As the cost of high-throughput biological assays fall, it has become more common to perform multimodal data integration, where different assays are performed on the same subjects and/or experimental conditions and the data analyzed together. The assumption is that such integrations would give us additional confidence on the biological insights, or perhaps even discover new ones.

Assumptions and caveats

Data integration presumes that the sum is better than its parts. For example, suppose you have an experiment where cells are treated with a drug, along with an unexposed control, and transcript and proteome data are collected. In the transcriptomic data, you see a number of genes being over-expressed in the treated group. Similarly in the proteomic data, you see the proteins from these genes also being over-expressed in the treated cells. So the two datasets are responding in the same direction: the exposed cells are up-transcribing some genes, then all those transcripts are being made into more proteins compared to the control.

However, it’s just as likely that the response in the two datasets are not in the same direction, at least not across all features (matched gene-protein in this case). Maybe only 80% of the up-regulated genes have increased protein counts, while the remaining have similar counts, or maybe even decreased protein counts. What does that mean ?

It is important to ensure that the conclusions you are drawing are not confounded by biological or statistical problems. Some questions to ask yourself:

• Were the samples processed in comparable ways across different assays ? Were all samples processed identically within the same assay ?
• Do the assays have the same sensitivity and dynamic range ? A large change in one assay may be undetectable in the other.
• Are the datasets normalized in the same way ? Are the same samples/features outliers in both datasets ?
• Do the timescales make sense ? Biological processes take time but many experiments are snapshots of cell states and your sampling may have missed the response.

If you have ruled all these problems out, perhaps your data has generated some new insights.

Approaches and implementations

There are multiple ways to integrate data from multiple assays. In order of increasing sophistication:

Conceptual overlap

This approach involves simply comparing individual genes or proteins that share the same function and/or response across different datasets. Many of the assumptions and caveats above apply.

Manually modeling the datasets together

The conceptual overlap approach above isn’t very sophisticated. For example, we only get which predictors may be relevant, not their relative importance. This can be addressed by manually specifying a linear model that covers both datasets, which would also give you quantitative results like confidence intervals and p-values. The usual modeling caveats of picking the correct underlying distributions, data normalization and scaling, and correctly specifying covariates apply.

Factor analysis

The final approach models the variance explained by the different datasets as latent factors, conceptually similar to principal component analysis. There are multiple software packages that implement this: mCIA, MOFA, SPEAR and Seurat, originally conceived to analyze single-cell RNASeq data. Similar to PCA, one caveat of this approach is that the factors explains the data numerically, and it can difficult to assign them biological meaning.