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

May combine two claims.

Identifying these groupings within latent factors is often a post-hoc process, and such domain-relevant groupings are not always identifiable within generated latent factors.

“Identifying these groupings within latent factors is often a post-hoc process, however, and such domain-relevant groupings are not always identifiable within generated latent factors.”

May combine two claims.

A principal challenge to multi-omic integration is missing data because all biomolecules are not measured in all samples.

“A principal challenge to multi-omic integration is missing data because all biomolecules are not measured in all samples.”

Classified as probably not a substantial claim. May combine two claims.

Hinderances to effective integration methods include the heterogeneous nature of data and distributional properties across multi-omics datasets, the typically large difference in the number of measured biomolecules compared to number of samples or replicates, and the complex and often noisy nature of biological data.

“Hinderances to effective integration methods include, but are not limited to, the heterogeneous nature of data and distributional properties across multi-omics datasets, the typically large difference in the number of measured biomolecules compared to number of samples or replicates, and the complex and often noisy nature of biological data.”

Classified as probably not a substantial claim.

Data for a biological sample may be missing for one or more 'omic technologies due to cost, instrument sensitivity, or other experimental factors.

“Due to either cost, instrument sensitivity, or other experimental factors, data for a biological sample may be missing for one or more ‘omic techologies.”

May combine two claims.

Additional complexities including protein 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.

“Furthermore, additional complexities including protein 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).”

• cites Goh and Wong, 2017Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

In 'omics studies, data are more commonly MNAR or MAR, considering that instruments are known to have certain detection thresholds (MNAR) and that biochemical dependencies exist between different biomolecules (MAR).

“In ‘omics studies, data are more commonly MNAR or MAR, considering that instruments are known to have certain detection thresholds (MNAR) and that biochemical dependencies exist between different biomolecules (MAR).”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

If an entire view is missing for one or more samples and must be imputed, any level of similarity specific to that view may be lost.

“If an entire view is missing for one or more samples and must be imputed, any level of similarity specific to that view may be lost.”

Classified as probably not a substantial claim. May combine two claims.

Missing data are common in 'omics data and can arise due to a variety of reasons, such as poor tissue quality, insufficient sample volume, measurement system limitations, budget restrictions, or subject dropout.

“Missing data are common in ‘omics data and can arise due to a variety of reasons, such as poor tissue quality, insufficient sample volume, measurement system limitations, budget restrictions, or subject dropout.”

May combine two claims.

When faced with missing data, these integration methods require a complete case analysis or the implementation of separate imputation steps beforehand.

“When faced with missing data, these methods thus require a complete case analysis or the implementation of separate imputation steps beforehand.”

May combine two claims.

Models incorporating gene-gene and methylation-methylation correlations were considered for FBM but had higher computational cost without substantive performance gains relative to independence models.

“However the authors note that models incorporating gene-gene and methylation-methylation correlations were additionally considered, but these models were associated with a higher computational cost without indicating substantive gains in performance relative to models based on the independence assumption.”

• cites Fang et al., 2018 sourceNot checked yet

  Not checked in the free run.

• AssumptionThe authors state this as an assumption: The FBM model assumes gene-gene and methylation-methylation independence, and imputed values reflect this assumption.

Probably not a substantial claim: may not pass S2 (checkable).

Given the prevalence of missing data across different 'omics, the approach that machine learning and artificial intelligence models use in handling missing observations is of great importance when considering multi-omics integration methods.

“Given the prevalence of missing data across different ‘omics, the approach that ML and AI models use in handling missing observations is of great importance when considering multi-omics integration methods.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

KNN imputation has been shown to work well for numerical imputation on single datasets, but the algorithm is sensitive to the choice of the number of nearest neighbors (k) and is limited in that it does not utilize the potential structure across the entire set of observations.

“KNN imputation has been shown to work well for numerical imputation on single datasets (Jadhav et al., 2019), but the algorithm is sensitive to the choice of the number of nearest neighbors (k) and is limited in that it does not utilize the potential structure across the entire set of observations.”

• cites Jadhav et al., 2019 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Available case analysis, while more robust than complete case analysis since less data are dropped, is still only valid under the MCAR assumption and generates conclusions based on different parts of the data, potentially leading to biased or conflicting inference.

“While more robust than a complete case analysis since less data are dropped, an available case analysis is still only valid under the MCAR assumption and generates conclusions based on different parts of the data, potentially leading to biased or conflicting inference.”

May combine two claims.

Zero-value or LOD imputation, like mean imputation, may distort feature distributions and lead to biased parameter and variability estimates.

“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).”

• cites Lubin et al., 2004 sourceNot checked yet

  Not checked in the free run.

• cites Succop et al., 2004 sourceNot checked yet

  Not checked in the free run.

• cites Webb-Robertson et al., 2015 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Middle integration applies some transformation to represent a complex combination of the datasets before applying downstream analyses.

“Middle integration applies some transformation to represent a complex combination of the datasets before applying downstream analyses.”

• cites Ritchie et al., 2015 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Missing samples in multi-omics data are rarely MCAR, and thus inference based on a complete case analysis is likely to be biased.

“Missing samples in multi-omics data are rarely MCAR, and thus inference based on a complete case analysis is likely to be biased.”

May combine two claims.

Both the Xie et al. (2019) and Xu et al. (2019) methods require complete input, and initial imputation is performed in both cases to accommodate.

“Both these methods require complete input, and initial imputation is performed in both cases to accommodate.”

• cites Xie et al., 2019 sourceNot checked yet

  Not checked in the free run.

• cites Xu et al., 2019Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Kernel-based methods rely on an imputation pre-processing step, which could possibly negatively affect the downstream analyses, as the relationships between nodes/samples/biomarkers may be biased toward similarities in imputation.

“These methods rely on an imputation pre-processing step, which could possibly negatively affect the downstream analyses, as the relationships between nodes/samples/biomarkers may be biased toward similarities in imputation.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Integration techniques generally assume measurements from the same samples on several markers, or features, from several 'omics-data types, which are more generally referred to as data views.

“Integration techniques generally assume measurements from the same samples on several markers, or features, from several ‘omics-data types, which are more generally referred to as data “views”.”

May combine two claims.

Mean imputation preserves feature means and grants access to the use of more data, but does not preserve relationships among variables and may lead to an underestimate of variability.

“While this simple approach preserves feature means and grants access to the use of more data, it does not preserve relationships among variables and may lead to an underestimate of variability (Gelman and Hill, 2006).”

• cites Gelman and Hill, 2006 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

In metabolomics, limited coverage of the known metabolome increases the risk of overlooking the metabolomic response of interest in targeted metabolomic analyses.

“In metabolomics, limited coverage of the known metabolome increases the risk of overlooking the metabolomic response of interest in targeted metabolomic analyses.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Early integration is primarily characterized by an initial concatenation of the features across all measured 'omics, followed by application of methods appropriate for analyzing the resulting high-dimensional dataset.

“Early integration is primarily characterized by an initial concatenation of the features across all measured ‘omics, followed by application of methods appropriate for analyzing the resulting high-dimensional dataset.”

• cites Ritchie et al., 2015 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Late integration analyzes each dataset separately and introduces a model or algorithm that combines the outputs of each individual analysis.

“Late integration analyzes each dataset separately and introduces a model or algorithm that combines the outputs of each individual analysis.”

• cites Ritchie et al., 2015 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Peptides may be missing from one or more samples due to a variety of underlying mechanisms, which can result in non-trivial amounts of potential observations being missing.

“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.”

• cites Daly et al., 2008 sourceNot checked yet

  Not checked in the free run.

• cites Karpievitch et al., 2009 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Given the ubiquity of partially observed cases in multi-omics datasets-and that most integrative multi-omics methods are not inherently compatible with partially observed data-imputation methods are often incorporated as a data pre-processing step within the multi-omics analytic pipeline, or alternatively, partially observed samples are dropped and a complete case analysis is performed.

“Given the ubiquity of partially observed cases in multi-omics datasets—and that most integrative multi-omics methods are not inherently compatible with partially observed data—imputation methods are often incorporated as a data pre-processing step within the multi-omics analytic pipeline, or alternatively, partially observed samples are dropped and a complete case analysis is performed.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Given the tendency of neural networks to overfit to the training data, it is possible that any bias in the imputation of the training data-which is likely due to the common violation of MAR in 'omics data-will be erroneously exploited by the network.

“Given the tendency of neural networks to overfit to the training data, it is possible that any bias in the imputation of the training data—which is likely due to the common violation of MAR in ‘omics data—will be erroneously exploited by the network.”

May combine two claims.

In the MOFA comparison, missing percentages ranged between approximately 10% and 80%, and all methods showed increased mean squared error of prediction with increased missing percentage.

“The missing percentages considered ranged between ~10–~80%, and all methods demonstrated increases to mean squared error of prediction with increased missing percentage.”

• cites Argelaguet et al., 2018 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Complete case analysis, while convenient, results in a decrease in sample size, throws away useful data, is only valid under the MCAR assumption and can otherwise lead to biased estimates.

“This approach, while convenient, results in a decrease in sample size, throws away useful data, is only valid under the MCAR assumption and can otherwise lead to biased estimates (Fang et al., 2018; Hawinkel et al., 2020).”

• cites Fang et al., 2018 sourceNot checked yet

  Not checked in the free run.

• cites Hawinkel et al., 2020 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Multi-omics data are uniquely characterized by several separate, yet related, datasets, which has motivated the development of newer approaches that leverage information across different 'omics datasets to inform imputations.

“Each of the previously described imputation methods have generally been developed within the context of a single dataset, however multi-omics data are uniquely characterized by several separate, yet related, datasets. This aspect of multi-omics has therefore motivated the development of several newer approaches that leverage information across different ‘omics datasets to inform imputations.”

May combine two claims.

iMODA's power to detect effects of interest varies with the proportion of missing data, and power decreases as missingness increases.

“The power of the iMODA framework to detect the effects of interest has been reported to vary according to the proportion of data that are missing. The authors considered cases where 50% and 30% of data were missing, finding that power decreases with increased missingness.”

• cites Lin et al., 2020 sourceNot checked yet

  Not checked in the free run.

• AssumptionRests on an assumption: The known causal hierarchy across biomolecules is that genotype affects omic expression, which influences phenotype.

May combine two claims.

Many methods described were motivated by or evaluated exclusively on sequencing-based omics platforms, and few considered the fuller breadth of omics platforms including metabolomics, proteomics, and lipidomics.

“Related to this point, many of the methods described here were motivated by or evaluated exclusively on sequencing-based omics platforms (e.g., transcriptomics, genomics, and epigenomics) and few considered the fuller breadth of omics platforms that includes metabolomics, proteomics, lipidomics, etc.”

May combine two claims.

Data of non-sequencing omics are typically generated by a mass spectrometer, an instrument whose use introduces several nuances to the data which may mean integration model assumptions reasonably made for sequencing-based data may not hold for mass spectrometry-based data.

“Contrasting with the sequencing-based ‘omics used by many of the methods reviewed in this work, data of other ‘omics are typically generated by a mass spectrometer, an instrument whose use introduces several nuances to the data which may mean integration model assumptions reasonably made for sequencing-based data may not hold for mass spectrometry-based data.”

May combine two claims.

The datasets considered by these approaches contain hundreds, if not thousands, of observations, sample sizes that may be practically prohibitive due to cost, biospecimen availability, or both.

“The datasets considered by these approaches contain hundreds, if not thousands, of observations, sample sizes that may be practically prohibitive due to cost, biospecimen availability, or both.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

MONET has no constraint on the number of detected modules and no upper bound on module size.

“Notably, there is no constraint on the number of detected modules, nor is there an upper bound on module size.”

• cites Rappoport et al., 2020 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

BIDIFAC+ assumes that shared structures are present across the integrated matrices, an assumption that may be violated in practice.

“However, Lock et al. (2022) note that BIDIFAC+ assumes that shared structures are present across the integrated matrices, an assumption that may be violated in practice.”

• cites Lock et al., 2022 sourceNot checked yet

  Not checked in the free run.

• AssumptionThe authors state this as an assumption: BIDIFAC+ assumes that shared structures are present across the integrated matrices.

May combine two claims.

MVAE's product-of-experts model architecture is robust to partially observed samples, combining only the omic-specific encodings of each partial sample's observed omics during model training.

“This PoE model architecture is robust to partially observed samples, combining only the ‘omic-specific encodings of each partial sample's observed ‘omics during model training.”

• cites Wu and Goodman (2018)Not checked yet

  Not checked in the free run.

• AssumptionThe authors state this as an assumption: MVAE assumes that data views are independent from one another, given a shared latent representation.

May combine two claims.

SUMO's range of adjusted rand index values was typically narrower than NEMO's, suggesting less variability in expected performance.

“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.”

• cites Sienkiewicz et al., 2022 sourceNot checked yet

  Not checked in the free run.

• cites Rappoport and Shamir, 2019 sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim. May combine two claims.

MONET's authors propose a heuristic algorithm that identifies a local maximum and recommend using the best solution from repeated runs to better approximate the globally maximal solution.

“The authors circumvent this issue by proposing a heuristic algorithm that identifies a local maximum and recommend using the best solution from repeated runs of the algorithm to better approximate the globally maximal solution.”

• cites Rappoport et al., 2020 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

DeepMF's perfect-accuracy evaluations were completed in a simulated data setting in which data were simulated as missing completely at random (MCAR) and ground truth feature subgroups were generated with simple discrimination boundaries.

“However, it should be noted that these evaluations were completed in a simulated data setting in which data were simulated as missing completely at random (MCAR) and ground truth feature subgroups were generated with simple discrimination boundaries.”

• cites Chen et al., 2019 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

These methods commonly investigate differences in early, middle, and late fusion, and work has been done to address view-missing and view-heterogeneity problems.

“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).”

• cites Nagrani et al., 2018 sourceNot checked yet

  Not checked in the free run.

• cites Wang et al., 2020 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

The criteria of sets of features from distinct views across a shared set of samples is applicable to any domain where the combination of different data sources provides a more complete representation of the system under investigation.

“The criteria of sets of features from distinct views across a shared set of samples is applicable to any domain where the combination of different data sources provides a more complete representation of the system under investigation.”

May combine two claims.

Dependencies exist between different omics data, but important differences between omics must also be accounted for in developed integrative approaches.

“On the other hand, though dependencies exist between different ‘omics data, important differences between ‘omics must also be accounted for in developed integrative approaches.”

May combine two claims.

In untargeted metabolomics and lipidomics, false discovery rates are difficult to compute and largely not accurate.

“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).”

• cites Stein, 1994Not checked yet

  Not checked in the free run.

• cites Jeong et al., 2011 sourceNot checked yet

  Not checked in the free run.

• cites Matsuda et al., 2013 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

MONET incorporates partially observed samples by representing their missing components as edgeless nodes within their corresponding edge-weighted graphs.

“This algorithm incorporates partially observed samples by representing their missing components as edgeless nodes within their corresponding edge-weighted graphs.”

• cites Rappoport et al., 2020 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Low-sample-size cases present the need for methods that generate reliable inference despite smaller sample sizes and avoid overfitting to the data.

“These cases present the need for methods that generate reliable inference despite smaller sample sizes and avoid overfitting to the data.”

Probably not a substantial claim: may not pass S3 (load-bearing).

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.

“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).”

• cites Carreira and Zisserman, 2017 sourceNot checked yet

  Not checked in the free run.

• cites Afouras et al., 2019 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

A study to identify biomarkers for Snyder Robinson syndrome obtained data on 16 patients, only three of which presented with the disorder of interest.

“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).”

• cites Abela et al., 2015 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Manual verification of metabolites following identification scoring is required for shotgun metabolomics because scoring methods are much less mature than for other omics data types.

“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).”

• cites Stein and Scott, 1994Not checked yet

  Not checked in the free run.

• cites Samaraweera et al., 2018 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The larger sample size requirement is implicit because it is rarely stated as a limitation but is implied by the sample sizes of the datasets these methods were successfully applied to.

“We describe this requirement as “implicit” given that it is rarely, if ever, stated as a limitation, but instead implied based on the sample sizes of the datasets that these methods were successfully applied to.”

May combine two claims.

Using AUROC on TCGA and CCLE benchmark datasets, DeepIMV's performance diminished with increased missingness but performed as well as or better than compared approaches.

“Using the area under the receiver-operator curve (AUROC) as a performance metric, the authors found that the performance of DeepIMV diminished with increased missingness but performed as well as or better than the compared approaches.”

• cites Lee and van der Schaar, 2021Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

BIDIFAC+ took nearly 24 hours to reach convergence when fitted to a combined TCGA dataset, and thus may not scale well with larger aggregations of datasets.

“As well, they noted that BIDIFAC+ took nearly 24 h to reach convergence when fitted to a combined TCGA dataset, and thus the method may not scale well with larger aggregations of datasets.”

• cites Lock et al., 2022 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Around 19 samples per group are needed to detect a fold change of 1.5 in microRNA data at 80% power and false discovery rate <10%.

“Kok et al. (2018) notes, for example, that one may need around 19 samples per group to detect a fold change of 1.5 in microRNA data at 80% power and false discovery rate <10%.”

• cites Kok et al., 2018 sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Insights from the omics domain, where some of these issues are most salient, could transfer to other domains.

“Insights from the ‘omics domain, where some of these issues are most salient, could transfer to other domains.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

The methodological choices inherent to the omics domain are usually the ones that offer the ability to constrain or elucidate relationships between biomarkers.

“The methodological choices inherent to the ‘omics domain are usually the ones that offer the ability to constrain or elucidate relationships between biomarkers.”

May combine two claims.

The groupings implicit within the latent factors generated by these models may be contextualized according to domain-specific knowledge, thus increasing their interpretability and allowing one to infer underlying relationships.

“the groupings implicit within the latent factors generated by these models may be contextualized according to domain-specific knowledge (e.g., shared molecular pathways), thus increasing their interpretability and allowing one to infer underlying relationships.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Common methodological choices in the omics domain include principal-component-like representations that might identify important groups of biomarkers, and injection of prior knowledge through techniques that constrain relationships between features, such as directed edges in graph representations, hard parameter constraints, or carefully chosen priors in Bayesian frameworks.

“Some common choices are the use of principal-component-like representations that might identify important groups of biomarkers, and the injection of prior knowledge through techniques that constrain the relationships between features such as directed edges in graph representations, hard parameter constraints, or carefully chosen priors in”

• Also stated in: § Discussion · plain

  “the relationships between features such as directed edges in graph representations, hard parameter constraints, or carefully chosen priors in Bayesian frameworks.”

Probably not a substantial claim: may not pass S2 (checkable).

Incorporating domain knowledge may be key toward bridging the interpretability gap apparent across these methods.

“incorporating domain knowledge may be key toward bridging the interpretability gap apparent across these methods.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

If the generation of these latent factors were instead guided by domain knowledge, resulting groupings may be more likely to hold relevance and rely less on post-hoc analyses for contextualization within the scientific domain.

“If the generation of these latent factors were instead guided by domain knowledge, resulting groupings may be more likely to hold relevance and rely less on post-hoc analyses for contextualization”

• Also stated in: § Discussion · hedged

  “instead guided by domain knowledge, resulting groupings may be more likely to hold relevance and rely less on post-hoc analyses for contextualization within the scientific domain.”