Fig. 1.

(A) DeePhase predicts the propensity of proteins to undergo phase separation by combining engineered features computed directly from protein sequences with protein sequence embedding vectors generated using a pretrained language model. The DeePhase model was trained using three datasets, namely two classes of intrinsically disordered proteins with a different LLPS propensity (${\mathrm{L}\mathrm{L}\mathrm{P}\mathrm{S}}^{+}$ and ${\mathrm{L}\mathrm{L}\mathrm{P}\mathrm{S}}^{-}$) and a set of structured sequences (${\mathrm{P}\mathrm{D}\mathrm{B}}^{*}$). (B) To generate the ${\mathrm{L}\mathrm{L}\mathrm{P}\mathrm{S}}^{+}$ and ${\mathrm{L}\mathrm{L}\mathrm{P}\mathrm{S}}^{-}$ datasets, the entries in the LLPSDB database (27) were filtered for single-protein systems. The constructs that phase separated at an average concentration below $c\hspace{0.17em}=\hspace{0.17em}\mathrm{100}\hspace{0.17em}\mathrm{\mu }\mathrm{M}$ were classified as having a high LLPS propensity (${\mathrm{L}\mathrm{L}\mathrm{P}\mathrm{S}}^{+}$; 137 constructs from 77 UniProt IDs) with the remaining 25 constructs together with constructs that had not been observed to phase separate homotypically classified as low-propensity dataset (${\mathrm{L}\mathrm{L}\mathrm{P}\mathrm{S}}^{-}$; 84 constructs from 52 UniProt IDs). (C) The 221 sequences clustered into 123 different clusters [Left, CD-hit clustering algorithm (28) with the lowest threshold of 0.4]. (Right) The 110 parent sequences showed high diversity by forming 94 distinct clusters. (D) The ${\mathrm{P}\mathrm{D}\mathrm{B}}^{*}$ dataset (1,563 constructs) was constructed by filtering the entries in the PDB (29) to fully structured full-protein single chains and clustering for sequence similarity with a single entry selected from each cluster.