Missing data in multi-omics integration: Recent advances through artificial intelligence

“Missing ‘omics data are accommodated through an expectation-maximization (EM) approach (Dempster et al., 1977) for parameter estimation.”

“For example, many of the deep-learning architectures mentioned in this review have similarities to ones used in analysis of digital media that try to combine multiple modalities to improve predictive performance of video content (Carreira and Zisserman, 2017; Afouras et al., 2019).”

“DeepIMV’s latent mappings are trained according to the informational bottleneck principle (Tishby et al., 2000; Alemi et al., 2016; Achille and Soatto, 2017), which seeks to define task-relevant data representations that respect the balance between parsimony and predictive power.”

“Finally, additional sources of uncertainty may arise from processes such as protein quantification in bottom-up proteomics (Rifai et al., 2006; Plubell et al., 2022) as measurements are made at the peptide level and isoforms can be difficult to identify (Forshed et al.”

“DeepIMV’s latent mappings are trained according to the informational bottleneck principle (Tishby et al., 2000; Alemi et al., 2016; Achille and Soatto, 2017), which seeks to define task-relevant data representations that respect the balance between parsimony and predictive power.”

“…‘omics data types, such as single cell ‘omics (Ma et al., 2020) or specific application areas, such as plant biology (Rajasundaram and Selbig, 2016), microbial communities (Fondi and Liò, 2015; Franzosa et al., 2015), cancer (Buescher and Driggers, 2016), and others (e.g., Subramanian et al., 2020).”

“, 2018) and its more recent improvement, MOFA+ (Argelaguet et al., 2020) are unsupervised integrative approaches that may be understood as generalizations of principal components analysis (PCA) to multi-omics data.”

“A sensible partition is of early, middle, and late integrative methods, alternatively denoted as concatenation, transformation-based, or model-based integration, respectively (Ritchie et al., 2015).”

“…of subsets of general AI and ML approaches to multi-omics data integration (e.g., Li et al., 2016; Huang et al., 2017; Kim and Tagkopoulos, 2018; Picard et al., 2021; Reel et al., 2021; Lee and Kim, 2022) with some attention given to limitations of methods, such as interpretability of models.”

“In untargeted metabolomics and lipidomics false discovery rates are difficult to compute and largely not accurate (Stein, 1994; Jeong et al., 2011; Matsuda et al., 2013; Kim and Zhang, 2014).”

“For example, many of the deep-learning architectures mentioned in this review have similarities to ones used in analysis of digital media that try to combine multiple modalities to improve predictive performance of video content (Carreira and Zisserman, 2017; Afouras et al., 2019).”

“When α = 0 the JACA loss reduces to that of CCA or SUMCOR multiset CCA (Kettenring, 1971) with sparsity regularization, depending on whether three or more ‘omics are under consideration.”

“, 2020) or specific application areas, such as plant biology (Rajasundaram and Selbig, 2016), microbial communities (Fondi and Liò, 2015; Franzosa et al., 2015), cancer (Buescher and Driggers, 2016), and others (e.”

“…a handful of reviews are dedicated to discussion of subsets of general AI and ML approaches to multi-omics data integration (e.g., Li et al., 2016; Huang et al., 2017; Kim and Tagkopoulos, 2018; Picard et al., 2021; Reel et al., 2021; Lee and Kim, 2022) with some attention given to limitations of…”

“Proteomics instrumentation can measure the relative abundance of proteins that are produced from gene expression and transcript translation (Graves and Haystead, 2002).”

“When an assumption of MCAR or MAR is used, the missingness is said to be ignorable and methods designed to handle missing values are applied to both cases (Gelman and Hill, 2006; Fang et al., 2018; Wei et al., 2018).”

“Picard et al. (2021) further separate the early-middle-late categorization into early, mixed, late, intermediate, and hierarchical.”

“However, data integration is one of the primary barriers to effective multi-omics research and reaching the desired holistic view of the biological system being studied (Misra et al., 2019).”

“As a joint-imputation approach, MOFA/MOFA + was compared to other imputation strategies including imputation by feature-wise mean, SoftImpute (Mazumder et al., 2010) and a k-nearest neighbor approach (Troyanskaya et al., 2001).”

“BIDIFAC+ builds upon BIDIFAC (Park and Lock, 2019) and SLIDE (Gaynanova and Li, 2019), two extensions of the JIVE method (Lock et al., 2013) that yields matrix factorizations that convey a decomposition of data variability into variability shared across multiple ‘omics and variability primarily attributable to single…”

“Machine learning (ML) and artificial intelligence (AI) have shown promise in overcoming some of these hindrances (e.g., Mirza et al., 2019; Zhu, 2020) and have gained popularity in recent years.”

“Some reviews focus on differentiating general data-driven integration approaches from a mathematical perspective (e.g., Ritchie et al., 2015; Bersanelli et al., 2016; Noor et al., 2019), focus on a particular type of method, such as dimension reduction (Meng et al.”

“Another example, MOLI (Sharifi-Noghabi et al., 2019), similarly processes each view separately through a DNN and concatenates The vast majority of these integration approaches require completely observed data for their implementation.”

“, 2019), focus on a particular type of method, such as dimension reduction (Meng et al., 2016), or compare data-driven models against biologically-informed models for a particular model type, such as networkmodels (Wanichthanarak et al.”

“Machine learning (ML) and artificial intelligence (AI) have shown promise in overcoming some of these hindrances (e.g., Mirza et al., 2019; Zhu, 2020) and have gained popularity in recent years.”

“, 2022) and a majority focused on cancer research (e.g., Wang and Gu, 2016; Biswas and Chakrabarti, 2020; Nicora et al., 2020; Cai et al., 2022).”

“This is as opposed to other (larger) transcriptomics platforms that would require at least 35 samples per group to detect the same magnitude fold change at the same power and false discovery rate levels (Krassowski et al., 2020).”

“Late integration examples include the method introduced by Sun et al. (2019) that processes each ‘omic type separately through a deep neural network (DNN) and then performs score fusion on the predictions.”

“These generally form viewspecific representations and assume a common latent space among views (Yang and Michailidis, 2015; Chalise and Fridley, 2017) or maximize some measure among the view-specific representations (Tenenhaus and Tenenhaus, 2011;Meng et al., 2014; Tenenhaus et al., 2017; Singh et al., 2019).”

“, stationary phase and ionization mode) induces increased instrumental sensitivity toward some substances and reductions toward others (Ribbenstedt et al., 2018).”

“These generally form view-specific representations and assume a common latent space among views (Yang and Michailidis, 2015; Chalise and Fridley, 2017) or maximize some measure among the view-specific representations (Tenenhaus and Tenenhaus, 2011; Meng et al., 2014; Tenenhaus et al., 2017; Singh et…”

“For example, the human proteome project reports that an estimated ∼20% of genes yield protein products that are not detected by mass spectrometry, an analytic platform for protein quantification (Paik et al., 2012; Baker et al., 2017).”

“Neighborhood based Multi-Omics clustering (NEMO; Rappoport and Shamir, 2019) is an algorithm that generates clusters based on individual data view similarity matrices.”

“On the other hand, many of these articles focus on integration for specific ‘omics data types, such as single cell ‘omics (Ma et al., 2020) or specific application areas, such as plant biology (Rajasundaram and Selbig, 2016), microbial communities (Fondi and Liò, 2015; Franzosa et al.”

“These generally form view-specific representations and assume a common latent space among views (Yang and Michailidis, 2015; Chalise and Fridley, 2017) or maximize some measure among the view-specific representations (Tenenhaus and Tenenhaus, 2011; Meng et al., 2014; Tenenhaus et al., 2017; Singh et…”

“…engineering ( Helmy et al., 2020), precision medicine (Hamamoto et al., 2019), amongst others (e.g., Mann et al., 2021; Lin et al., 2022; Zhou et al., 2022) and a majority focused on cancer research (e.g., Wang and Gu, 2016; Biswas and Chakrabarti, 2020; Nicora et al., 2020; Cai et al., 2022).”

“It has been shown that missing values in large-scale ‘omics data can hinder downstream analyses (Ouyang et al., 2004; Jörnsten et al., 2005).”

“DeepIMV (Lee and van der Schaar, 2021) incorporates some of the principles utilized by Wu and Goodman (2018) to develop a deep variational information bottleneck approach that maps integrated, multi-omics data into a lower-dimensional subspace of task-relevant embeddings.”

“It is not uncommon to have 20–50% of the possible peptide values that are not quantified (WebbRobertson et al., 2015; Brenes et al., 2019).”

“Included in these are articles discussing examples of method application (e.g., Holzinger et al., 2019; Li et al., 2021) and challenges in applying AI and ML techniques to multi-omics data (e.g., Kang et al., 2021; Termine et al., 2021).”

“However, like mean imputation, imputed values from these approaches may distort feature distributions and lead to biased parameter and variability estimates (Lubin et al., 2004; Succop et al., 2004; Webb-Robertson et al., 2015; Lazar et al., 2016; Bramer et al., 2020).”

“…2019), focus on a particular type of method, such as dimension reduction (Meng et al., 2016), or compare data-driven models against biologically-informed models for a particular model type, such as network models (Wanichthanarak et al., 2015) with little attention on particular ‘omics data types.”

“…form view-specific representations and assume a common latent space among views (Yang and Michailidis, 2015; Chalise and Fridley, 2017) or maximize some measure among the view-specific representations (Tenenhaus and Tenenhaus, 2011; Meng et al., 2014; Tenenhaus et al., 2017; Singh et al., 2019).”

“Similar co-regularization approaches for semisupervised multi-omic or (more generally) multi-view learning have been previously introduced (Sindhwani et al., 2005; Brefeld et al., 2006; Sun and Shawe-Taylor, 2010).”

“These methods also commonly investigate differences in “early”, “middle”, and “late” fusion and work has even been done to address the view-missing and view-heterogeneity problems (Nagrani et al., 2018; Wang et al., 2020).”

“Included in these are articles discussing examples of method application (e.g., Holzinger et al., 2019; Li et al., 2021) and challenges in applying AI and ML techniques to multi-omics data (e.g., Kang et al., 2021; Termine et al., 2021).”

“For contrast, another deep learning-based method (Xu et al., 2019) trains separate networks that see only the data from a particular view and then fuses the hidden representations at some intermediate layer.”

“It has been shown that missing values in large-scale ‘omics data can hinder downstream analyses (Ouyang et al., 2004; Jörnsten et al., 2005).”

“…‘omics data types, such as single cell ‘omics (Ma et al., 2020) or specific application areas, such as plant biology (Rajasundaram and Selbig, 2016), microbial communities (Fondi and Liò, 2015; Franzosa et al., 2015), cancer (Buescher and Driggers, 2016), and others (e.g., Subramanian et al., 2020).”

“Other common imputation approaches include random forest imputation (Pantanowitz and Marwala, 2009) and expectation maximization (Dempster et al.”

“For example, batch effects are common and often much more pronounced in mass spectrometry-based data (Mertens, 2016; Phua et al., 2022; e.g., Han and Li, 2020), even when the same physical instrument is used and the same data type is being generated.”

“Xie et al. (2019) use deep learning to perform survival analysis, which involves feature selection by removing observations with high missingness, early concatenation of raw features to form the input to their model, and implicit feature extraction using deep neural networks, as the value at each…”

“For example, the human proteome project reports that an estimated ∼ 20% of genes yield protein products that are not detected by mass spectrometry, an analytic platform for protein quantification (Paik et al., 2012; Baker et al., 2017).”

“Finally, additional sources of uncertainty may arise from processes such as protein quantification in bottom-up proteomics (Rifai et al., 2006; Plubell et al., 2022) as measurements are made at the peptide level and isoforms can be difficult to identify (Forshed et al., 2011; Webb-Robertson et al.,…”

“…methods such as graph representations or other hard constraints on parameter representations with machine learning methods are ongoing (Noor et al., 2019; Lee and Kim, 2022), and evaluation and development of mechanisms for handling missing data within these frameworks will be…”

“…engineering ( Helmy et al., 2020), precision medicine (Hamamoto et al., 2019), amongst others (e.g., Mann et al., 2021; Lin et al., 2022; Zhou et al., 2022) and a majority focused on cancer research (e.g., Wang and Gu, 2016; Biswas and Chakrabarti, 2020; Nicora et al., 2020; Cai et al., 2022).”

“BIDIFAC + builds upon BIDIFAC (Park and Lock, 2019) and SLIDE (Gaynanova and Li, 2019), two extensions of the JIVE method (Lock et al., 2013) that yields matrix factorizations that convey a decomposition of data variability into variability shared across multiple ‘omics and variability primarily attributable to single…”

“…focus on integration for specific ‘omics data types, such as single cell ‘omics (Ma et al., 2020) or specific application areas, such as plant biology (Rajasundaram and Selbig, 2016), microbial communities (Fondi and Liò, 2015; Franzosa et al., 2015), cancer (Buescher and Driggers, 2016), and…”

“Aggregation of these normal distributions is achieved through the process of statistical conflation (Hill, 2011), which involves taking the normalized product of the aggregated probability density functions.”

“Other studies that involve rare diseases, the collection of difficult-to-access tissue, or smaller grants that do not allow for large consortium samples are characterized by similarly low sample sizes (Sirrs et al., 2015; Abela et al., 2017).”

“Peptides may be missing from one or more samples due to a variety of underlying mechanisms (Daly et al., 2008; Karpievitch et al., 2009), which can result in non-trivial amounts of potential observations being missing.”

“Trans-Omics Block Missing Data Imputation (TOBMI; Dong et al., 2018) proposes a similar method, but imputed values are generated based on a weighted average of neighbors, with closer neighbors having greater weights.”

“However, like mean imputation, imputed values from these approaches may distort feature distributions and lead to biased parameter and variability estimates (Lubin et al., 2004; Succop et al., 2004; Webb-Robertson et al., 2015; Lazar et al., 2016; Bramer et al., 2020).”

“, 2021) and challenges in applying AI and ML techniques to multi-omics data (e.g., Kang et al., 2021; Termine et al., 2021).”

“For example, a study to identify biomarkers for Snyder Robinson syndrome, a rare neurological disorder, obtained data on 16 patients, only three of which presented with the disorder of interest (Abela et al., 2015).”

“For example, batch effects are common and often much more pronounced in mass spectrometry-based data (Mertens, 2016; Phua et al., 2022; e.g., Han and Li, 2020), even when the same physical instrument is used and the same data type is being generated.”

“In untargeted metabolomics and lipidomics false discovery rates are difficult to compute and largely not accurate (Stein, 1994; Jeong et al., 2011; Matsuda et al., 2013; Kim and Zhang, 2014).”

“For small molecule measurements, such as metabolomics, manual verification of metabolites following identification scoring is required for shotgun metabolomics, as scoring methods are much less mature than for other ‘omics data types (Stein and Scott, 1994; Samaraweera et al., 2018).”

“, 2022) as measurements are made at the peptide level and isoforms can be difficult to identify (Forshed et al., 2011; Webb-Robertson et al., 2014).”

“…metabolic engineering ( Helmy et al., 2020), precision medicine (Hamamoto et al., 2019), amongst others (e.g., Mann et al., 2021; Lin et al., 2022; Zhou et al., 2022) and a majority focused on cancer research (e.g., Wang and Gu, 2016; Biswas and Chakrabarti, 2020; Nicora et al., 2020; Cai et al.,…”

“When an assumption of MCAR or MAR is used, the missingness is said to be ignorable and methods designed to handle missing values are applied to both cases (Gelman and Hill, 2006; Fang et al., 2018; Wei et al., 2018).”

“…isolation and solubilization, sequence ambiguity, and a lack of standards in statistical thresholds and algorithms result in inconsistent detection of large proportions of protein products that are detected, thus limiting the reproducibility of data collection and analysis (Goh and Wong, 2017).”

“, 2022) as measurements are made at the peptide level and isoforms can be difficult to identify (Forshed et al., 2011; Webb-Robertson et al., 2014).”

“effects are common and often much more pronounced in mass spectrometry-based data (Mertens, 2016; Phua et al., 2022; e.g., Han and Li, 2020), even when the same physical instrument is”

“, 2019), amongst others (e.g., Mann et al., 2021; Lin et al., 2022; Zhou et al., 2022) and a majority focused on cancer research (e.”

“Other studies that involve rare diseases, the collection of difficult-to-access tissue, or smaller grants that do not allow for large consortium samples are characterized by similarly low sample sizes (Sirrs et al., 2015; Abela et al., 2017).”

“In untargeted metabolomics and lipidomics false discovery rates are difficult to compute and largely not accurate (Stein, 1994; Jeong et al., 2011; Matsuda et al., 2013; Kim and Zhang, 2014).”

“A more recent approach, Multiple Similarity Network Embedding (MSNE; Xu et al., 2021) modifies NEMO’s original framework to overcome limitations in the patterns of missing data that can be accommodated.”

“Multi Omics clustering by Non-Exhaustive Types (MONET; Rappoport et al., 2020) is a clustering algorithm that defines “modules”, which are clusters formed based on subsets of different data views.”

“…of these review articles have focused specifically on the availability of software tools and discussions of preprocessing considerations (e.g., Arakawa and Tomita, 2013; Misra et al., 2019), or on providing perspective on challenges in multi-omics integration (e.g., Gomez-Cabrero et al.,…”

“Traditional multi-omics considers the integration of several different ‘omics for the same set of samples, whereas the integration considered by BIDIFAC + involves combining different ‘omics and different sample sets.”

“DeepMF (Chen et al., 2019) casts traditional matrix factorization within a deep learning framework to map concatenated multiomics data to a lower-dimensional latent space.”

“Introduced by Das and Mukhopadhyay (2021), TiMEG (Tool for integrating Methylation, gene Expression and Genotype) is a supervised learning approach for the identification of disease-associated biomarkers in case-control multi-omics studies.”

“Joint association and classification analysis (JACA; Zhang and Gaynanova, 2021) is a semi-supervised approach that enables quantification of inter-omic associations with respect to known classifications or data subtypes.”

“Unique from most other approaches described in this review, BIDIFAC + (Lock et al., 2022) is a matrix factorization approach for bi-dimensionally linked matrices.”

“As well, the range of adjusted rand index values (the performance metric used) was typically observed to be narrower for SUMO, suggesting less variability in expected performance.”

“Emphasizing analyses more consistent with a causal framework, iMODA (Lin et al., 2020) proposes a hierarchical regression framework for quantifying (i) the effects of genotypes on quantitative ‘omics, (ii) the effects of quantitative ‘omics on response phenotypes of interest, and (iii) the direct and indirect effects…”

“In untargeted metabolomics and lipidomics false discovery rates are difficult to compute and largely not accurate (Stein, 1994; Jeong et al., 2011; Matsuda et al., 2013; Kim and Zhang, 2014).”

“Last, Ni et al. (2022) introduce scLRTD, an imputation approach based on approximating a third-order tensor representation of the multi-omics dataset with a low-rank approximation.”

“, 2018), Compositional Omics ModelBased Integration (COMBI; Hawinkel et al., 2020) introduces model extensions better suited for modeling compositional and sequence count data.”

“The former results in a reduced sample size while not allowing the full usage of all information contained in the dataset, and the samples with missing values may not be a random subset of the data [101].”

“Moreover, AI and machine learning techniques can be leveraged to analyze the large and complex datasets generated by these methods, revealing hidden patterns and predictive models of plant phenotypes [56–58].”

“Due to the size and complexity of the data produced by various omics technologies, AI and data processing are essential elements in multi ‐ omics analysis [16].”