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_5352_Библиотеки_им_академика_М_И_Перельмана
.pdf
372 Chemistry and Biology of Beta-Lactams
https://t.me/med1917
to detect K. pneumoniae producing extended-spectrum beta-lactamase enzymes very quickly.76 The purpose of this study was to evaluate the potential of infrared microspectroscopy along with machine learning algorithms for the rapid detection and identication of ESBL-producing K. pneumoniae in samples
obtained from patients with urinary tract infections. 285 ESBL+ and 365 ESBL− K. pneumoniae samples, gathered from cultured colonies, were examined. As a result of the ndings of the experiment, it
was found that it was possible to determine that K. pneumoniae is ESBL+ within 20 minutes following
the initial culture of the bacteria, with an accuracy of 89%, a sensitivity of 88%, and a specicity of 89%.
There are limited treatment options available as a result of antibiotic resistance, which underscores the
need to optimize the diagnostics that are currently available. The genome sequence of some bacteria species can be used to predict AMR in an unambiguous manner. Machine learning-enabled molecular diagnostics for the prediction of AMR in P. aeruginosa was studied.77 As part of the study, the genomes and
transcriptomes of 414 drug-resistant clinical isolates of P. aeruginosa were sequenced. It was possible to
generate predictive models and identify biomarkers of resistance to four commonly used antimicrobial
drugs by using machine learning classiers trained on the data about the presence or absence of genes,
their sequence variation, and their expression proles. As a result, when these data types were used alone
or in combination, the sensitivity and predictive values were high (0.8–0.9) or very high (>0.9). Overall,
gene expression information improved diagnostic performance for all drugs except for ciprooxacin.
The results provide the basis for the development of a molecular resistance proling tool that is capable
of reliably predicting antimicrobial susceptibility based on the analysis of genomic and transcriptomic
data. By incorporating a molecular susceptibility test system into routine microbiology diagnostics, it is
hoped that earlier and more detailed information will be available about antibiotic resistance proles of
bacterial pathogens, which could change how physicians treat bacterial infections in the future.
A growing AMR in uropathogens is a clinical challenge for emergency physicians as antibiotics should
be selected before an infecting pathogen or the AMR prole of that pathogen can be conrmed. It has
been shown that machine learning can be used to predict ciprooxacin resistance in patients with UTIs
in the emergency department (ED) and the presence of ESBL.78 During January 2020 and June 2021, a
single-center retrospective study was conducted on patients who were diagnosed with UTI in the emergency department. In order to train the model to predict resistance to ciprooxacin and the presence of
extended-spectrum beta-lactamases in urinary pathogens, 39 variables were used in the training process.
Gradient-boosted decision tree (GBDT) has been used for the construction of the model along with
evaluation of its performance. Additionally, the study used SHapely Additive Explanations to visualize
the importance of features. Following two steps of customization of threshold adjustment and feature
selection, the nal model was compared with that of the original prescribers in the emergency department based on the degree of ineffectiveness of the antibiotic that was selected. As a result of customizing
the threshold for decision-making in the GBDT model, the probability of using ineffective antibiotics in
the ED has been signicantly reduced by 20%. In addition to that, it was possible to reduce the number of
predictors down to 20 and 5 variables with high importance while maintaining the same level of performance for the model. In the emergency department, an ML model has the potential to be used to predict
antibiotic resistance in order to improve the effectiveness of empirical antimicrobial treatment given to
patients with UTI. There is a possibility that the model could be used as a point-of-care decision support
tool to assist clinicians in prescribing antibiotics based on individual clinical needs.
Antibiotic resistance is fast becoming one of the most pressing threats to global health, which is estimated to cause 700,000 deaths each year globally as a result of its spread. Besides its surrogates, ARGs
of various kinds can be transmitted easily from food to water, from animal to animal, and from human
to human, thereby reducing the efcacy of antibiotics. To understand the ecology as well as the transmission of ARGs from environmental reservoirs to human-associated reservoirs, accurate identication of
ARGs is an essential step. Despite the fact that previous computational methods for identifying ARGs
were mainly based on the alignment of sequences, these methods are not capable of identifying novel
ARGs, and their application is restricted by the current lack of knowledge about ARGs.
HMD-ARG is a hierarchical multitask deep learning method that was proposed to annotate ARGs.79
As a result of taking raw sequence encoding as an input, HMD-ARG is able to identify multiple ARG
properties simultaneously, without querying existing sequence databases. Among these properties are

373Articial Intelligence on Beta-Lactam Research
https://t.me/med1917
whether the input protein sequence is an ARG, and if it is, what antibiotic family it belongs to, what
resistance mechanism the ARG uses, and if it is an intrinsic ARG or acquired ARG. The HMD-ARG can
further predict the subclass of the beta-lactamase that the ARG is likely to be resistant to if the predicted
antibiotic family is beta-lactamase. Several comprehensive experiments have been conducted in order
to validate the proposed method, including cross-fold validation, third-party validation of the human
gut microbiota dataset, wet-experimental functional validation, and structural investigation of predicted
conserved sites. The results of these experiments show that the proposed method is not only superior to
the existing state-of-the-art methods, but also robust and effective.
It is proposed that a hierarchical multitask method, known as HMD-ARG (Figure 12.6 and Figure
12.7), be developed, which is based on deep learning and is able to inform about the surface characteristics of ARGs from three crucial points of view: resistant antibiotic class, resistant mechanism, and
gene mobility. The HMD-ARG tool is believed to be one of the most powerful tools available to identify
ARGs and therefore mitigate the global threat of antibiotic resistance.
12.3 Beta-Lactam Identification and Quantitation
A photochromic sensor provides the advantages of multiple isomers for multi-analysis, which means that
they provide more sensory information and have more recognition units as well as greater sensitivity to
external stimulation, but they also present an enormous amount of complexity with respect to various
stimulations. There is no doubt that deep learning (DL) algorithms provide a huge advantage when analyzing nonlinear and multidimensional data, but they suffer from their nontransparent inner networks,
which can be described as “black boxes”. There were recent advances in the development of explainable
deep learning-assisted photochromic sensors that can identify beta-lactam antibiotics.80 As a part of
the research, the explainable DL approach was used to process and explicate the photochromic sensing
process. A spirooxazine metallic complex has been used to prepare a multistate analysis array for identifying and quantifying beta-lactams.
The convolutional neural network (CNN) operation has been conducted on a dataset comprised of
2520 unduplicated uorescence intensity images. The method was capable of clearly discriminating six
beta-lactams with a prediction accuracy of 97.98% and was capable of providing rapid quantication
for concentration ranges ranging from 1 mg/L to 100 mg/L. As a result of a molecular simulation and
a class activation mapping analysis, the photochromic sensing mechanism was veried to explain the
mechanism by which a CNN model assesses the importance of photochromic sensor states and makes a
discrimination decision based on them. DL-assisted analysis provides an end-to-end strategy for ascertaining and verifying the complicated sensing mechanism of a device for device optimization and even
for making new scientic discoveries based on the explainable DL-assisted analysis method.
12.4 Artificial Intelligence in Drug Discovery
Drug development has been a long-standing concern for the industry for a number of reasons, including the lengthy process, the high costs, and the inefciency of the process. It takes 10–15 years to
complete the entire process of developing an innovative drug, from the beginning to the end, from
the point of research and development to the point of marketing. It normally requires an investment
of $200 million. Recent advancements in the area of computer technology, particularly in the eld of
articial intelligence, have been making signicant progress in the eld of drug development in recent
years. Articial intelligence drugs have become increasingly widespread as a result of this, which has
led to the widespread adoption of articial intelligence drug design (AIDD), which uses articial intelligence systems and software that can be used to process, interpret, and predict input data using machine
learning (ML) algorithms. There are various approaches to machine learning, which include supervised
learning, unsupervised learning, semi-supervised learning, reinforcement learning, and deep learning
approaches. There are a number of algorithms that are commonly used in the drug development process

374 Chemistry and Biology of Beta-Lactams
https://t.me/med1917
FIGURE 12.6 Overview of HMD-ARG. Adapted with permission from Li Y et al. (2021).

375Articial Intelligence on Beta-Lactam Research
https://t.me/med1917
FIGURE 12.7 HMD-ARG database composition and HMD-ARG database construction pipeline. Adapted with permis-
sion from Li Y et al. (2021).
79
such as recurrent neural networks (RNNs), convolutional neural networks (CNNs), and graph convolutional networks (GCNs).
As a result of the application of articial intelligence in drug research and development over the past
5 years, many companies have been excited to invest in this growing sector. The number of companies
entering the AI pharmaceutical industry has increased to over 150 from 2015 to 2022, with investment
amounts increasing every year. As of 2020, Relay Therapeutics and Atomwise have raised $460 million
and $123 million, respectively, in the AI pharmaceutical industry. By 2021, the total amount of funding
for seven companies, including Exscientia and Benevolent-AI, will have exceeded $2.6 billion. There
have been a number of companies that have utilized AI to design numerous compounds that have the
potential for drug development and have been successful in developing several promising small-molecule
drug candidates, demonstrating the enormous potential of AI in drug development in the coming years.
As a matter of fact, it is important to acknowledge that no technology-based drugs have yet been introduced to the market as of yet. For instance, Exscientia’s AI-driven development of three small-molecule
drugs, DSP-1181, EXS21546, and DSP-0038, has only resulted in DSP-0038 and the cyclin-dependent
kinase 7 (CDK7) inhibitor GTAEXS617 entering phase I clinical trials this year, as stated on the ofcial
website. The development of the other two compounds has been halted, raising concerns about the novelty of these compounds, which has led to their abandonment. As a result, it remains unclear whether AI
can help improve the efciency and speed of the drug discovery process in a meaningful way. It is still
unclear when AI will be able to fully take over drug development as a whole.
It has been reported that articial intelligence can be used in the discovery of small-molecule drugs.81
Articial intelligence can be utilized in the drug design process in order to identify targets and develop
new drugs in an advanced manner. By integrating AI techniques into drug development, signicant efciencies can be achieved in early-stage drug discovery, and the workload involved in drug development
can be reduced signicantly. As part of the study, articial intelligence (AI) has been utilized to design
small drugs with a focus on four key areas in particular: protein structure prediction, molecular virtual

376 Chemistry and Biology of Beta-Lactams
https://t.me/med1917
screening, molecular design, and absorption, distribution, metabolism, excretion, and toxicity (ADMET)
prediction. It is important to point out that although AI has the potential to bring numerous benets to the
early stages of drug development, the direction and quality of decision-making should still be taken into
consideration, as it should be considered a tool rather than a decisive factor, in determining the outcome
of a drug development.
Antibiotic-resistant bacteria are becoming more prevalent by the day, and it has become imperative
to discover new antibiotics in order to combat them. In order to address this challenge, a deep learning
approach to antibiotic discovery (Figure 12.8) has been reported in the literature.82 As a result of performing predictions on multiple chemical libraries, a molecule was identied from the drug repurposing
hub named halicin that is structurally different from conventional antibiotics, but also displays bactericidal activity against a wide spectrum of pathogens such as M. tuberculosis and Enterobacteriaceae that
are resistant to carbapenems. In murine models, halicin is also effective at treating infections caused by
C. difcile and pan-resistant A. baumannii. There are eight antibacterial compounds that are structurally
distant from known antibiotics, which were identied by the model based on a discrete set of 23 empirically tested predictions from a set of over 107 million molecules curated from the ZINC15 database. By
discovering structurally distinct antibacterial molecules using deep learning approaches, this research
demonstrates the utility of deep learning approaches to expand antibiotic arsenals by discovering new
molecules with antibacterial activity.
As beta-lactamases are widely disseminated among pathogenic bacteria, the effectiveness of betalactam antibiotics is greatly limited. As a result of the use of beta-lactamase inhibitors, the activity of
beta-lactam antibiotics has been restored to a considerable extent; however, there is a lack of effective
clinically approved inhibitors against class B metallo-beta-lactamases. It is well known that S. malto-
philia, an opportunistic pathogen that utilizes beta-lactam resistance to survive, produces an enzyme
called L1, which is part of the class B3 enzymes. As far as structural features are concerned, L1 is a
tetramer, with two elongated loops (α3-β7 and β12-α5) around each monomer’s active site. Substrate–
inhibitor binding is inuenced by residues in these two loops. There are a number of gating interactions
that control the conformational changes in the loop of the L1 metallo-beta-lactamase active site.
83
In order to study how the conformational changes of the elongated loops affect the active site in each
monomer, enhanced sampling molecular dynamic simulations were carried out, Markov state models
were built, and convolutional variational autoencoder-based deep learning was applied. An evaluation of
the activity of the generated L1 variants in cell-based experiments was carried out based on mutations at
key residues identied in the previous experiment (D150a, H151, P225, Y227, and R236). As a result of
FIGURE 12.8 Machine learning in antibiotic discovery. Adapted with permission from Stokes JM et al. (2020).

377Articial Intelligence on Beta-Lactam Research
https://t.me/med1917
the data analysis, it has been demonstrated that the gating interactions between loops α3-β7 and β12-α5
are extremely signicant. The gating interactions together with the conformational changes of the key
residues play an important role in the structural remodeling of the active site when they are taken together
as a whole. There is a potential for novel drugs to be developed based on these gating interactions, as
these observations offer insights into the potential for novel drug development.
Transmission electron microscopy (TEM) enzymes are among the most commonly encountered betalactamases with different catalytic abilities against different antibiotics. Even though many studies have
been carried out to investigate the catalytic mechanisms of TEM beta-lactamases, it has been found that
the binding modes of these enzymes against ligands in different functional catalytic states are largely
ignored. However, the binding modes of these proteins may play an important role in the function and
even in the evolution of these proteins over time. A machine learning classication model is presented
for the classication of functional binding modes of the TEM-1 beta-lactamase.84 TEM-1 beta-lactamase
binding modes with respect to penicillin were compared in this study by utilizing a new machine learning
analysis strategy that is capable of recognizing protein dynamic states in different catalytic states as well
as analyzing the binding mode between penicillin and TEM-1 beta-lactamase.
Despite the fact that conventional analysis methods, such as principal component analysis (PCA), are
incapable of discriminating between different binding modes of TEM-1, the application of a machine
learning method produced excellent classication models able to differentiate between these states.
There is also evidence to suggest that both reactant/product states and apo/product states are more differentiable from one another than the apo/reactant states. The feature importance generated by the training
procedure of the machine learning model was utilized to evaluate the contribution from residues at active
sites and in different secondary structures. A key role is played by two key active site residues, Ser70
and Ser130, when it comes to distinguishing reactant/product states, whereas other active site residues
are more important to differentiate apo/product states. There is no doubt that this study provides new
insights into the different dynamical functional states of the beta-lactamase TEM-1 and may open up a
new avenue for the study of beta-lactamases and their evolution in general.
12.5 Determining the Formation of β-Lactam Antibiotic
Complexes with Cyclodextrins
An approach based on machine learning for determining the formation of beta-lactam antibiotic complexes with cyclodextrins using multispectral analysis has been reported.85 As a result of the use of
machine learning on multispectral images, the problem of understanding the formation of complexes
of beta-lactam antibiotics with cyclodextrins (CDs) and the interactions that occur during this process
has been addressed. Complexes of beta-lactam antibiotics, including cefuroxime axetil, cefetamet pivoxil, and pivampicillin, as well as CDs, including αCD, βCD, γCD, hydroxypropyl-αCD, methyl-βCD,
hydroxypropyl-βCD, and hydroxypropyl-γCD, were prepared in all combinations. With the use of dif-
ferential scanning calorimetry, thermograms were obtained that conrmed the formation of cyclodextrin complexes. Transmission Fourier transform infrared (tFTIR) and complementary attenuated total
reectance FTIR (ATR) coupled with machine learning were techniques chosen as a nondestructive
alternative.
Using the machine learning algorithm, it was possible to predict the formation of complexes in samples based solely on their tFTIR and ATR spectra at the prediction stage. In order to support the analysis
by molecular modeling of the complexes, parameterized method 7 (PM7) was used. In cross-validation,
the machine learning model of separating samples that have formed complexes from those that have not
formed complexes had an accuracy of 90.4%. According to the analysis of the contribution of spectral
bands to the model, it was indicated that there were some interactions between the esters of beta-lactam
antibiotics and CDs, as well as some interactions between cephem rings in cefetamet pivoxil and penam
moiety in pivampicillin (Figure 12.9 and Table 12.2). In order to explain experimental results and to suggest possible binding mechanisms, molecular modeling with PM7 was used.
The current generation of machine learning techniques allows robust association of biological signals
with measured phenotypes, but these techniques are not able to identify causal relationships between

378 Chemistry and Biology of Beta-Lactams
https://t.me/med1917
FIGURE 12.9 Optimized structure of prodrug beta-lactam analogs and their maps of electrostatic potentials (MEPs).
Adapted with permission from Mizera M et al. (2019).
TAB LE 12.2
Binding Enthalpies of Complex Conformations Acquired According to Machine Learning Results
(Experimentally Favored) in Relation to Conformations Favored in a Simulated Water Environment
Simulation
α
β
γ
HPα
Mβ
HPβ
HPγ
Source: Adapted with permission from Mizera M et al. (2019).
favored
-157.31 -134.83 -239.54 -235.76 -197.79 -149.85
-238.22 -145.17 -279.69 -131.33 -263.39 -195.71
-255.57 -196.71 -291.79 -178.73 -243.19 -125.94
486.92 558.60 -250.64 -182.03 -235.03 -180.66
-256.15 -168.48 -185.90 -144.22 -236.37 -173.36
-904.60 352.80 -268.65 -259.74 -193.11 -223.94
-352.92 -230.65 -282.78 -263.54 -246.71 -293.35
Cefuroxime axetil Cefetamet pivoxil Pivampicillin
Experimentally
favored
Simulation
favored
Experimentally
favored
85
Simulation
favored
Experimentally
favored

379Articial Intelligence on Beta-Lactam Research
https://t.me/med1917
these biological signals and these measured phenotypes. It is possible to reveal antimicrobial mechanisms
of action using a white-box machine learning approach.86 As part of this project, an integrated “whitebox” biochemical screening, network modeling, and machine learning approach has been developed in
order to reveal causal mechanisms and to apply this approach to understanding antibiotic efcacy. A
genome-scale metabolic network model will be used in order to counterscreen diverse metabolites in E.
coli against bactericidal antibiotics, as well as to simulate their corresponding metabolic states. Using
regression analysis of the measured screen data against model simulations, it has been found that purine
biosynthesis plays a role in antibiotic lethality, which has been validated experimentally. It has been
observed that antibiotic-induced adenine limitation leads to an increase in ATP demand, which leads to
an increase in carbon metabolism activity and oxygen consumption, increasing the killing effects of antibiotics. As a result of this study, it was demonstrated how prospective network models can be combined
with machine learning to identify complex causal mechanisms that contribute to drug efcacy.
12.6 Morphological Deconvolution of Beta-Lactam Polyspecificity in E. coli
Beta-lactam molecules are capable of covalently inhibiting enzymes from the PBP family, which are
essential to the construction of the cell wall of microorganisms. Consequently, beta-lactams cause a dramatic change to the morphology of the cell, the nature of which varies depending on the range of PBPs
that are involved at the same time. The traditional method of examining beta-lactam polyspecicity is
through the use of a gel-based binding assay, which has a low throughput and is typically run ex situ in
cell extracts to determine the degree of polyspecicity. A study on the morphological deconvolution of
beta-lactam polyspecicity in E. coli has been carried out.
The objective of the study was to describe a medium-throughput, image-based assay that can be used,
along with machine learning methods, to automatically prole the activity of beta-lactams in E. coli
cells. It has been demonstrated that having tested for morphological changes across a panel of strains
with perturbations to individual PBP enzymes, this method is capable of automatically and quantitatively
report different beta-lactam antibiotics according to their preference for individual PBPs in cells. The
study suggested that the approach could be used to guide the design of novel inhibitors toward different PBP-binding proles by predicting the mechanisms for two recent inhibitors of PBP that have been
reported.
It was found that treatment led to four forms of cellular morphology previously identied in PBP
literature: ovoid, lament, spindle, and lysed. Further, it was considered that there was an untreated
morphology and an enlarged morphology that were associated with treatments with low concentrations
of compounds (Figure 12.10). As a result of this analysis, it was suggested that this phenotype might
represent the transitional morphology between the untreated morphology and the more canonical morphology caused by PBP inhibitor treatment. From pixels to morphologies with deep learning is shown
in Fig u re 12.11.
87
12.7 Identification of β-Lactamase Proteins
It has been demonstrated that beta-lactamase, which is produced by different bacteria, confers resistance
to beta-lactam-containing drugs. There is a gene that encodes beta-lactamase that is plasmid-borne and
can be easily transferred from one bacterium to another during conjugation. As a result of these transformations, the recipient also acquires resistance to drugs that belong to the class beta-lactams. There is
no doubt that beta-lactam antibiotics can play a vital role in the clinical treatment of devastating diseases
like soft tissue infections, gonorrhea, skin infections, urinary tract infections, and bronchitis. βLactPred is a computer program that predicts the presence of beta-lactamase using statistical moments and
PseAAC via a ve-step rule.
It is the primary amino acid sequence structure that is used as input to the computational model. From
the primary structure, a set of metrics is derived in order to form a feature vector based on those metrics.
88

380 Chemistry and Biology of Beta-Lactams
https://t.me/med1917
FIGURE 12.10 Representative images of each morphological class. Membranes stained with FM4-64FX. DNA was
stained with Hoechst 34580. Nucleic acids were stained with Syto9. Images were acquired at 100X magnication. Adapted
with permission from Godinez WJ et al. (2019).
The experimental data of positive and negative beta-lactamases has been collected and transformed
into feature vectors. As part of the training process, an algorithm based on the articial neural network
is used by integrating the position relative features and the sequence statistical moments in PseAAC.
Several types of approaches have been employed to validate the results of the proposed computational
model, such as self-consistency testing, jackknife tests, cross-validation tests, and independent tests. The
overall accuracy of the predictor for self-consistency, jackknife testing, cross-validation, and independent testing shows 99.76%, 96.07%, 94.20%, and 91.65%, respectively, for the proposed model. Based on
impressive experimental results, it has been demonstrated that the proposed predictor “Lact-Pred” has
surpassed the results of existing methods in terms of accuracy. An articial neural network (ANN) is
one of the most signicant tools for addressing the issues, since it mimics the process of preparing data
as shown in Figure 12.12. In order to clarify the fundamental shape of every residue within a protein, a
neural network is used.
It is possible to estimate the association between antibiotic exposure and colonization with extendedspectrum beta-lactamase-producing Gram-negative bacteria using machine learning methods.89
In this study, the aim was to determine the inuence of prolonged antibiotic administration on the

381Articial Intelligence on Beta-Lactam Research
https://t.me/med1917
FIGURE 12.11 From pixels to morphologies with deep learning. Adapted with permission from Godinez WJ et al. (2019).
FIGURE 12.12 Graphical representation of the articial neural network for βLact-Pred. Adapted with permission from
Ashraf MA et al. (2021).
acquisition of colonization by Gram-negative bacteria producing extended-spectrum beta-lactamases
(ESBL-GNBs) by taking into account individual- and group-level confounding by using machine
learning methods.
A total of 10,034 patients were screened, and 28,322 rectal swab samples were collected. In the absence
and with antibiotic treatment, the incidence of new ESBL-GNB colonization was 22/1,000 and 9/1,000
exposure-days, respectively. Based on the results of the adjusted regression analyses, it was found that
antibiotic exposure, age 60–69 years, and spring season were independently associated with new colonizations. In terms of promoting ESBL-GNB colonization, monotherapy ranks higher than combination
Соседние файлы в папке Библиотека им академика М.И. Перельмана
