Workflow-dependent conformational landscape fragmentation and quantitative failure modes in 3D molecular embedding of...

The automated conversion of 2D molecular structures into physical 3D Cartesian coordinates is a foundational prerequisite for high-throughput virtual screening, structure-based drug design, and geometric deep learning. However, default 2D-to-3D embedding workflows exhibit operational biases and severe structural failure modes that are rarely quantified across large chemical libraries. Here, we present a systematic benchmark evaluating eight execution strategies across five computational suites-RDKit (two strategies), Open Babel (two strategies), HyperChem 8.0, Schrödinger LigPrep (OPLS4), and Wavefunction Spartan (two strategies)-over 33,715 biologically active compounds. To ensure evaluation on a common molecular-mechanics energy scale, all coordinates were rescored using the same MMFF94s potential energy protocol (obenergy). For each molecule, the portfolio global minimum was defined as the lowest MMFF94s energy identified within the finite collection of geometries produced by the eight evaluated strategies; it does not denote the mathematically exact minimum over the complete conformational space or a model-independent physical minimum. Our findings reveal an extreme potential energy surface (PES) sampling bias: 31,121 molecules (92.31%) had portfolio minima uniquely recovered by one strategy, whereas multi-strategy consensus occurred for 2594 molecules (7.69%). The corrected non-disjoint analysis yielded 37,516 portfolio-minimum recovery occurrences because every strategy satisfying the energy-equivalence criterion received recovery credit. Furthermore, we establish a quantitative taxonomy of structural pathologies, cataloging severe energy anomalies and probable steric overlaps (E>10³kcal/mol), planar coordinate retention or collapse (max_i|q_i|<10^-7Å for at least one Cartesian coordinate q∈{x,y,z}), and macrocyclic ring interpenetrations. The rdkit0 strategy produced the largest portfolio-minimum recovery count, with 13,867 recoveries (41.13%), followed by rdkit1, with 11,109 recoveries (32.95%), and obabel1, with 6482 recoveries (19.23%). Severe energy anomalies were concentrated in obabel0, which produced 1679 structures above the adopted threshold, whereas no such anomalies were detected for obabel1, rdkit0, rdkit1, or schrodinger. These results demonstrate that standard embedding pipelines operate in strongly workflow-dependent conformational basins and highlight the necessity of multi-strategy generation, common energy evaluation, explicit structural quality control, and subsequent conformational expansion for scalable, high-fidelity dataset curation.