"First Steps in AI Drug Development" is an introductory book that explains how artificial intelligence is changing the paradigm of the pharmaceutical and biotechnology industries.
Traditionally, developing new drugs takes more than a decade, is associated with enormous costs, and numerous failures. However, AI is gaining attention as an innovative tool that analyzes massive amounts of data to discover candidate substances, predict protein structures and drug actions, and increase clinical success rates.
This book covers the basic concepts of new drug development, the interactions between proteins, diseases, and drugs, and the process of drug discovery, optimization, and clinical trials.
It also connects deep learning principles and neural network structures such as CNN, RNN, and GNN, and even new drug design using generative AI, helping readers understand the overall flow in three dimensions.
This book serves as a guide covering both fundamentals and practical applications for students, researchers, and industry professionals, and serves as a starting point for AI-based pharmaceutical and biotechnology innovation.

I hope that through this book, readers will achieve two key goals.
First, it's about understanding the new drug development process, and second, it's about experiencing the potential of AI technology to revolutionize new drug development.
[...] In the middle part (Chapters 4-5), we explore in earnest how to apply AI models to new drug development problems.
Learn how to handle various data types, from SMILES representing molecular structures to 3D structures, using representative deep learning models.
Through this, you will learn the principles and practice of 'virtual exploration', which predicts drug properties and efficiently finds effective substances in vast databases.

In the latter part (Chapters 6-7), we move on to ‘generative model-based drug design,’ which can be considered the pinnacle of AI drug development.
You'll gain cutting-edge experience in this field by learning how to design novel molecules using relatively simple generative models and optimize them for desired properties.
Finally, we present our vision of an autonomous laboratory that will further accelerate new drug development by automating the so-called design-synthesis-test-analysis cycle through the combination of AI and robotics.
--- From the "Preface"

Eroom's Law is a concept that explains the phenomenon in which, despite advances in science and technology, the efficiency of new drug development is continuously decreasing.
The term is derived from Moore's Law spelled backwards, and refers to the fact that, contrary to the prediction that processing power would double as technology advances, the number of new drugs approved per billion dollars invested in drug development halves roughly every nine years.

One reason for this decline in effectiveness is the drastic tightening of drug regulations following the thalidomide incident in the 1960s.
The incident involved a pregnant woman taking a sedative that resulted in fetal malformation, and since then, regulatory agencies, including the FDA, have been demanding thorough clinical evidence and safety verification during the drug approval process.
This has made new drug development more time-consuming and costly.
--- 「Chapter 1.
From “Basic Concepts of New Drug Development”

In AI drug discovery, deep learning models are commonly used to analyze and predict various biological data, such as drug-target prediction, drug activity prediction, and toxicity prediction.
These predictive tasks involve complex, nonlinear relationships and require modeling the relationship between a drug's chemical properties and the resulting biological responses.
[...] Increasing the depth of neural networks in high-dimensional problems such as AI drug discovery is essential for modeling complex biological and chemical interactions.
Deep neural networks can more accurately predict complex relationships between drugs and biological targets, significantly increasing the likelihood of successful drug discovery.
--- 「Chapter 2.
From "Introduction to deep learning"

Regularization refers to a constraint or auxiliary technique added to a learning algorithm to reduce the generalization error of the model.
When a model overfits its training data, regularization can improve generalization performance and enhance predictive power on new data.
Regularization aims to reduce the generalization error without significantly increasing the training error.
--- 「Chapter 3.
Among the “Regularization Methods”

To understand the structure and principles of deep learning application models (CNN, RNN, etc.) used in new drug development, we must first understand how molecular structures are expressed in a form that computers can process.
Accordingly, this chapter first covers molecular representations before discussing the model in earnest.

A significant portion of modern new drugs are composed of 'small molecules', which are the core substances with pharmacological activity and occupy the center of new drug development.
Small molecule compounds generally have relatively simple structures with a molecular weight of 500 Da or less, and exhibit therapeutic effects by selectively binding to specific biological targets (e.g., proteins, enzymes, etc.).
--- 「Chapter 4.
From "Deep learning models 1"

The central goal of this chapter is to explore the role of graph neural networks and inductive bias in improving the efficiency of deep learning models.
These two topics are crucial for designing models that can effectively analyze complex data structures, thereby improving learning speed and accuracy.
In this course, you will understand how deep learning models recognize relationships and patterns in data, for each topic.
This will provide you with the foundational knowledge needed to solve complex tasks such as molecular structure prediction.
--- 「Chapter 5.
From "Deep learning models 2"

Generative adversarial networks (GANs) are a powerful tool for generating novel molecules in drug design, capable of generating entirely new molecular structures not found in existing data. GANs are generative models that utilize two neural networks to compete to generate data. Instead of explicitly estimating the probability density function of the data, GANs implicitly model the data using neural networks.
The two main components are the generator (G) and the discriminator (D), which learn by competing with each other.
--- 「Chapter 6.
From “Generative AI for drug design”

For AI-based drug development to be truly successful, highly accurate predictive models, high-quality experimental data, and a validation-focused research environment must be organically combined.
For AI to comprehensively support the entire new drug development process in the future, not only technological precision but also integration with biological and clinical interpretations are essential. When a well-established collaborative structure between AI and humans is established, the productivity and success rate of new drug development can be dramatically improved.
In the future, the technologies mentioned above are expected to converge and revolutionize the entire new drug development process.
This will lead to tangible results, such as increasing the success rate of developing treatments for intractable diseases and shortening the development period.

--- 「Chapter 7.
From “Future Outlook”