Fig. 7: The impact of incorporating information entropy into the high-energy atom detector (HEAD) approach for identifying abnormal molecular conformations.
From: Assessing conformation validity and rationality of deep learning-generated 3D molecules
a–d Examples of conformations identified as having high information entropy. In each panel, the molecular 3D conformation is shown as sticks, with the information entropy denoted by H, and accompanied by a bar chart of atomic energies (atom names on the horizontal axis and energy values on the vertical axis). Atoms with relatively high energies that contribute to elevated information entropy are highlighted by red bars and indicated by red arrows. Details of the identification procedure are described in the Methods section (Information Entropy). e Weighted F1 scores demonstrating the performance trade-off of introducing entropy-based metric. While the discriminative power of HEAD improves for conformations with low ΔE, there is a corresponding decrease in weighted F1 scores for conformations with high ΔE. The performance of PoseBusters8 is provided for reference.
