Highly Multiplexed Imaging Uncovers Changes in Compositional Noise within Assembling Focal Adhesions

Abstract

Integrin adhesome proteins bind each other in alternative manners, forming within the cell diverse cell-matrix adhesion sites with distinct properties. An intriguing question is how such modular assembly of adhesion sites is achieved correctly solely by self-organization of their components. Here we address this question using high-throughput multiplexed imaging of eight proteins and two phosphorylation sites in a large number of single focal adhesions. We found that during the assembly of focal adhesions the variances of protein densities decrease while the correlations between them increase, suggesting reduction in the noise levels within these structures. These changes correlate independently with the area and internal density of focal adhesions, but not with their age or shape. Artificial neural network analysis indicates that a joint consideration of multiple components improves the predictability of paxillin and zyxin levels in internally dense focal adhesions. This suggests that paxillin and zyxin densities in focal adhesions are fine-tuned by integrating the levels of multiple other components, thus averaging-out stochastic fluctuations. Based on these results we propose that increase in internal protein densities facilitates noise suppression in focal adhesions, while noise suppression enables their stable growth and further density increase-hence forming a feedback loop giving rise to a quality-controlled assembly.

Fig 1. High-throughput CycIF imaging of cell-matrix adhesion sites.

(a) Imaging procedure and an example of the images obtained for a cell. Scale bar, 10 μm. (b) Top, the mean Pearson correlation coefficients (r, n = 6 datasets; see S2 Table) between the densities of the components. Bottom, superposition of these correlations with the reported vertical positions of the components across focal adhesions [26] (left) and a comparison with their reported physical associations in the cytosol [33] (right).

Fig 2. Inferring changes in noise levels in the molecular content of focal adhesions.

(a) An assembly process with competing binding interactions. (b) Higher diversity in the local levels of a recruiting protein leads to a stronger correlation between the recruited proteins, while higher noise causes the opposite. (c) Simulated CV and r² of the densities of the dark-blue and red components as a function of binding noise and diversity in the density of the yellow component among focal adhesions. (d) Inferring changes in noise levels based on ΔCV and Δr². Changes between focal adhesion categories exemplified in (c) are indicated. The inference approach was validated by systematic screen of diversity and noise levels for competitive and non-competitive assembly processes.

Fig 3. Noise decrease in focal adhesions is coupled to their size and internal density independently.

(a) The workflow of inferring changes in noise levels between two categories of focal adhesions (e.g. small versus big focal adhesions). (b) Changes in noise levels as a function of focal adhesions area. (c) Changes in noise levels as a function of focal adhesions age. (d) Focal adhesions were sub-categorized according to both their area and age. Changes in noise levels were inferred within each age category as a function of area and vice versa. (e) Same as (d), using eccentricity instead of age. (f) Same as (d), using density instead of age. Error bars indicate standard error of the mean between datasets (see S1 Fig and S2 Table).

Fig 4. Density-dependent high-order statistical relations between components in focal adhesions.

(a) The artificial neural network architecture, exemplified for the case of four input proteins. (b) The coefficients of determinations between the predicted and observed densities of the target component as obtained by artificial neural networks versus Random Forests, shown for all possible combinations of input and target proteins. (c) A scatter plot showing the extent of high-order relations identified for the indicated target components. The horizontal axis indicates the average number of input components in the identified high-order relations. The vertical axis indicates the average coefficient of determination between the predicted levels of the target protein, based on the artificial neural network analysis, and its actual levels in the focal adhesions. High-order relations with average coefficient of determination lower than 0.6 are omitted from the plot. The diameter of the circles indicates the number of identified high-order relations (see Materials and Methods). (d) A positive feedback model for the emergence of noise suppression in focal adhesions.