When Is a Molecule a Duplicate? Identity Policy Determines What a Benchmark Audit Finds
- 1Department of Pharmacy
- 2Daffodil International University
Abstract
Reports of duplicate or redundant records in molecular machine-learning benchmarks are routinely presented as properties of the datasets. They are not. A duplicate count is the output of an operational definition of molecular identity applied to a dataset, and different defensible definitions return different answers from the same file. We audited seven MoleculeNet [3] benchmarks (52,144 records) under a four-level identity hierarchy — canonical molecular graph, fragment parent, charge-normalised parent and tautomer-normalised parent — with the Standard InChIKey as an orthogonal comparator rather than a rung, and with derived identity withheld outside a declared domain of applicability. Duplicate-group counts moved with the level in every benchmark, and for three the audit changed its answer to whether the benchmark contains duplicate compounds at all. Domain choice mattered as much: HIV reports 330 duplicate groups at the tautomer-normalised level when identity is asserted for every record, against 124 on the declared domain. A post hoc sensitivity analysis found that domain under-covered main-group and f-block metal-associated structures, giving 109 without displacing 124 as the prespecified result. Of 295 merge events, 179 were decidable from structure alone and 116 were policy-dependent; across that complete census, 19 events (16.4%; 41 records) received a different benchmark-equivalence decision from transition-appropriate official-InChI arbitration under an endpoint rulebook fixed before final classification, all 19 in one benchmark's stereochemical subset. We publish the identity policy as an executable, fingerprinted contract reproducing every reported count exactly (434/434), so the assumptions behind an audit statistic can be inspected and re-executed rather than inferred.
Related
- Fast Process Development Made Possible Using an Easy to Duplicate Integrated Microfluidic Platform
- A modular mechanism determines the proton-pumping stoichiometry of respiratory complex I
- On how cell shape determines bactericidal effects of nanopatterned surfaces
- Macromolecular association creates dispersity of components that determines liquid–liquid phase separation of oppositely charged polyions
- Pressing Mode Determines Densification Uniformity and Dimensional Stability in Sulfide Solid-State Batteries
BibTeX
@misc{roktim2026when,
title = {{When Is a Molecule a Duplicate? Identity Policy Determines What a Benchmark Audit Finds}},
author = {Md. Rahul Reza Roktim},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009099/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009099/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - When Is a Molecule a Duplicate? Identity Policy Determines What a Benchmark Audit Finds
AU - Md. Rahul Reza Roktim
PY - 2026
DA - 2026/09/20
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009099/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009099/v1
AB - Reports of duplicate or redundant records in molecular machine-learning benchmarks are routinely presented as properties of the datasets. They are not. A duplicate count is the output of an operational definition of molecular identity applied to a dataset, and different defensible definitions return different answers from the same file. We audited seven MoleculeNet [3] benchmarks (52,144 records) under a four-level identity hierarchy — canonical molecular graph, fragment parent, charge-normalised parent and tautomer-normalised parent — with the Standard InChIKey as an orthogonal comparator rather than a rung, and with derived identity withheld outside a declared domain of applicability. Duplicate-group counts moved with the level in every benchmark, and for three the audit changed its answer to whether the benchmark contains duplicate compounds at all. Domain choice mattered as much: HIV reports 330 duplicate groups at the tautomer-normalised level when identity is asserted for every record, against 124 on the declared domain. A post hoc sensitivity analysis found that domain under-covered main-group and f-block metal-associated structures, giving 109 without displacing 124 as the prespecified result. Of 295 merge events, 179 were decidable from structure alone and 116 were policy-dependent; across that complete census, 19 events (16.4%; 41 records) received a different benchmark-equivalence decision from transition-appropriate official-InChI arbitration under an endpoint rulebook fixed before final classification, all 19 in one benchmark's stereochemical subset. We publish the identity policy as an executable, fingerprinted contract reproducing every reported count exactly (434/434), so the assumptions behind an audit statistic can be inspected and re-executed rather than inferred.
ER -