Добавил:
Sekretar
kiopkiopkiop18@yandex.ru
t.me/Prokururor I Вовсе не секретарь, но почту проверяю
Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз:
Предмет:
Файл:Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_5338_Библиотеки_им_академика_М_И_Перельмана
.pdf
xii Contents
https://t.me/med1917
11 Electronic-Structure Informatics for Drug Development .......... 187
Manabu Sugimoto
12 Data-Driven Chemistry for Developing Organic Synthesis
Routes for Functional Chemicals ............................... 209
Kenji Hori, Shohei Majima, and Toru Yamaguchi
13 “Quantum-Chemoinformatics” for Design and Discovery
of New Molecules and Reactions ................................ 225
Hiroko Satoh, Vincenz-Maria Steiner, and Jürg Hutter
14 Toxicity Prediction System for Chemical Substances Based
on Toxicity Expression Mechanisms—AI-SHIPS ................. 247
Kimito Funatsu
15 Data Assimilation to Integrate High-Speed Atomic Force
Microscopy with Biomolecular Simulations: Characterization
of Drug Target Functions ...................................... 255
Sotaro Fuchigami and Shoji Takada
16 Potential of High-Spatiotemporal Resolution Live Cell
Imaging for Drug Discovery and Development ................... 273
Yuko Mimori-Kiyosue, Tomonobu Koizumi, and Takashi Washio
17 Design of Biomaterials Using Informatics ........................ 313
Tomohiro Hayashi
18 Monitoring and Controlling in Continuous Manufacturing
Process ....................................................... 323
Kimito Funatsu
19 Formulation Using Hansen Solubility Parameters ................ 335
Hiroshi Yamamoto

Chapter 1
https://t.me/med1917
The AI Trends in Chemical Space
for Drug Discovery
Takuto Koyama and Yasushi Okuno
1.1 Introduction
The process of pharmaceutical development requires significant time (often over
10 years) and cost (exceeding $1 billion), potentially causing delays in bringing
new drugs to the market, and depriving patients of opportunities to receive novel
treatments. Moreover, high development costs contribute to increased drug prices
and heightened risks in research and development. The accumulation of big data in
the field of drug discovery and chemistry through repositories like PubChem [
2], ChEMBL [3], coupled with advancements in Artificial Intelligence (AI),
ZINC [
has garnered attention to the utilization of drug discovery AI technology leveraging
vast compound databases.
In general, small organic compounds are estimated to reach the order of 1060,
making it unrealistic for humans to thoroughly explore this immense chemical space,
which potentially harbors countless undiscovered drug candidates. Hence, AI is anticipated to play a crucial role in the efficient exploration of chemical spaces and the
discovery of new drug candidates. However, challenges in practical applications
arise from the scarcity of experimental compounds that serve as training data, which
affects the predictive accuracy and adaptability range of AI-driven drug discovery.
1],
T. Koy ama · Y. Okuno (B)
Graduate School of Medicine, Kyoto University, 53 Kawahara-Cho, Shogoin Sakyo-Ku,
Kyoto 606-8507, Japan
e-mail: okuno.yasushi.4c@kyoto-u.ac.jp
T. Koy ama
e-mail:
koyama.takuto.82j@st.kyoto-u.ac.jp
Y. Okuno
HPC- and AI-driven Drug Development Platform Division, RIKEN Center for Computational
Science, Kobe 650-0047, Japan
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024
H. Satoh et al. (eds.), Drug Development Supported by Informatics,
https://doi.org/10.1007/978-981-97-4828-0_1
1

2 T. Koyama and Y. Okuno
https://t.me/med1917
These challenges must be overcome for effective utilization in pharmaceutical
settings. This chapter focuses on elucidating recent approaches that have gained attention in addressing the aforementioned challenges faced by current drug discovery AI
technology.
1.2 Representation Methods for Chemical Structures
To handle compounds as data for AI drug discovery, it is necessary to process them
as numerical information. Compound representation methods are techniques that
mathematically and informatively transform the structures and characteristics of
compounds, allowing AI models to process compounds as numerical data. Major
compound representation methods are discussed below (Fig.
1.1).
1.2.1 Representation Method Using Fingerprints
Representation methods using fingerprints are the most fundamental approaches for
identifying compounds (Fig.
fixed-length bit vectors, where each bit position indicates the presence of a specific
substructure or pattern. Representative methods include ECFP [
]. Although fingerprints do not align well with recently developed deep learning
5
[
architectures, they are still widely used in analyses involving input formats and the
compound similarities of traditional machine learning models.
1.1(i)). Fingerprints represent molecular features as
] and MACCS key
4
Fig. 1.1 Molecular representation

1 The AI Trends in Chemical Space for Drug Discovery 3
https://t.me/med1917
1.2.2 String-Based Representation Method
In a string-based compound representation, the molecular structure of a compound is
represented by a string (Fig.
and InChI [
into numerical sequence information, the string undergoes tokenization, where it is
divided into atomic or substructure units and assigned a unique numbering. This
operation, called tokenization, is a transformation method widely used in natural
language processing (NLP). Tokenized SMILES strings can be transformed into
numerical sequence information using techniques such as one-hot encoding or word
embedding, making them suitable inputs for neural networks capable of processing
sequence information, such as Convolutional Neural Networks (CNNs) and language
models.
7], with SMILES being widely used. To convert strings, such as SMILES,
1.1(ii)). Representative notations include SMILES [6]
1.2.3 Graph-Based Representation Method
A graph-based representation method for compounds represents compound structures as graphs (Fig.
structure consisting of vertices (nodes) and edges (lines), with nodes representing
atoms, and edges representing bonds. Graph-based notations emphasize the topology
of compound structures providing intuitive representation of stereochemical information and chemical properties. Moreover, by incorporating features such as atomic
species, electronic states, and bond orders into the nodes and edges of the graph,
various types of compound-related information can be integrated. Graph Neural
Networks (GNNs) are generally used as neural networks to handle graph structures.
1.1(iii)). In the chemical context, a graph is a mathematical
1.3 Construction of Drug Discovery Foundation Models
Using Self-Supervised Learning
Generally, in compound property prediction tasks using machine learning, the scale
of the compound data used for learning significantly affects the performance of the
constructed predictive model. However, in the field of drug discovery and chemistry, training data consists of experimental data obtained through bioassays, making
it challenging and time-consuming to collect such data on a large scale. Therefore, self-supervised learning using large-scale unlabeled data has recently gained
attention. Self-supervised learning is a representation learning method that learns
features related to the input data from a vast amount of unlabeled data. In drug
discovery and chemistry, it is common to conduct pre-training using a large amount
of unlabeled compound data through self-supervised learning and then fine-tune the
intended task using labeled data (Fig.
1.2). A major advantage of self-supervised

4 T. Koyama and Y. Okuno
https://t.me/med1917
Fig. 1.2 Self-supervised learning and fine-tuning
learning is the effective utilization of large amounts of existing unlabeled data, such
as ZINC or PubChem, without relying on labeled data, which involves high experimental costs. Consequently, this approach is expected to achieve high predictive
performance in downstream tasks and expand the scope of applications, making selfsupervised learning-based pre-training models a focus in AI-driven drug discovery.
In this section, the most used self-supervised learning techniques, (A) Attribute
Masking and (B) Contrastive Learning are introduced.
1.3.1 Attribute Masking
In the language domain, Masked Language Models (MLMs) constitute a significant
self-supervised task, forming a fundamental technology in large language models.
MLMs train the model by randomly masking tokens in the input sentence and
predicting the masked tokens using the remaining tokens. Attribute Masking generalizes the concept of MLMs to molecules, involving the masking of certain attributes
(e.g., atoms, bonds, and fragments) within the compound structure. The model is then
trained to predict or restore these masked components from the remaining structure
1.3). Models such as ChemBERTa [8
(Fig.
parts of the SMILES strings, and then restore these masked parts based on the output
of the transformer. In the context of pre-training for graph-based structures, the
proposed method involves randomly masking input atoms or bonds and pre-training
GNNs to predict them using prominent approaches such as GROVER [
11].
[
] or SMILES-BERT [9] randomly mask
10
] and GEM

1 The AI Trends in Chemical Space for Drug Discovery 5
https://t.me/med1917
Fig. 1.3 Outline of attribute masking
1.3.2 Contrastive Learning
Contrastive learning is a technique widely used in the field of image AI, pre-training
models to align representations between pairs of similar inputs, such as two different
augmentations or descriptors of the same molecule. In image AI, various data
augmentation techniques (e.g., rotation, flipping, and zooming) have been employed
to increase the number of image samples. These augmentations, originating from the
same sample, should yield similar representations, and hence, the model learns them
accordingly. In the context of compounds, numerous pre-training methods using
contrastive learning have been proposed (Fig.
13
GraphCL [
train the model to bring the resulting latent representations closer to each other.
Augmentations may involve node masking, edge masking, and subgraph extraction.
However, unlike imaging, augmentation of molecules may pose the risk of changing
the compound itself; therefore, caution is necessary when selecting augmentation
methods.
] input different augmentations of molecular graphs into GNNs and
1.4). Models like MolCLR [12]or
1.4 Applications of Self-Supervised Learning-Based Drug
Discovery Foundation Models
As an application of leveraging self-supervised learning-based pre-training models,
we discuss (A) Molecular property prediction and (B) Prediction of compoundprotein interactions.

6 T. Koyama and Y. Okuno
https://t.me/med1917
Fig. 1.4 Outline of contrastive learning
1.4.1 Molecular Property Prediction
In drug discovery, considering various molecular properties such as solubility,
membrane permeability, and metabolic stability for potential drug candidates is
crucial. Acquiring experimental data related to these compound properties is expensive, thus making self-supervised learning-based pre-training models an approach
that can improve the generalization performance and expand the applicability domain
in drug discovery AI models. In molecular property prediction, benchmark sets like
14
MoleculeNet [
via self-supervised learning often demonstrate higher generalization performance
than supervised learning models such as Graph Convolutional Networks (GCNs).
] are commonly used for model evaluation, and models pre-trained
1.4.2 Prediction of Compound-Protein Interactions
In hit screening for drug development, the interaction between candidate compounds
and target proteins (compound-protein interaction: [CPI]) is a critical element. Hence,
AI-based CPI prediction models are expected to accelerate the drug discovery
process. In CPI prediction AI models, a common approach involves inputting the
molecular structure of compounds and primary sequences of proteins into the AI
model, which is known as Chemical Genomics-Based Virtual Screening (CGBVS)
15] (Fig. 1.5). Recent deep learning-based CGBVS models use separate encoders for
[
compounds and proteins to integrate their hidden layers in the later stages. Therefore,
applying self-supervised learning to a compound encoder can potentially improve the
generalization performance of the CPI prediction. Additionally, for protein encoders,
CNNs or RNNs are generally used; in this case, the effectiveness of representation
learning via self-supervised learning can be expected.

1 The AI Trends in Chemical Space for Drug Discovery 7
https://t.me/med1917
Fig. 1.5 Outline of CPI prediction with AI
1.5 Conclusion
In this chapter, we discussed the current state of AI technology in drug development,
focusing on self-supervised learning and drug discovery foundation models. Selfsupervised learning provides an effective approach to address the bottleneck of insufficient training data in AI applications in drug discovery and chemistry. However,
the impact of the choice of compound datasets for pre-training and the differences in
self-supervised learning methods on AI models in drug discovery remains unclear.
Hence, the continued development of new self-supervised learning methods and
advancements in AI research are expected to further advance AI in drug discovery
and chemistry, promising a bright future outlook.
References
1. Kim S, Chen J, Cheng T, Gindulyte A, He J, He S, et al (2023) PubChem 2023 update. Nucleic
Acids Res 51(D1):D1373–D1380.
2. Sterling T, Irwin TJ (2015) ZINC 15-Ligand Discovery for Everyone. J Chem Inf Model
55(11):2324–2337.
3. Gaulton A, Bellis LJ, Bento AP, Chambers J, Davies M, Hersey A et al (2012) ChEMBL: a
large-scale bioactivity database for drug discovery. Nucleic Acids Res 40(D1):D1100–D1107.
https://doi.org/10.1093/nar/gkr777
4. Rogers D, Hahn M (2010) Extended-connectivity fingerprints. J Chem Inf Model 50(5):742–
754.
https://doi.org/10.1021/ci100050t
5. Durant JL, Leland BA, Henry DR, Nourse JG (2002) Reoptimization of MDL keys for use in
drug discovery. J Chem Inf Comput Sci 42(6):1273–1280.
6. Weininger D (1988) SMILES, a Chemical Language and Information Systems. 1. Introduction
to Methodology and Encoding Rules. J Chem Inf Comput Sci 28:31–36.
1021/ci00057a005
7. Heller S, McNaught A, Stein S, Tchekhovskoi D, Pletnev I (2013) InChi – the worldwide
chemical structure identifier standard. J Cheminformatics 5:7.
2946-5-7
https://doi.org/10.1093/nar/gkac956
https://doi.org/10.1021/acs.jcim.5b00559
https://doi.org/10.1021/ci010132r
https://doi.org/10.
https://doi.org/10.1186/1758-

8 T. Koyama and Y. Okuno
https://t.me/med1917
8. Chithrananda S, Grand G, Ramsundar B (2020) ChemBERTa: Large-Scale Self-Supervised
Pretraining for Molecular Property Prediction. Preprint at
09885
9. Wang S, Guo Y, Wang Y, Sun H, Huang J (2019) SMILES-BERT: Large Scale Unsupervised
Pre-Training for Molecular Property Prediction. In: Shi X, Buck M. (eds) BCB’19: Proceedings
of the 10
Health Informatics, Niagara Falls NY, September 2019, Association for Computing Machinery,
New York, p 429–436.
10. Rong Y, Bian Y, Xu T, Xie W, Wei Y, Huang W, et al (2020) Self-Supervised Graph Transformer on Large-Scale Molecular Data. In: Larochelle H, Ranzato M, Hadsell R, Balcan MF,
Lin H (eds) Advanced in Neural Information Proceeding Systems 33 (NeurIPS 2020), vol
33. Curran Associates, Inc. p 12559–12571.
2020/file/94aef38441efa3380a3bed3faf1f9d5d-Paper.pdf. Accessed 4 Feb 2024
11. Fang X, Liu L, Lei J, He D, Zhang S, Zhou J, et al (2022) Geometry-enhanced molecular
representation learning for property prediction. Nat Mach Intell 4:127–134.
1038/s42256-021-00438-4
12. Wang Y, Wang J, Cao Z, Farimani BA, (2022) Molecular contrastive learning of representations
via graph neural networks. Nat Mach Intell 4:279–287.
00447-x
13. You Y, Chen T, Sui Y, Chen T, Wang Z, Shen Y (2020) Graph Contrastive Learning with
Augmentations. NeurIPS, 33:5812.
14. Wu Z, Ramsundar B, Feinberg EN, Gomes J, Geniesse C, Pappu AS, et al (2018) MoleculeNet:
a benchmark for molecular machine learning. Chem Sci 9(2):513–530.
C7SC02664A
15. Hamanaka M, Taneishi K, Iwata H, Ye J, Pei J, et al (2017) CGBVS-DNN: Prediction
of compound-protein Interactions Based on Deep Learning. Mol Inform. 36(1–2):1600045.
https://doi.org/10.1002/minf.201600045
th
ACM International Conference on Bioinformatics, Computational Biology and
https://doi.org/10.1145/3307339.3342186
https://proceedings.neurips.cc/paper_files/paper/
https://doi.org/10.48550/arXiv.2010.13902
https://doi.org/10.48550/arXiv.2010.
https://doi.org/10.
https://doi.org/10.1038/s42256-022-
https://doi.org/10.1039/

Chapter 2
https://t.me/med1917
Screening Methods for Drugs Using
Chemoinformatics Methods
for Beginners
Hiroshi Yamamoto
2.1 Introduction
Machine learning has progressed rapidly since 2012, when deep learning enabled
computers to recognize pictures of cats. Image recognition and machine translation
have advanced dramatically in the past decade. Such deep learning requires a very
large amount of training data. In the area of chemistry, on the contrary, there is a
wide variety of types: there are 30 million compounds with CAS numbers, 100,000
compounds that are industrially produced, and 5000 compounds that are produced at
a rate of more than 1000 tons per year. However, there is surprisingly little physical
property data on toxicity values and experimental values. There are many reference
books on machine learning, but it is difficult to see how it can be used in actual
research on chemical systems.
This chapter deals with molecular structures and targets materials development,
where available data is very limited. I will show how the various tools can be applied
with actual examples.
The following calculation methods will be explained:
•
Multiple regression methods, (variable selection multiple regression, LASSO
regression, logistic regression, principal component regression)
•
Neural network (NN) methods
•
K-Means method
•
Self-organizing map (SOM) method
•
Decision tree
The subject of this study is the molecular design of halogenated compounds. Fluorine atoms are introduced into many pharmaceutical drugs. However, compounds
H. Yamamoto (B)
53-61 Enokigaoka, Aoba-Ku, Yokohama City Kanagawa-Pref, Japan
e-mail: yamahiro@pirika.com
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024
H. Satoh et al. (eds.), Drug Development Supported by Informatics,
https://doi.org/10.1007/978-981-97-4828-0_2
9
Соседние файлы в папке Библиотека им академика М.И. Перельмана
