Добавил:
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_5606_Библиотеки_им_академика_М_И_Перельмана.pdf
Скачиваний:
0
Добавлен:
02.09.2026
Размер:
21 Мб
Скачать
90 D. S. de Sousa et al.

4 Limitations

ML is a powerful approach to solving a variety of complex problems, but it is determined to recognize and understand its limitations. Like any tool, ML has constraints that can affect its effectiveness in different contexts. The following are some of the most signicant limitations associated with ML.
4.1 Bias
A signicant limitation of ML lies in the unavoidable presence of biases. These biases can be inadvertently incorporated into models due to the nature of the data used in training. When data sets ree ct historical inequalities or mirror the biases of their creators, ML algorithms tend to reproduce and, in some cases, amplify these biases [93, 94].
The source of biases often stems from human decisions during the data collection, feature selection, and labeling process. Additionally, ML algor ithms can unintentionally magnify existing biases in the data, resulting in discriminatory out­comes in areas such as drug efcacy predictions, patient stratication, or adverse effect proling [9395].
The complexity of these issues is exacerba ted by the fact that ML algorithms often operate as black boxes, making it difcult to understand how decisions are made. This makes it challenging to identify and rectify biases, especially when results are presented without a clear explanation [9395].
Mitigating biases in ML models requires a holistic approach that involves developersawareness of ethical challenges, the implementation of more equitable data collection practices, and the exploration of advanced methods to identify and correct biases existing in models [95].
4.2 Overtting and Undertting
Overtting and undertting are two opposing phenomena that can occur when training ML models. Overtting occurs when a model is trained too well on the training data but fails to generalize to new, unseen data. In other words, the model learnsthe training data so well that it ends up capturing specic patterns from that set, including noise, instead of learning general patterns that would apply to other data sets. This often results in lower performance when the model is confronted with data that were not used during training [89, 9395].
On the other hand, undertting occurs when a model is too simple to capture the complexity of the training data. This means that the model fails to learn the underlying patterns and cannot adjust adequately to the training data. As a result, the model performs poorly on both the training data and new data [9698].
4 Machine Learning and Neural Network Methods Applied to Drug Discovery 91
Finding the right balance between overtting and undertting is fundamental for developing an effective ML model. This often involves adjusting the models complexity using techniques such as regularization, data augmentation, cross­validation, and hyperparameter tuning. Cross-validation, for example, can help evaluate the models performance on unseen data and ne-tune its behavior to ensure proper generalization [9395].
4.3 Interpretability
Interpretability is an essential characteristic in ML models, referring to the ability to understand and explain the decisions made by the model. As complex algorithms, such as deep neural networks, gain popularity, the ability to interpret how and why a model makes a particular prediction becomes increasingly important. The lack of interpretability in ML models can be a barrier, especially in critical applications such as target identication, drug design, and pharmacokinetic prole. The inability to explain model decisions can lead to distrust in the results, hindering the acceptance and adoption of these technologies in sensitive environments [9395 , 99].
Several challenges contribute to the lack of interpretability. Complex models, like deep neural networks, often operate as black boxes, making it difcult to understand how a specic input translates into an output. Additionally, interpretability is often sacriced in pursuit of performance, especially in more advanced models. Approaches to improve interpretability include the use of simpler models, such as decision trees, which can be more easily understood. Additionally, post hoc inter­pretability methods, such as Lime and SHAP, have been developed to provide insights into model decisions, even when they are intrinsically complex [100, 101].
Interpretability is not just a technical concern but also an ethical issue. In sectors where decisions directly impact peoples lives, it is decisive that users can under­stand and trust the model predictions. Therefore, advancing research and practices in interpretability is essential to ensure the responsible and ethical use of articial intelligence [100, 101].
4.4 Computational Cost
Computational cost is a decisive consi deration in many aspects of computer science and software development. It refers to the resource-intensive nature of resources such as the processing time of the Central Processing Unit (CPU), memory, and energy associated with the execution of operations or algorithms on a computer system. As problems and data sets addressed by ML algorithms increase in scale and complexity, computational cost becomes a central concern. Adva nced models, such as deep neural networks, may require specialized hardware and substantial comput­ing resources for efcient training.
92 D. S. de Sousa et al.
In addition, the utilization of big data, particularly in virtual high-throughput screening and deep learning applications, especially CNNs, necessitates efciency to ensure prompt and effective responses. Strategies for addressing computational costs encompass algorithm optimization, task parallelization, judicious selection of hard­ware architectures, and the implementation of techniques such as pruning (eliminat­ing insignicant connections in neural networks) to streamline model complexity [102].
The evolution of hardware technology, such as Graphics Processing Units (GPUs) and Tensor Processing Units (TPUs), has played a signicant role in mitigating computational costs in data-intensive tasks like ML model training. Computational cost is not just a technical consideration but also has economic implications, especially in cloud environments where resources are often paid for based on usage. Therefore, computational efciency becomes necessary for opti­mizing costs and ensuring the economic viability of ML-based solutions [102, 103].
4.5 Data Dependency
Data dependency is a central concept in ML, emphasizing the critical inuence that input data have on the performance and effectiveness of models. The quality and representativeness of the data used during model training directly impact the models ability to generalize to new, unseen data. If the training data are not representative of the real-world domain of the problem or is biased in some way, the model may fail to make accurate predictions in real-world situations [9395].
Furthermore, data dependen cy is also related to the need for sufcient data to train models effectively. Complex models, such as deep neural networks, often require large data sets to learn meaningful patterns and avoid overtting. The quality and diversity of data are also determining for addressing potential biases. If the data reect existing inequalities or prejudices, the model may perpetuate or amplify these biases, leading to unfair or discriminatory outcomes. Data dependency is not limited to the training phase; it is also relevant during the inference phase when the model makes predictions or decisions based on input data. If the input data are of low quality, incorrect, or incomplete, the models predictions can be inaccurate or inappropriate [9395].
4.6 Robustness
Robustness in ML refers to the ability of a model to make accurate and consistent predictions across a variety of conditions and situations. However, ML models face several limitations in terms of robustness. One of the primary limitations is sensi­tivity to perturbations in input data. Small changes or noise in the data can lead to signicant variations in model predictions, especially in complex models like deep
4 Machine Learning and Neural Network Methods Applied to Drug Discovery 93
neural networks. This makes models more prone to errors when confronted with data that slightly differ from the training data [93 95 ].
Furthermore, the robustness of ML models is often compromised by adversarial attacks. Adversaries can intentionally manipulate input data subtly to deceive the model and induce errors. This vulnerability to adversarial attacks is a signicant concern in crit ical applications such as security, health, and nance. Another limi­tation is related to the dist ribution of data. If the model is trained on a data set that does not fully represent the real-world application domain, it may struggle to generalize to new data, resulting in a drop in performance. Additionally, robustness is also affected by the presence of outliers in the data. Models sensitive to outliers may have impaired performance when faced with atypical or extreme examples [93
95].

5 Applications in Drug Discovery

The drug discovery proces s is marked by complex stages, precisely because it deals with biological systems that possess numerous pieces of information, many of which are often not completely mapped. This information is now being systematically measured and mined at unprecedented levels using a plethora of omics techniques (proteomics, genomics, epigenomics, metabolomics, and transcriptomics). Cer­tainly, these areas have a high volume of data and variables, and therefore, the application of ML techniques is of great importance [14, 15, 38, 99, 104].
In the drug development stages, one can highlight the identication of targets, lead disco very, pre-clinical development, and clinical development. ML algorithms are applied in all these stages, each with a distinct function and interest, with their applicability dened by the specic characteristics of each algorithm. Figure 4.8 illustrates this relationship in a summarized way [14, 15, 38, 99, 104].
5.1 Target Identication
Target identication is an essential step in the drug discovery process, involving the identication and validation of specic biological molecules or pathways that play a key role in a disease. This intricate process requires a comprehensive understanding of the molecular mechanisms underlying the pathological condition. Researchers employ various techniques such as genomics, proteomics, and bioinformatics to pinpoint potential targets, seeking proteins or genes that are aberrantly expressed or dysregulated in the diseased state. Successful target identication lays the founda­tion for the development of targeted therapies, allowing for the design of drugs that precisely intervene in the disease-associated pathways, potentially leading to more effective and tailored treatments with minimal side effects [14, 15, 20, 99].
Undoubtedly, the process of target identication involves establishing a causal link between the identied target and the associated disease, necessitating the
94 D. S. de Sousa et al.
Target identification Lead discovery
Identification of alternative targets.
Prediction of novel therapeutic targets through target-gene associations
Target druggability
prediction
synthesis plans
creation
Synthesizability
Virtual screening
Molecule
druggability
prediction
prediction
De Novo drug design
Drug
metabolites
prediction
QSAR
Drug
repurposing
Bind affinities
prediction
ADMETox properties prediction
Natural product
identification
Identification
of hits
Preclinical and clinical
development
Drug sensitivity prediction
Metabolic sites
prediction
Patient
stratification
Prediction of tumor
responses
Prediction of biomarkers
Prediction of
drug
mechanisms
Image analysis
Fig. 4.8 Mainly applications of ML algorithms in the drug discovery process
demonstration that manipulating the target yields a discernible impact on the disease. This approach demands diverse data sets with varying dimensionalities and magni­tudes. In addition to high-reliability literature data, it requires comprehensive omics data representing both disease and normal states. The application of ML algorithms further facilitates the treatment, understanding, and elucidation of the causal rela­tionship between the disease and the target [14, 15, 20, 99].
In pursuit of this objective, commonly employed ML algorithms include classi­ers such as SVM, Nearest Neighbors, gradient boosting, and Natural Language Processing (NLP) methods. These classiers are primarily applied for establishing associations among target-disease-drug, identifying druggable and nondruggable targets for specic diseases, and predicting novel therapeutic targets through target– gene associations. Additionally, regression methods, including ANNs, and RF, among others, are applied for target identication. These methods are typically utilized to assess the druggability of targets based on pharmacokinetic properties, protein structure or sequence, and exploration of novel targets such as those associ­ated with Huntingtons disease and the identication of alternative targets [14, 15].
5.2 Lead Discovery
Lead discovery represents one of the most extensive stages in the drug discovery process, typically the primary focus of medicinal chemists. This stage encompasses the discovery of hits, progressing from hit-to-lead to lead optimization. Its focus is on the identication of promising chemical compounds with the potential to become
4 Machine Learning and Neural Network Methods Applied to Drug Discovery 95
the foundation for new therapeutic agents. This intricate stage involves screening vast libraries of compounds to pinpoint molecules exhibiting desirable biological activities against a specic target or pathway associated with a disease. Traditionally, the lead discovery process involves extensive laboratory assays that are time­consuming and expensive. The application of ML in this phase optimizes the process, accelerating the identication of promising drug candidates [14, 15].
In the rst phase corresponding to the identication of hits, virtual screening and drug repurposing stand out. Drug repurposing aims to identify new purposes for existing drugs. In this process, the use of Deep Neural Networks (DNNs) is common, but methods such as naive Bayes, SVM, KNN, and RF can also be found. Addi­tionally, these methods can be applied to predict binding afnities in the virtual screening process based on score functions from docking, without the need to apply physical concepts and functions (which are more complex). Algorithms based on deep learning, such as CNNs, shine in handling 3D structures [14, 15].
In the hit-to-lead process, two approaches heavily involve ML algorithms, namely QSAR and De Novo design. In QSAR, essentially any regression algorithm can be used for predicting biological activity, showcasing the techniques versatility across various methods. In De Novo drug design, RNNs play a signicant role in creating new compounds with desirable biological activity [14, 15]. For more details on the application of QSAR and ML models, see Chap. 6.
In the nal phase corresponding to lead optimization, various methods are used for predicting ADME properties, toxicity, and ultimately the druggability of biolog­ically active compounds. The most employed methods fall within the realm of deep learning such as DNNs and CNNs. Conventional ML methods like SVM can also be found, but their performance is markedly inferior to DL methods. Another deter­mining factor is assessing synthesizability and creating synthesis plans for the envisioned compounds, a process often performed by reinforcement learning algo­rithms. Furthermore, ML algorithms nd applications in predicting drug metabolites and metabolic sites, as well as natural product identication [14, 15].
5.3 Preclinical and Clinical Development
The evolution of ML application in clinical and pre-clinical stages is constantly progressing, and already showing signicant advancements. In the pre-clinical phase, ML algorithms stand out for their ability to identify and predict biomarkers, predominantly utilizing regression and classication approaches. These biomarkers are of utmost relevance, enabling a deeper understanding of drug mechanisms and efciently differentiating them for specic patients. Commonly used traditional ML methods for this purpose include SVM and RF. These models are applied in the initial stages of clinical trials, aiming to reduce both the time and costs associated with the nal stages [14, 15].
In the clinical context, the application of ML model s for biomarker discovery and prediction of therapeutic responses continues to stand out. The emphasis on
96 D. S. de Sousa et al.
classication approaches is evident, where algorithms are employed to stratify patients, identify potential indications, and suggest drug mechanisms [14, 15].
Another critical application is the construction of drug sensitivity models, where supervised learning techniques, such CNNs, are employed. These models are par­ticularly applicable in digital pathology, where image analysis allows the prediction of tumor responses to different treatments. Drug sensitivity models and their corresponding biomarkers are validated by independent test data sets, covering both pre-clinical trials and early clinical stages [14, 15].

6 Resources and Tools

The discovery of new drugs represents a complex challenge, but recent advances in ML implicated accelerating this process. Var ious software and online platforms have been developed to integrate ML algorithms in the analysis of vast sets of biological and chemical data, providing valuab le insights for scientists and researchers. These tools offer a variety of applications, ranging from target identication to drug prioritization, prediction of properties, and assistance in generating synthesis path­ways. In this dynamic and challenging eld, the diversity of available tools is notable, as detailed in the accompanying Table 4.5 , which provides a comprehensive overview of various platforms, their availability, software format, and specic applications.
Beyond tools, databases are fundamental in constructing suitable models for drug discovery, serving as a critical starting point for the ML process. These repositories of information are necessary for the collection and organization of relevant data, providing a robust foundation for the implementation of ML algorithms. Table 4.6 presents the main databases along with their respective descriptions used in the eld of drug discovery for building ML models.

7 Challenges and Perspectives

The integration of ML into drug discovery processes holds immense promise, offering unprecedented advancements in identifying potential therapeutic com­pounds and accelerating the development of new drugs. However, this transforma­tive approach is not without its challenges and ethical considerations. Among the challenges, stand out:
Experimental Validation While ML offers powerful predictive capabilities, it is not a substitute for traditional experimental methods. Experimental validation remains indispensable to conrm and interpret ML predictions accurately. The integration of ML with experimental approaches can synergize the drug discovery process. ML can efciently analyze vast data sets and propose potential drug
4 Machine Learning and Neural Network Methods Applied to Drug Discovery 97
Table 4.5 Some key ML-based tools used in the drug discovery process
Name Availability
Open targets [105] Free Web
admetSAR [106] Free Web
ACD/Percepta [107]
pkCSM [108] Free Web
IBM RXN [109] Free/
DeepChem [110] Free Library Multi-application in drug design
RDKit [111] Free Library Multi-application in drug design
AutoQSAR [112] Commercial Installable QSAR modeling
PaDEL-Descriptor [113]
SYBYL-X [114] Commercial Installable Molecular descriptors calculation/QSAR
MatLab [115] Commercial Installable QSAR modeling
Commercial Installable Properties
Commercial
Free Installable Molecular descriptors calculation/QSAR
Software format Application
platform
platform
platform
Web platform
Target identication and drug prioritization
Properties Prediction
Prediction
Properties Prediction
Synthesis pathway generation
modeling
modeling
candidates, but human researchers must validate and interpret these ndings. The collaborative efforts of ML and human expertise can optimize the drug discovery process, leading to the accelerated development of new drugs [127]. For more details on experimental assays, see Chap. 12.
Regulation and Ethics The integration of ML in drug discovery raises ethical considerations and regulatory challenges. Decision-making processes impacti ng peoples health, such as drug development choices and clinical trial selections, are increasingly inuenced by AI algorithms. Ensuring fairness, transparency, and unbiased outcomes becom es paramount. Biases in AI algorithms could lead to unequal access to medical treatment, contradicting principles of equality and justice. Additionally, the potential automation of certain tasks in the pharmaceutical industry may raise concerns about job losses. Ethical AI implementation involves regulatory compliance, periodic audits for bias, and the development of protocols ensuring data privacy and security [128].
Generalization The challenge of generalization involves the ability of ML models to apply learned knowledge to new, unseen data. ML models trained on specic data sets may struggle to generalize their predictions to diverse patient populations or different experimental conditions. Overtting to a particular data set can limit the models adaptability. Techniques such as transfer learning, where models leverage knowledge gained from one domain to enhance performance in another, can help address generalization challenges in drug discovery ML [15].
98 D. S. de Sousa et al.
Table 4.6 Main databases used in the drug discovery process
Name Description URL
ChEMBL [24] Information on bioactive molecules with drug-
DrugBank [23] Information on drugs and drug targets https://go.drugbank.com/
PubChem [22] Information on chemical structures, identiers,
UniProt [116] Information on protein sequences and their
RCSB PDB [117]
DrugCentral [118]
ADReCS [119] Adverse Drug Reaction Classication System http://bioinf.xmu.edu.cn/
CTD [120] Comparative Toxicogenomics Database https://ctdbase.org/
DisGeNET [121]
UK BioBank [122]
STRING data­base [123]
ZINC [21] Free database of commercially available com-
Therapeutic Tar­get Database [124]
Promiscuous 2.0 [125]
ClinicalTrials. gov [126]
like properties
properties, activities, patents, safety, toxicity data
functions, many entries from genome sequenc­ing projects
Archive of 3D structure data for large biological molecules (proteins, DNA, RNA)
Comprehensive information on drugs and their targets
Discovering gene–disease associations https://www.disgenet.org/
United Kingdom Biobank: Large-scale bio­medical database
Protein–protein interaction information https://string-db.org/
pounds for virtual screening
Information on the known and explored thera­peutic protein and nucleic acid targets
Database of promiscuous proteins and compounds
Database of privately and publicly funded clin­ical studies conducted worldwide
https://www.ebi.ac.uk/ chembl/g/
https://pubchem.ncbi.nlm. nih.gov/
https://www.uniprot.org/
https://www.rcsb.org/
https://drugcentral.org/
ADReCS/
https://www.ukbiobank. ac.uk/
https://zinc.docking.org/
https://idrblab.net/ttd/
https://bioinf-applied. charite.de/promiscuous2/ index.php
https://clinicaltrials.gov/
Data Integration In the dynamic landscape of drug discovery, the integration of data emerges as a critical hurdle, encompassing both heterogeneous inputs from omicsdomains and the complexities of homogeneous databases. Robust pipeline architectures, powered by Extract Transform Load (ETL) tools, are indispensable for orchestrating the seamless convergence of diverse data sets from public repositories. This integration process demands not only technical prowess but also the application of advanced technologies, including ML and big data analytics. Even within homo­geneous data sets, challenges persist, spanning testing, logical considerations, cross­platform normalization, and statistical intricacies. ML and big data analytics serve as invaluable assets in navigating these challenges, providing a cohesive framework for integrating diverse strands of information. In the pharmaceutical realm, where
4 Machine Learning and Neural Network Methods Applied to Drug Discovery 99
research spans from molecular insights to individual patient data, the challenge of integrating heterogeneous data becomes particularly pronounced. The need for a sophisticated level of AI becomes evident, ensuring a comprehensive understanding of data gathered across varying contex ts and scales. Modern data connectors are essential for centralizing dissimilar data, offering a strategi c solution to allocate and manage this diverse information effectively [15].
7.1 Future Trends
At the crossroads of science and technology, drug discovery is undergoing a signicant revolution, thanks to the increasing application of ML techniques. This revolution not only accelerates the process of identifying new drugs but also pro­vides a deeper and more personalized understanding of therapies. In this context, several emerging trends act as promising catalysts for the future of drug discovery.
Transfer Learning This approach stands out as a dominant trend in the eld, characterized by the widespread adoption of pre-trained machine learning models. This methodology involves utilizing models that have been initially trained on extensive data sets, which may not necessarily pertain to the domains of biology or medicine. However , these pre-trained models can be ne-tuned and adapted to specic tasks within drug discovery. By capitalizing on the knowledge acquired in a diverse range of domains, these models facilitate the accelerated analysis of intricate biological data. This, in turn, provides a substantial advantage in pinpointing potential therapeutic targets and identifying promising compounds [129].
Explainable AI (XAI) The reliance on ML-driven drug discoveries crucially depends on the ability to interpret and understand model decisions. The rise of Explainable Articial Intelligence (XAI) is, therefore, a response to this need. As ML models become more complex, ensuring transparency and interpretability becomes paramount. XAI techniques provide insights into how models make deci­sions, allowing researchers to validate and better understand predictions, ensuring more reliable outcomes [130].
Use of Imaging Data and Molecular Structure The integration of imaging and molecular structure data marks another innovative frontier in drug discovery. Advances in imaging techniques, such as high-resolution microscopy and computed tomography, are now complemented by ML algorithms that can extract crucial information from these images. Similarly, the analysis of complex molecular struc­tures, aided by ML models, allows for a more rened understanding of interactions between chemical compounds and biological targets [14].
Deep Learning Advancements Advancements in Deep Learning continue to shape the landscape of drug discovery. More complex models, such as deep neural networks, have the ability to learn more abstract hierarchical representations, cap­turing subtle nuances in data. This capability for extracting complex features is