Sequence-Based Prediction of Protein Phase Separation: The Role of Beta-Pairing Propensity

Abstract

The formation of droplets of bio-molecular condensates through liquid-liquid phase separation (LLPS) of their component proteins is a key factor in the maintenance of cellular homeostasis. Different protein properties were shown to be important in LLPS onset, making it possible to develop predictors, which try to discriminate a positive set of proteins involved in LLPS against a negative set of proteins not involved in LLPS. On the other hand, the redundancy and multivalency of the interactions driving LLPS led to the suggestion that the large conformational entropy associated with non specific side-chain interactions is also a key factor in LLPS. In this work we build a LLPS predictor which combines the ability to form pi-pi interactions, with an unrelated feature, the propensity to stabilize the β-pairing interaction mode. The cross-β structure is formed in the amyloid aggregates, which are involved in degenerative diseases and may be the final thermodynamically stable state of protein condensates. Our results show that the combination of pi-pi and β-pairing propensity yields an improved performance. They also suggest that protein sequences are more likely to be involved in phase separation if the main chain conformational entropy of the β-pairing maintained droplet state is increased. This would stabilize the droplet state against the more ordered amyloid state. Interestingly, the entropic stabilization of the droplet state appears to proceed according to different mechanisms, depending on the fraction of "droplet-driving" proteins present in the positive set.

Keywords: amyloid aggregates; liquid-liquid phase separation; protein droplets.

Figure 1

Schematic diagram showing $\beta$-pairings between two amino acid chains for parallel and anti-parallel orientations. The pairing amino acids are connected by yellow lines. The register shifts in both cases are indicated below the diagrams. In this work the two chains share the same sequence.

Figure 2

Pictorial illustration of (a) sequence partition according to classifier outcome; (b) different examples of ROC curves.

Figure 3

Normalised distributions for the values of AUC on the (a) training set and the (b) test set as obtained from the cross validation procedure. LLPS is the positive set and hsnLLPS is the negative set. The AUC values for PScore shown here were obtained on the test sets of cross validation.

Figure 4

ROC curves corresponding to different predicting scores. The dashed line represents the random classifier. (a) LLPS positive set, hsnLLPS negative set. (b) PP positive set, hsnLLPS negative set. (c) LLPS positive set, nsLLPS negative set. (d) PP positive set, nsLLPS negative set.

Figure 5

Correlation plot between the PASTA energy E and the score $s_4$. Each point represents a sequence in the positive set. (a) LLPS positive set. (b) PP positive set.