Equivalent Distance Geometry Error for Molecular Conformation Comparison
This addresses a bottleneck in drug design tasks like 3D-QSAR and virtual screening by improving conformation generation models, though it appears incremental as it builds on existing backbones.
The paper tackles the problem of poor local geometry in molecular conformation generation by proposing Equivalent Distance Geometry Error (EDGE), a loss function that equivalently optimizes bond lengths, angles, and dihedral angles, and experiments show it performs effectively and efficiently compared to existing methods.
Straight-forward conformation generation models, which generate 3-D structures directly from input molecular graphs, play an important role in various molecular tasks with machine learning, such as 3D-QSAR and virtual screening in drug design. However, existing loss functions in these models either cost overmuch time or fail to guarantee the equivalence during optimization, which means treating different items unfairly, resulting in poor local geometry in generated conformation. So, we propose Equivalent Distance Geometry Error (EDGE) to calculate the differential discrepancy between conformations where the essential factors of three kinds in conformation geometry (i.e. bond lengths, bond angles and dihedral angles) are equivalently optimized with certain weights. And in the improved version of our method, the optimization features minimizing linear transformations of atom-pair distances within 3-hop. Extensive experiments show that, compared with existing loss functions, EDGE performs effectively and efficiently in two tasks under the same backbones.