Distinguishing between direct and indirect regulation

Analysis of single-cell protein variation may also enable a long elusive goal: the quantitative characterization of the direct protein interactions that weave the signal transduction networks in our cells. While genomics identifies many causal genetic associations, these associations are indirect [1]. Such indirect associations are mediated by many unobserved molecules and can be explained by very many different models. This multiplicity of models limits the utility of indirect causal associations [1].

By contrast, single-cell measurements of proteins may enable inference of direct regulatory interactions with minimal assumptions, as shown in Fig 1. Analysis of joint distributions of protein abundances can directly reveal regulatory interactions without depending on model assumptions. Specifically, conditioning the joint distribution of proteins X and Y and a confounder protein Z may reveal that while X and Y are correlated, they do not directly regulate each other (Fig 1A). Alternatively, the joint distribution of A and B may not be correlated unless conditioned on protein Z, which suggests a different regulatory model (Fig 1B). Importantly, this analysis does not require assuming specific types of models or functional dependencies. Rather, it requires quantitatively accurate protein measurements across many single cells. This requirement is becoming feasible due the technological advances discussed below.

Technological requirements and frontiers

Learning from the patterns of single-cell protein variation requires accurate single-cell protein measurements. Indeed, ideas described here have previously surfaced in the literature [8], but their implementation has remained limited by the type and accuracy of the available high-throughput single-cell measurements. This is because the data analysis outlined above, such as conditioning joint distributions on confounders, is much less tolerant to measurement noise than the currently popular single-cell analysis methods, such as cell type clustering and dimensionality reduction.

Therefore, realizing these possibilities requires that we advance the technology. First, we must make the state-of-the-art methods widely accessible via detailed protocols, computational resources, and community standards [9]. Second, we must continue to increase the accuracy of the measurements by (i) increasing the sampling efficiency of protein molecules and thus reducing counting errors; (ii) reducing the potential for measurement interferences; and (iii) using control and reference samples that may allow reducing the impact of experimental artifact, as in the case of ratiometric measurements. This second requirement will take advantage of exciting opportunities for innovation that can increase the depth and accuracy of proteome profiling by intelligent data acquisition and parallel analysis of both peptides and single cells [10].

In addition to increasing the accessibility and accuracy, single-cell proteomic technologies will develop toward measuring protein dynamics, activities, and localization. For example, protein synthesis and degradation dynamics may be encoded into the proteomes of single cells by pulsing at different time points amino acids labeled with heavy isotopes. These encoded dynamics can be decoded at the end of the experiment by MS measurements of the abundances of proteins labeled with each isotopic composition. Protein localization may be measured by physically isolating cellular organelles, while posttranslational modifications (PTMs) may be analyzed by using PTM-enriched isobaric carriers [2,9]. Such analysis requires future technological advances, which will be powered by major opportunities for innovation [10].

Acknowledgments

I thank Prof. Alexander Franks for constructive feedback.