Many proteins contain disordered regions of low-sequence complexity, which cause aging-associated diseases because they are prone to aggregate. Here, we study FUS, a prion-like protein containing intrinsically disordered domains associated with the neurodegenerative disease ALS. We show that, in cells, FUS forms liquid compartments at sites of DNA damage and in the cytoplasm upon stress. We confirm this by reconstituting liquid FUS compartments in vitro. Using an in vitro "aging" experiment, we demonstrate that liquid droplets of FUS protein convert with time from a liquid to an aggregated state, and this conversion is accelerated by patient-derived mutations. We conclude that the physiological role of FUS requires forming dynamic liquid-like compartments. We propose that liquid-like compartments carry the trade-off between functionality and risk of aggregation and that aberrant phase transitions within liquid-like compartments lie at the heart of ALS and, presumably, other age-related diseases.

Proteins such as FUS phase separate to form liquid-like condensates that can harden into less dynamic structures. However, how these properties emerge from the collective interactions of many amino acids remains largely unknown. Here, we use extensive mutagenesis to identify a sequence-encoded molecular grammar underlying the driving forces of phase separation of proteins in the FUS family and test aspects of this grammar in cells. Phase separation is primarily governed by multivalent interactions among tyrosine residues from prion-like domains and arginine residues from RNA-binding domains, which are modulated by negatively charged residues. Glycine residues enhance the fluidity, whereas glutamine and serine residues promote hardening. We develop a model to show that the measured saturation concentrations of phase separation are inversely proportional to the product of the numbers of arginine and tyrosine residues. These results suggest it is possible to predict phase-separation properties based on amino acid sequences.

The organization of a cell emerges from the interactions in protein networks. The interactome is critically dependent on the strengths of interactions and the cellular abundances of the connected proteins, both of which span orders of magnitude. However, these aspects have not yet been analyzed globally. Here, we have generated a library of HeLa cell lines expressing 1,125 GFP-tagged proteins under near-endogenous control, which we used as input for a next-generation interaction survey. Using quantitative proteomics, we detect specific interactions, estimate interaction stoichiometries, and measure cellular abundances of interacting proteins. These three quantitative dimensions reveal that the protein network is dominated by weak, substoichiometric interactions that play a pivotal role in defining network topology. The minority of stable complexes can be identified by their unique stoichiometry signature. This study provides a rich interaction dataset connecting thousands of proteins and introduces a framework for quantitative network analysis.

Prion-like RNA binding proteins (RBPs) such as TDP43 and FUS are largely soluble in the nucleus but form solid pathological aggregates when mislocalized to the cytoplasm. What keeps these proteins soluble in the nucleus and promotes aggregation in the cytoplasm is still unknown. We report here that RNA critically regulates the phase behavior of prion-like RBPs. Low RNA/protein ratios promote phase separation into liquid droplets, whereas high ratios prevent droplet formation in vitro. Reduction of nuclear RNA levels or genetic ablation of RNA binding causes excessive phase separation and the formation of cytotoxic solid-like assemblies in cells. We propose that the nucleus is a buffered system in which high RNA concentrations keep RBPs soluble. Changes in RNA levels or RNA binding abilities of RBPs cause aberrant phase transitions.

Despite our rapidly growing knowledge about the human genome, we do not know all of the genes required for some of the most basic functions of life. To start to fill this gap we developed a high-throughput phenotypic screening platform combining potent gene silencing by RNA interference, time-lapse microscopy and computational image processing. We carried out a genome-wide phenotypic profiling of each of the ,21,000 human protein-coding genes by two-day live imaging of fluorescently labelled chromosomes. Phenotypes were scored quantitatively by computational image processing, which allowed us to identify hundreds of human genes involved in diverse biological functions including cell division, migration and survival. As part of the Mitocheck consortium, this study provides an in-depth analysis of cell division phenotypes and makes the entire high-content data set available as a resource to the community.To target the ,21,000 protein-coding genes in the human genome, we used a chemically synthesized short interfering RNA (siRNA) library designed to uniquely target each gene with 2-3 independent sequences (Supplementary Methods). The siRNAs in this library were tested individually and reduced the messenger RNAs of targeted genes to below 30% of original levels (to an average of 13%) for 97% of more than 1,000 genes tested (Supplementary Table 1). To allow high-throughput phenotyping of each individual siRNA in triplicates by live-cell imaging, we used a previously established workflow for solid-phase transfection using siRNA microarrays coupled to automatic time-lapse microscopy 1 . As a high-content phenotypic assay we chose to monitor fluorescent chromosomes in a human cell line stably expressing core histone 2B tagged with green fluorescent protein (GFP) 1 . After seeding on the siRNA microarrays, on average 67 (630) cells for each siRNA of the library were imaged in triplicates for 2 days, thus documenting many of their basic functions such as cell division, proliferation, survival and migration. Image processing reveals mitotic hitsThis resulted in a large data set of ,190,000 time-lapse movies providing time-resolved records of over 19 million cell divisions. To automatically score and annotate phenotypes in this large data set, we developed a computational pipeline 2 ( Fig. 1) extending previously established methods of morphology recognition by supervised machine learning [3][4][5][6] . In brief, after segmentation, about 200 quantitative features were extracted from each nucleus and used for classification into one of 16 morphological classes ( Fig. 1 and Supplementary Movies 1-30) by a support vector machine classifier previously trained on a set of ,3,000 manually annotated nuclei (Supplementary Methods). This classifier automatically recognizes changes in nuclear morphology due to the cell cycle, cell death or other phenotypic changes with an overall accuracy of 87% (Supplementary Fig. 1) and allows us to convert each time-lapse movie into a phenotypic profile that quantifies the response to each siRNA ...