MoleculeACE is a widely used benchmark that annotates activity-cliff compounds in 30 ChEMBL datasets and reports that machine-learning models predict them poorly. Its “soft consensus” criterion admits a compound pair when any of three similarities reaches 0.9: ECFP4 Tanimoto similarity, Tanimoto similarity of anonymized molecular graphs, or normalized Levenshtein similarity between SMILES strings. The first two are computed after parsing; the third operates on the supplied strings, which the benchmark code does not canonicalize. This study held molecules, potencies, thresholds and code fixed and varied only how the same molecules werewritten. A re-implementation reproduced the distributed labels exactly in 27 of 30 datasets, and canonical round-tripping changed amedian of 0.00% of labels. Re-serializing themolecules with two independent cheminformatics toolkits, with every rewritten string verified to encode the original molecule, changed a median of 8.98% (Indigo) and 9.48% (Open Babel) of compound labels; Kekulé SMILES from RDKit changed 5.23%, and randomized SMILES 17.05%. The graph-derived criteria were invariant across all rewrites. Because MoleculeACE stratifies its train/test split on the cliff label, changing the labels caused split membership to approach that of an independent redraw (31.3% of compounds reassigned, against a 32.21% ceiling). An end-to-end repair, with serialization-invariant labels, re-derived splits and retrained descriptorbased models, moved the pooled cliff penalty by −0.006 to +0.020 log units; every confidence interval included zero against a minimum detectable effect of 0.041–0.057, and for SVM the shift fell within the variability of same-definition split redraws of comparable magnitude. The distributed annotation is therefore reproducible but not invariant to valid serialization. Benchmarks that retain string-based criteria should specify and enforce the canonicalization contract (toolkit, version and output mode); otherwise, labels should be defined on parsed molecular representations. Serialization-invariant labels, splits and per-molecule predictions are released.

In cheminformatics and molecular machine learning, frozen embeddings from pretrained SMILES encoders are an attractive low-cost feature source for property prediction, but their practical value depends on whether they improve over strong classical baselines. A common frozen-feature workflow is evaluated in which ChemBERTa or MoLFormer embeddings are cached once and then paired with inexpensive downstream classifiers for ADMET and CNS bioactivity classification. The comparators are Morgan fingerprints and RDKit descriptors, which remain strong practical cheminformatics baselines when combined with random forests, XGBoost, or logistic regression. To avoid conflating representation quality with downstream classifier choice, each representation family is paired with logistic regression, random forest, and XGBoost classifiers where available. The revised benchmark contains 2580 completed fits across 8 curated binary tasks, 12 model configurations, random splits, scaffold splits, and low-data training budgets of 50, 100, 250, and 500 molecules. Frozen-vs-classical family comparisons are available for 43 dataset-regime rows; in this retrospective oracle comparison, the best classical baseline wins 36 rows and the best frozen embedding wins 7 rows. Across 215 paired seed-level comparisons, the mean AUROC difference for best frozen embedding minus best classical baseline is -0.026 with a 95% bootstrap confidence interval of [-0.032, -0.021]. A clustered bootstrap over dataset-regime rows gives a similar mean difference of -0.025 [-0.032, -0.018]. The average gap is modest but does not support default replacement of Morgan/RDKit baselines by frozen SMILES embeddings. These results provide a practical benchmarking recommendation for small-data cheminformatics and ADMET modeling workflows

How much physiological information does the geometry of a cortical neuron's dendritic arbor carry, and does the answer differ between mouse and human? We re-derived arborization morphometrics for all 701 three-dimensional reconstructions in the Allen Cell Types Database (537 mouse, 164 human), validated the re-derivation against the released summaries, and linked morphology to intrinsic physiology cell by cell, treating the donor - not the cell - as the unit of inference. Coupling between total dendritic length and input resistance was negative in both released samples but far stronger among the human reconstructions (rho = -0.87, 95% CI -0.91 to -0.80; n = 162 cells, 35 donors) than among the mouse reconstructions (rho = -0.25, CI -0.38 to -0.12; n = 236 cells, 163 donors). Distribution matching and common-support restriction attenuate roughly half of the species contrast in Fisher-z units, while direct donor-aware contrasts at the shared restriction confirm that a substantial residual difference remains (Fisher-z difference -0.42 to -0.57; all intervals exclude zero), with wide uncertainty at the strictest restriction (human -0.67; n = 49 cells, 18 donors). Under comparable sampling distributions, the mouse gain in prediction from arbor geometry beyond size (+0.34) fell to +0.03 to +0.06 within mouse excitatory cells, and the human value (-0.05) was not stable under restriction; apical predominance was not supported. These findings quantify how sampling composition and morphological range shape species differences in this release; the residual coupling-magnitude contrast remains an observational comparison between released samples that differ in preparation and size range. Reproducible analysis code is openly archived with the study.