Fig. 2: Comparison of predicted CPI contact maps across different models. | Communications Chemistry

Fig. 2: Comparison of predicted CPI contact maps across different models.

From: Generalizable compound protein interaction prediction with a model incorporating protein structure aware and compound property aware language model representations

Fig. 2

A, B For two representative protein-compound complexes, the interaction sites of their experimentally determined structures and structural formula of the compound, the ground-truth contact maps, and the predicted contact maps from Cross–Interaction, PSC–CPI, and GenSPARC are illustrated from left to right. In the structure shown in the left panel, the compound is depicted using a ball-and-stick representation, while the side chains and Cα atoms of the interacting amino acids are represented as sticks. Protein structures and compound structural formulae are annotated with residue numbers and atomic numbers corresponding to the contact map. The contact maps depict interactions between residue sites (proteins) and atomic sites (compounds), with darker hues indicating higher prediction accuracy. The AUPRC scores (higher is better) are displayed for each method. Prothrombin (PDB: 1BCU; A) and macrophage metalloelastase (PDB: 3RTT; B) were used to generate the displayed experimental structures and ground-truth contact maps.

Back to article page