Optimizing artificial neural network models for metabolomics and systems biology: an example using HPLC retention index data

Application of artificial intelligence to quantitative structure-retention relationship calculations in chromatography

Quantitative structure-retention relationship (QSRR) is an important tool in chromatography. QSRR examines the correlation between molecular structures and their retention behaviors during chromatographic separation. This approach involves developing models for predicting the retention time (RT) of analytes, thereby accelerating method development and facilitating compound identification. In addition, QSRR can be used to study compound retention mechanisms and support drug screening efforts. This review provides a comprehensive analysis of QSRR workflows and applications, with a special focus on the role of artificial intelligence-an area not thoroughly explored in previous reviews. Moreover, we discuss current limitations in RT prediction and propose promising solutions. Overall, this review offers a fresh perspective on future QSRR research, encouraging the development of innovative strategies that enable the diverse applications of QSRR models in chromatographic analysis.

Exploring the chemical subspace of RPLC: A data driven approach

BACKGROUND

The chemical space is comprised of a vast number of possible structures, of which an unknown portion comprises the human and environmental exposome. Such samples are frequently analyzed using non-targeted analysis via liquid chromatography (LC) coupled to high-resolution mass spectrometry often employing a reversed phase (RP) column. However, prior to analysis, the contents of these samples are unknown and could be comprised of thousands of known and unknown chemical constituents. Moreover, it is unknown which part of the chemical space is sufficiently retained and eluted using RPLC.

RESULTS

We present a generic framework that uses a data driven approach to predict whether molecules fall 'inside', 'maybe' inside, or 'outside' of the RPLC subspace. Firstly, three retention index random forest (RF) regression models were constructed that showed that molecular fingerprints are able to predict RPLC retention behavior. Secondly, these models were used to set up the dataset for building an RPLC RF classification model. The RPLC classification model was able to correctly predict whether a chemical belonged to the RPLC subspace with an accuracy of 92% for the testing set. Finally, applying this model to the 91 737 small molecules (i.e., ≤1 000 Da) in NORMAN SusDat showed that 19.1% fall 'outside' of the RPLC subspace.

SIGNIFICANCE AND NOVELTY

The RPLC chemical space model provides a major step towards mapping the chemical space and is able to assess whether chemicals can potentially be measured with an RPLC method (i.e., not every RPLC method) or if a different selectivity should be considered. Moreover, knowing which chemicals are outside of the RPLC subspace can assist in reducing potential candidates for library searching and avoid screening for chemicals that will not be present in RPLC data.

Exploring the Chemical Space of the Exposome: How Far Have We Gone?

Around two-thirds of chronic human disease can not be explained by genetics alone. The Lancet Commission on Pollution and Health estimates that 16% of global premature deaths are linked to pollution. Additionally, it is now thought that humankind has surpassed the safe planetary operating space for introducing human-made chemicals into the Earth System. Direct and indirect exposure to a myriad of chemicals, known and unknown, poses a significant threat to biodiversity and human health, from vaccine efficacy to the rise of antimicrobial resistance as well as autoimmune diseases and mental health disorders. The exposome chemical space remains largely uncharted due to the sheer number of possible chemical structures, estimated at over 1060 unique forms. Conventional methods have cataloged only a fraction of the exposome, overlooking transformation products and often yielding uncertain results. In this Perspective, we have reviewed the latest efforts in mapping the exposome chemical space and its subspaces. We also provide our view on how the integration of data-driven approaches might be able to bridge the identified gaps.

A perspective-driven and technical evaluation of machine learning in bioreactor scale-up: A case-study for potential model developments

Bioreactor scale-up and scale-down have always been a topical issue for the biopharmaceutical industry and despite considerable effort, the identification of a fail-safe strategy for bioprocess development across scales remains a challenge. With the ubiquitous growth of digital transformation technologies, new scaling methods based on computer models may enable more effective scaling. This study aimed to evaluate the potential application of machine learning (ML) algorithms for bioreactor scale-up, with a specific focus on the prediction of scaling parameters. Factors critical to the development of such models were identified and data for bioreactor scale-up studies involving CHO cell-generated mAb products collated from the literature and public sources for the development of unsupervised and supervised ML models. Comparison of bioreactor performance across scales identified similarities between the different processes and primary differences between small- and large-scale bioreactors. A series of three case studies were developed to assess the relationship between cell growth and scale-sensitive bioreactor features. An embedding layer improved the capability of artificial neural network models to predict cell growth at a large-scale, as this approach captured similarities between the processes. Further models constructed to predict scaling parameters demonstrated how ML models may be applied to assist the scaling process. The development of data sets that include more characterization data with greater variability under different gassing and agitation regimes will also assist the future development of ML tools for bioreactor scaling.

Combined mechanistic modeling and machine-learning approaches in systems biology - A systematic literature review

BACKGROUND AND OBJECTIVE

Mechanistic-based Model simulations (MM) are an effective approach commonly employed, for research and learning purposes, to better investigate and understand the inherent behavior of biological systems. Recent advancements in modern technologies and the large availability of omics data allowed the application of Machine Learning (ML) techniques to different research fields, including systems biology. However, the availability of information regarding the analyzed biological context, sufficient experimental data, as well as the degree of computational complexity, represent some of the issues that both MMs and ML techniques could present individually. For this reason, recently, several studies suggest overcoming or significantly reducing these drawbacks by combining the above-mentioned two methods. In the wake of the growing interest in this hybrid analysis approach, with the present review, we want to systematically investigate the studies available in the scientific literature in which both MMs and ML have been combined to explain biological processes at genomics, proteomics, and metabolomics levels, or the behavior of entire cellular populations.

METHODS

Elsevier Scopus®, Clarivate Web of Science™ and National Library of Medicine PubMed® databases were enquired using the queries reported in Table 1, resulting in 350 scientific articles.

RESULTS

Only 14 of the 350 documents returned by the comprehensive search conducted on the three major online databases met our search criteria, i.e. present a hybrid approach consisting of the synergistic combination of MMs and ML to treat a particular aspect of systems biology.

CONCLUSIONS

Despite the recent interest in this methodology, from a careful analysis of the selected papers, it emerged how examples of integration between MMs and ML are already present in systems biology, highlighting the great potential of this hybrid approach to both at micro and macro biological scales.

Predicting RP-LC retention indices of structurally unknown chemicals from mass spectrometry data

Non-target analysis combined with liquid chromatography high resolution mass spectrometry is considered one of the most comprehensive strategies for the detection and identification of known and unknown chemicals in complex samples. However, many compounds remain unidentified due to data complexity and limited number structures in chemical databases. In this work, we have developed and validated a novel machine learning algorithm to predict the retention index (r[Formula: see text]) values for structurally (un)known chemicals based on their measured fragmentation pattern. The developed model, for the first time, enabled the predication of r[Formula: see text] values without the need for the exact structure of the chemicals, with an [Formula: see text] of 0.91 and 0.77 and root mean squared error (RMSE) of 47 and 67 r[Formula: see text] units for the NORMAN ([Formula: see text]) and amide ([Formula: see text]) test sets, respectively. This fragment based model showed comparable accuracy in r[Formula: see text] prediction compared to conventional descriptor-based models that rely on known chemical structure, which obtained an [Formula: see text] of 0.85 with an RMSE of 67.

A pilot study of ion current estimation by ANN from action potential waveforms

Experiments using conventional experimental approaches to capture the dynamics of ion channels are not always feasible, and even when possible and feasible, some can be time-consuming. In this work, the ionic current-time dynamics during cardiac action potentials (APs) are predicted from a single AP waveform by means of artificial neural networks (ANNs). The data collection is accomplished by the use of a single-cell model to run electrophysiological simulations in order to identify ionic currents based on fluctuations in ion channel conductance. The relevant ionic currents, as well as the corresponding cardiac AP, are then calculated and fed into the ANN algorithm, which predicts the desired currents solely based on the AP curve. The validity of the proposed methodology for the Bayesian approach is demonstrated by the R (validation) scores obtained from training data, test data, and the entire data set. The Bayesian regularization's (BR) strength and dependability are further supported by error values and the regression presentations, all of which are positive indicators. As a result of the high convergence between the simulated currents and the currents generated by including the efficacy of a developed Bayesian solver, it is possible to generate behavior of ionic currents during time for the desired AP waveform for any electrical excitable cell.

In depth investigation of the retention behavior of structurally related β-blockers on RP-HPLC column: Quality by design and quantitative structure-property relationship complementary approaches for optimization and validation

The persistent introduction of new β-blockers motivates the demand for optimizing RP-HPLC well-designed analytical procedures that could be applied to this structurally related and commonly prescribed pharmacological group in order to reduce time and chemicals consumption in quality control units. Betoxolol HCl (BEX) and Carvidolol (CAR) were selected as representative examples to conduct predictive studies based on two complementary approaches, Quality by design (QBD) and Quantitative structure property relationship (QSPR). In concern QBD, a Box-Behnken design was adopted at variable chromatographic parameters to achieve the most proper conditions that might be applied for efficient analysis of the majority of group members. On the other hand, the retention time was chosen as the target property in the QSPR study that was conducted onto seven β. blockers (the two investigated drugs in addition to five other β. blockers) to find the best correlated molecular descriptors to the retention behavior. Both external and internal validation studies have comparable quality with training levels. Hence a simple selection algorithm of conventional features provides robust confirmatory predictive QBD and QSPR models. Derringer's desirability function as as a multi-criteria approach was applied for getting the optimum chromatographic analysis conditions. Efficient analysis of BET and CAR was achieved at column temperatures of 26.00 and 27.50 °C, respectively using acetonitrile and phosphate buffer (pH 4.55) 70:30 v/v as a mobile phase with a flow rate of 1.00 mL/min, and UV detection at 220 nm. The method was validated in accordance to ICH guidelines, and had exhibited acceptable precision, accuracy, linearity, and robustness.

Prediction of Liquid Chromatographic Retention Time with Graph Neural Networks to Assist in Small Molecule Identification

The predicted liquid chromatographic retention times (RTs) of small molecules are not accurate enough for wide adoption in structural identification. In this study, we used the graph neural network to predict the retention time (GNN-RT) from structures of small molecules directly without the requirement of molecular descriptors. The predicted accuracy of GNN-RT was compared with random forests (RFs), Bayesian ridge regression, convolutional neural network (CNN), and a deep-learning regression model (DLM) on a METLIN small molecule retention time (SMRT) dataset. GNN-RT achieved the highest predicting accuracy with a mean relative error of 4.9% and a median relative error of 3.2%. Furthermore, the SMRT-trained GNN-RT model can be transferred to the same type of chromatographic systems easily. The predicted RT is valuable for structural identification in complementary to tandem mass spectra and can be used to assist in the identification of compounds. The results indicate that GNN-RT is a promising method to predict the RT for liquid chromatography and improve the accuracy of structural identification for small molecules.

A novel Chebyshev neural network approach for solving singular arbitrary order Lane-Emden equation arising in astrophysics

The motivation of this investigation is to develop a single-layer Chebyshev Neural Network (ChNN) model to handle singular fractional (arbitrary)-order Lane-Emden type equations. These equations are well-known application problems of astrophysics and quantum mechanics. Fractional Lane-Emden equations are singular so it is very difficult to solve analytically. Thus, an efficient method is required to handle the above equations. Here, our main aim is to use a single-layer ChNN model for solving fractional Lane-Emden equations. ChNN model is one kind of Functional Link Neural Network (FLNN) in which the hidden layer is replaced by a functional expansion block of the input pattern using orthogonalshifted Chebyshev polynomials (SChP). Thus, the network parameters of ChNN are less than the Multi-Layer Artificial Neural Network (MLANN). We have considered factional-order singular nonlinear problems of astrophysics to show the computational effort of the proposed method. Back Propagation algorithm of the unsupervised version has been considered for minimizing the error function and updating the weights of the ChNN model. Computed results are displayed in terms of tables and graphs.

The application of artificial neural networks in metabolomics: a historical perspective

BACKGROUND

Metabolomics data, with its complex covariance structure, is typically modelled by projection-based machine learning (ML) methods such as partial least squares (PLS) regression, which project data into a latent structure. Biological data are often non-linear, so it is reasonable to hypothesize that metabolomics data may also have a non-linear latent structure, which in turn would be best modelled using non-linear equations. A non-linear ML method with a similar projection equation structure to PLS is artificial neural networks (ANNs). While ANNs were first applied to metabolic profiling data in the 1990s, the lack of community acceptance combined with limitations in computational capacity and the lack of volume of data for robust non-linear model optimisation inhibited their widespread use. Due to recent advances in computational power, modelling improvements, community acceptance, and the more demanding needs for data science, ANNs have made a recent resurgence in interest across research communities, including a small yet growing usage in metabolomics. As metabolomics experiments become more complex and start to be integrated with other omics data, there is potential for ANNs to become a viable alternative to linear projection methods.

AIM OF REVIEW

We aim to first describe ANNs and their structural equivalence to linear projection-based methods, including PLS regression. We then review the historical, current, and future uses of ANNs in the field of metabolomics.

KEY SCIENTIFIC CONCEPT OF REVIEW

Is metabolomics ready for the return of artificial neural networks?

Challenges in Identifying the Dark Molecules of Life

Metabolomics is the study of the metabolome, the collection of small molecules in living organisms, cells, tissues, and biofluids. Technological advances in mass spectrometry, liquid- and gas-phase separations, nuclear magnetic resonance spectroscopy, and big data analytics have now made it possible to study metabolism at an omics or systems level. The significance of this burgeoning scientific field cannot be overstated: It impacts disciplines ranging from biomedicine to plant science. Despite these advances, the central bottleneck in metabolomics remains the identification of key metabolites that play a class-discriminant role. Because metabolites do not follow a molecular alphabet as proteins and nucleic acids do, their identification is much more time consuming, with a high failure rate. In this review, we critically discuss the state-of-the-art in metabolite identification with specific applications in metabolomics and how technologies such as mass spectrometry, ion mobility, chromatography, and nuclear magnetic resonance currently contribute to this challenging task.

Using LC Retention Times in Organic Structure Determination: Drug Metabolite Identification

Background:

There is a continued need for improvements in the efficiency of metabolite structure elucidation.

Objective:

We propose to take LC Retention Time (RT) into consideration during the process of structure determination.

Methods:

Herein, we develop a simple methodology that employs a Chromatographic Hydrophobicity Index (CHI) framework for standardizing LC conditions and introduce and utilize the concept of a pre-dictable CHI change upon Phase 1 biotransformation (CHIbt). Through the analysis of literature exam-ples, we offer a Quantitative Structure-Retention Relationship (QSRR) for several types of biotransfor-mation (especially hydroxylation) using physicochemical properties (clogP, hydrogen bonding).

Results:

The CHI system for retention indexing is shown to be practical and simple to implement. A da-tabase of CHIbt values has been created from re-incubation of 3 compounds and from analysis of an addi-tional 17 datasets from the literature. Application of this database is illustrated.

Conclusion:

In our experience, this simple methodology allows complementing the discovery efforts that saves resources for in-depth characterization using NMR.

Computational Methods for the Discovery of Metabolic Markers of Complex Traits

Metabolomics uses quantitative analyses of metabolites from tissues or bodily fluids to acquire a functional readout of the physiological state. Complex diseases arise from the influence of multiple factors, such as genetics, environment and lifestyle. Since genes, RNAs and proteins converge onto the terminal downstream metabolome, metabolomics datasets offer a rich source of information in a complex and convoluted presentation. Thus, powerful computational methods capable of deciphering the effects of many upstream influences have become increasingly necessary. In this review, the workflow of metabolic marker discovery is outlined from metabolite extraction to model interpretation and validation. Additionally, current metabolomics research in various complex disease areas is examined to identify gaps and trends in the use of several statistical and computational algorithms. Then, we highlight and discuss three advanced machine-learning algorithms, specifically ensemble learning, artificial neural networks, and genetic programming, that are currently less visible, but are budding with high potential for utility in metabolomics research. With an upward trend in the use of highly-accurate, multivariate models in the metabolomics literature, diagnostic biomarker panels of complex diseases are more recently achieving accuracies approaching or exceeding traditional diagnostic procedures. This review aims to provide an overview of computational methods in metabolomics and promote the use of up-to-date machine-learning and computational methods by metabolomics researchers.

Evaluation of an Artificial Neural Network Retention Index Model for Chemical Structure Identification in Nontargeted Metabolomics

Liquid chromatography coupled with electrospray ionization tandem mass spectrometry (LC-ESI-MS/MS) is a major analytical technique used for nontargeted identification of metabolites in biological fluids. Typically, in LC-ESI-MS/MS based database assisted structure elucidation pipelines, the exact mass of an unknown compound is used to mine a chemical structure database to acquire an initial set of possible candidates. Subsequent matching of the collision induced dissociation (CID) spectrum of the unknown to the CID spectra of candidate structures facilitates identification. However, this approach often fails because of the large numbers of potential candidates (i.e., false positives) for which CID spectra are not available. To overcome this problem, CID fragmentation predication programs have been developed, but these also have limited success if large numbers of isomers with similar CID spectra are present in the candidate set. In this study, we investigated the use of a retention index (RI) predictive model as an orthogonal method to help improve identification rates. The model was used to eliminate candidate structures whose predicted RI values differed significantly from the experimentally determined RI value of the unknown compound. We tested this approach using a set of ninety-one endogenous metabolites and four in silico CID fragmentation algorithms: CFM-ID, CSI:FingerID, Mass Frontier, and MetFrag. Candidate sets obtained from PubChem and the Human Metabolite Database (HMDB) were ranked with and without RI filtering followed by in silico spectral matching. Upon RI filtering, 12 of the ninety-one metabolites were eliminated from their respective candidate sets, i.e., were scored incorrectly as negatives. For the remaining seventy-nine compounds, we show that RI filtering eliminated an average of 58% from PubChem candidate sets. This resulted in an approximately 2-fold improvement in average rankings when using CFM-ID, Mass Frontier, and MetFrag. In addition, RI filtering slightly increased the occurrence of number one rankings for all 4 fragmentation algorithms. However, RI filtering did not significantly improve average rankings when HMDB was used as the candidate database, nor did it significantly improve average rankings when using CSI:FingerID. Overall, we show that the current RI model incorrectly eliminated more true positives (12) than were expected (4-5) on the basis of the filtering method. However, it slightly improved the number of correct first place rankings and improved overall average rankings when using CFM-ID, Mass Frontier, and MetFrag.

Analyzing omics data by pair-wise feature evaluation with horizontal and vertical comparisons

Feature relationships are complex and may contain important information. k top scoring pairs (k-TSP) studies feature relationships by the horizontal comparison. This study examines feature relationships and proposes vertical and horizontal k-TSP (VH-k-TSP) to identify the discriminative feature pairs by evaluating feature pairs based on the vertical and horizontal comparisons. Complexity is introduced to compute the discriminative abilities of feature pairs by means of these two comparisons. VH-k-TSP was compared with support vector machine-recursive feature elimination, relative simplicity-support vector machine, k-TSP and M-k-TSP on nine public genomics datasets. For multi-class problems, one-to-one method was used. The experiments showed that VH-k-TSP outperformed the four methods in most cases. Then, VH-k-TSP was applied to a metabolomics data of liver disease. An accuracy rate of 88.11 ± 3.30% in discrimination between cirrhosis and hepatocellular carcinoma was obtained by VH-k-TSP, better than 77.39 ± 4.10% and 79.28 ± 3.73% obtained by k-TSP and M-k-TSP, respectively. Hence combining the vertical and horizontal comparisons could define more discriminative feature pairs.

Retention Index Prediction Using Quantitative Structure-Retention Relationships for Improving Structure Identification in Nontargeted Metabolomics

Structure identification in nontargeted metabolomics based on liquid-chromatography coupled to mass spectrometry (LC-MS) remains a significant challenge. Quantitative structure-retention relationship (QSRR) modeling is a technique capable of accelerating the structure identification of metabolites by predicting their retention, allowing false positives to be eliminated during the interpretation of metabolomics data. In this work, 191 compounds were grouped according to molecular weight and a QSRR study was carried out on the 34 resulting groups to eliminate false positives. Partial least squares (PLS) regression combined with a Genetic algorithm (GA) was applied to construct the linear QSRR models based on a variety of VolSurf+ molecular descriptors. A novel dual-filtering approach, which combines Tanimoto similarity (TS) searching as the primary filter and retention index (RI) similarity clustering as the secondary filter, was utilized to select compounds in training sets to derive the QSRR models yielding R² of 0.8512 and an average root mean square error in prediction (RMSEP) of 8.45%. With a retention index filter expressed as ±2 standard deviations (SD) of the error, representative compounds were predicted with >91% accuracy, and for 53% of the groups (18/34), at least one false positive compound could be eliminated. The proposed strategy can thus narrow down the number of false positives to be assessed in nontargeted metabolomics.

Development of a Reverse Phase HPLC Retention Index Model for Nontargeted Metabolomics Using Synthetic Compounds

The MolFind application has been developed as a nontargeted metabolomics chemometric tool to facilitate structure identification when HPLC biofluids analysis reveals a feature of interest. Here synthetic compounds are selected and measured to form the basis of a new, more accurate, HPLC retention index model for use with MolFind. We show that relatively inexpensive synthetic screening compounds with simple structures can be used to develop an artificial neural network model that is successful in making quality predictions for human metabolites. A total of 1955 compounds were obtained and measured for the model. A separate set of 202 human metabolites was used for independent validation. The new ANN model showed improved accuracy over previous models. The model, based on relatively simple compounds, was able to make quality predictions for complex compounds not similar to training data. Independent validation metabolites with feature combinations found in three or more training compounds were predicted with 97% sensitivity while metabolites with feature combinations found in less than three training compounds were predicted with >90% sensitivity. The study describes the method used to select synthetic compounds and new descriptors developed to encode the relationship between lipophilic molecular subgraphs and HPLC retention. Finally, we introduce the QRI (qualitative range of interest) modification of neural network backpropagation learning to generate models simultaneously based on quantitative and qualitative data.

Machine Learning Methods for Analysis of Metabolic Data and Metabolic Pathway Modeling

Machine learning uses experimental data to optimize clustering or classification of samples or features, or to develop, augment or verify models that can be used to predict behavior or properties of systems. It is expected that machine learning will help provide actionable knowledge from a variety of big data including metabolomics data, as well as results of metabolism models. A variety of machine learning methods has been applied in bioinformatics and metabolism analyses including self-organizing maps, support vector machines, the kernel machine, Bayesian networks or fuzzy logic. To a lesser extent, machine learning has also been utilized to take advantage of the increasing availability of genomics and metabolomics data for the optimization of metabolic network models and their analysis. In this context, machine learning has aided the development of metabolic networks, the calculation of parameters for stoichiometric and kinetic models, as well as the analysis of major features in the model for the optimal application of bioreactors. Examples of this very interesting, albeit highly complex, application of machine learning for metabolism modeling will be the primary focus of this review presenting several different types of applications for model optimization, parameter determination or system analysis using models, as well as the utilization of several different types of machine learning technologies.