structure and Dynamics of Molecular Networks: A Novel Paradigm of Drug Discovery; A Comprehensive Review, *Pharmacology & Therapeutics*, 138:333–408, 2013.
Todeschini, R., et al., Data Analysis in Chemistry and Biomedical Sciences, *Int. J. Mol. Sci.*, 17(12):E2105, 2016.
Tetko, I. V., et al., BIGCHEM: Challenges and Opportunities for Big Data Analysis in Chemistry, *Mol. Inform.*, 35(11–12):615–621, 2016.
Cheng, F., et al., Machine Learning-Based Prediction of Drug–Drug Interactions by Integrating Drug Phenotypic, Therapeutic, Chemical, and Genomic Properties, *J. Am. Med. Inform. Assoc.*, 21:e278–e286, 2014.
Gawehn, E., et al., Deep Learning in Drug Discovery, *Mol. Inform.*, 35(1):3–14, 2016.
Angermuller, C., et al., Deep Learning for Computational Biology, *Molecular Systems Biology*, 12:878, 2016.
Ravì, D., et al., Deep Learning for Health Informatics, *IEEE J. Biomed. Health Inform.*, 21(1):4–21, 2017.
Tsuguchika Kaminuma., Application of Control Theory to Medicine (*Advances in Life Science*, Japan Medical Association, Vol. 5, Shunjusha, 1978), pp. 26–46.
Kaggal, V. C., et al., Toward a Learning Healthcare System – Knowledge Delivery at the Point of Care Empowered by Big Data and NLP, *Biomed. Inform. Insights*, 8(Suppl 1):13–21, 2016.
Miller, R. S., CancerLinQ Update, *Journal of Oncology Practice*, 12(10), 2016.
RIKEN Center for Brain Science (CBS) (ed.), *Connected Neuroscience*, Kodansha, 2016.
Nishimori, H., and Ozeki, M., *Quantum Computers Accelerate Artificial Intelligence*, Nikkei BP, 2016.
Schuld, M., Sinayskiy, I., and Petruccione, F., An Introduction to Quantum Machine Learning, *Physical Review A*, 94:022342, 2016.
PROFILE
Tsuguchika Kaminuma
Born in Kanagawa Prefecture, Japan, in 1940. Educated at International Christian University, Yale University, and the University of Hawaii. Received a Ph.D. in Physics. Since 1971, has worked at Hitachi Information Systems Research Institute, the Tokyo Metropolitan Institute of Medical Science, and the National Institute of Health Sciences. Conducted research in pattern recognition, medical artificial intelligence, medical information systems, bioinformatics, and chemical safety. In 1981, founded an industry-government-academia research exchange organization (now CBI Society) aimed at theoretical drug design. Later engaged in interdisciplinary human resource development at Hiroshima University and Tokyo Medical and Dental University. Established the NPO Cyber Bond Research Institute in 2011.
