Construction of representative fragment sets based on mutual 3D structural similarity and docking feasibility for fragment-based virtual screening
Satoshi YONEYAMA *, Keisuke YANAGISAWA, Yutaka AKIYAMA
School of Computing, Institute of Science Tokyo
Purpose: Fragment-based virtual screening (FBVS) has gained attention as a method that decomposes compounds into fragments and reuses computational results for shared substructures, enabling faster and more efficient screening. However, the computational cost of FBVS remains high, and further acceleration is required. To reduce this cost, some studies have proposed selecting representative fragments based on 2D structural similarity. While these approaches improve efficiency, they still suffer from limited accuracy, partly due to the possible inability of 2D similarity to reflect docking-relevant structural features.
Method: To address the issue, we focus on incorporating 3D structure and actual docking scores, which were not considered in prior methods, to improve the accuracy of representative fragment selection in FBVS. We propose two novel similarity metrics to improve the accuracy of FBVS: 3D structural similarity between fragments and docking score-based similarity derived from docking results across multiple proteins. Specifically, the 3D similarity was calculated by generating 3D conformations for each fragment, performing molecular structure alignments, and computing the Tanimoto similarity of volume overlap. For the docking score similarity, we conducted docking simulations against receptor targets in the DUD-E diverse subset and obtained docking scores, which were then treated as feature vectors for similarity calculation. These similarity metrics aim to capture spatial proximity and docking feasibility that are difficult to assess using conventional 2D structure-based similarities. We also developed a method for constructing representative fragment sets using these novel similarity metrics.
Results and Discussion: We first analyzed the correlation between the proposed 3D similarity and docking score differences, both computed between fragment pairs. The analysis revealed a maximum coefficient of determination of r² = 0.55, indicating a significantly stronger correlation compared to conventional 2D similarity metrics (r² = 0.38), suggesting that the 3D similarity more accurately reflects variations in docking scores. Furthermore, we constructed representative fragment sets using the proposed 3D similarity and docking score similarity, and evaluated their performance in virtual screening. Our method was compared with conventional 2D similarity-based approaches, as well as individual approaches using only 3D similarity or docking score similarity. As a result, the overall average AUC increased by 0.175. We also evaluate them with EF10%. These results suggest that integrating 3D structural information and docking score profiles into fragment selection can enhance the effectiveness of FBVS. Moreover, this integration holds promise for developing more generalizable similarity metrics applicable to a broader range of targets.