A grouped split silently removes the chemistry entangled with its grouping variable: the case of MOF topology hold-out
Year: 2026
Type: preprint
Abstract: Machine learning screens metal–organic frameworks (MOFs) for CO2 capture, and 2026 studies on enumerated databases have converged on topology hold-out as a test of topological transferability, presupposing untested separability of topology and chemistry. They are not: on hMOF (324,426 structures), 74.7% of coarse compositions (61.3% of structures) occur in a single topology, and for seven of hMOF's eleven topologies removing one topology leaves every structure with no chemical counterpart in training; the confound replicates on ARC-MOF (99.3% coarse single-topology), which is not independent of hMOF, so only a leave-one-source-out bound (98.6–99.8%) is claimed. On the isoreticular pcu ↔ nbo pair, removing the matched topology degrades nbo-side MAE by 70.0% while a sizematched control leaves it unchanged, and across the shared 25,000-sample support set the top-1% hit rate falls from 50.4% to 19.3% (five seeds). Yet the topology penalty is not explained by chemistry alone: holding out the composite chemistry key instead retains R2 = 0.674 against topology's 0.379, so the composition counts fix the identity of the withheld chemistry, not its cost. Topology hold-out should therefore not be read as a clean test of topological transferability, and grouped-split reports should first disclose the confound structure and chemical granularity, for which we provide a reporting checklist.
Source: ChemRxiv
Author: Tao Wang
Institution: Nanjing Tech University
Language: English
Subfield: Inorganic Chemistry
Field: Chemistry
Domain: Physical Sciences
Open Access status: green
Funder: Nanjing Tech University
OpenAlex ID: W7223057179