Comparison of Biomolecular Condensate Localization and Protein Phase Separation Predictors

Abstract

Research in the field of biochemistry and cellular biology has entered a new phase due to the discovery of phase separation driving the formation of biomolecular condensates, or membraneless organelles, in cells. The implications of this novel principle of cellular organization are vast and can be applied at multiple scales, spawning exciting research questions in numerous directions. Of fundamental importance are the molecular mechanisms that underly biomolecular condensate formation within cells and whether insights gained into these mechanisms provide a gateway for accurate predictions of protein phase behavior. Within the last six years, a significant number of predictors for protein phase separation and condensate localization have emerged. Herein, we compare a collection of state-of-the-art predictors on different tasks related to protein phase behavior. We show that the tested methods achieve high AUCs in the identification of biomolecular condensate drivers and scaffolds, as well as in the identification of proteins able to phase separate in vitro. However, our benchmark tests reveal that their performance is poorer when used to predict protein segments that are involved in phase separation or to classify amino acid substitutions as phase-separation-promoting or -inhibiting mutations. Our results suggest that the phenomenological approach used by most predictors is insufficient to fully grasp the complexity of the phenomenon within biological contexts and make reliable predictions related to protein phase behavior at the residue level.

Shift of the insoluble content of the proteome in the aging mouse brain

“…S6 H ) ( 41 , 42 ). We recently established several computational tools to predict localization of proteins in MLOs, such as the Membraneless organelle and Granule Z-Score (MaGS) ( 31 , 32 , 43 ). Accordingly, proteins in the young dataset have strikingly higher MaGS values relative to both the proteome and the old dataset indicating that they are more likely to be found in MLOs ( Fig.…”

Accurate Predictions of Liquid-Liquid Phase Separating Proteins at Single Amino Acid Resolution

“…After we encode sequence-and structure-based features into a vector, we train multiple binary classifiers and we select the best performing model, which we call ROBOT, to define a LLPS propensity score for a protein. We tested catGRANULE 2.0 ROBOT on proteins belonging to different condensates in human and in different organisms and we show that it outperforms previous methods, both based on sequence features [7,30,31,59] but also on structural features [23]. Moreover, we provide an orthogonal validation of our predictions using thousands of antibody-based immunofluorescence (IF) confocal microscopy images obtained from the Human Protein Atlas [63].…”

“…One of the first LLPS predictors, catGRANULE 1.0, computes the protein propensity for granule formation based on structural disorder and nucleic acid-binding propensities [7]. Following this, the MaGS method was developed using a variety of features including protein abundance, intrinsic disorder percentage, phosphorylation site annotations, PScore, Camsol score, RNA interaction, and the composition of leucine and glycine [30,31]. A more recent method is PICNIC, which uses both sequence-based and structure-based features derived from AlphaFold2 models, focusing on sequence complexity, disorder scores, and amino acid co-occurrences [23].…”

“…The extensive training dataset used in catGRANULE 2.0 ROBOT comprises human proteins documented to undergo LLPS, sourced from various databases and resources [31,39,43,53,65,68,75,76]. It also includes a selection of negative proteins-those highly unlikely to undergo phase separation, specifically excluding known interactors of LLPS proteins [45].…”

RNA: The Unsuspected Conductor in the Orchestra of Macromolecular Crowding

“…The role of proteins in the formation of RNP condensates may be determined, with high efficiency, by analyzing their amino acid sequence and identifying specific regions such as low complexity regions or RNA binding domains. 76 However, this kind of approximation is not feasible with RNA, since all molecules are made up of only four different building block nucleotides, whereas proteins are chains of up to 20 different amino acids able to generate many different combinations. Nevertheless, the elements modulating the RNA interactivity go further than just the four-nucleotide code (Figure 2).…”