Methods and references
This page describes the data behind the Results, how the predictions were tested and what the tests leave open. The design of the 3D models is not published at this stage.
Activity data
All activity values are IC50 values from published studies. Snowdrop Science has made no measurements of its own. Each value is converted to pIC50, which is 6 minus the base-10 logarithm of the IC50 in µM. A compound with an IC50 of 1 µM has a pIC50 of 6, and a more potent compound has a higher value.
The copper dataset holds 26 complexes measured in MCF-7 cells, from ten studies [1–10]. These studies come from different laboratories, and their assay conditions differ. The ruthenium dataset holds 49 complexes measured in A2780 cells, from two studies that used the same assay [12, 13]. The two datasets are kept separate in every result on this site.
Each complex is counted once, however many structures or repeat measurements exist for it. Where a study reported only a limit, such as an IC50 above 100 µM, the limit is kept as a limit and is never turned into a value.
Structures
The 3D structures come from published crystal structures and from structures built in the software and optimised with GFN2-xTB in implicit water [15, 16]. The geometry results show how closely GFN2-xTB reproduces 13 ruthenium crystal structures.
Several of the copper complexes can exist as cis and trans isomers. Their published IC50 values were measured on bulk samples, and each complex is represented in the dataset by one isomer. The isomer used for prediction may therefore differ from the form that was active in the assay.
How the predictions are tested
Every error on this site comes from complexes that the model did not see during training. The dataset is divided into groups. Each group is held out in turn, and models trained on the remaining complexes predict it. Any scaling of the inputs or the activity values is set from the training complexes alone. The copper predictions are the average of three models trained from different random starting points.
Two measures are reported. The mean absolute error (MAE) is the average difference between predicted and measured pIC50. Because pIC50 is a logarithm, an MAE of 0.351 corresponds to a factor of 100.351, about 2.25, between predicted and measured IC50. This factor is a geometric mean over the held-out complexes, and individual predictions can be much further out. The second measure is the number of pairs of held-out complexes that a model places in the correct order of activity.
Each 3D model is compared with two baselines on the same held-out complexes. The first predicts the median or mean pIC50 of the training compounds for every complex. The second is a random forest [11] trained on summary descriptors of each complex. The comparison shows whether the 3D structure adds anything beyond these simpler methods.
Limits of these results
The datasets are small, with 26 and 49 complexes. All of the tests are retrospective. The measured values already existed when the models were built, and the test in Figure 2 of the Results was chosen for display after its outcome was known.
The same complexes appear in several held-out groups and the training sets overlap, so the comparisons between methods are correlated [25]. No significance test has been applied, and small differences between methods in the tables may be due to chance.
An IC50 measured in one cancer cell line describes cytotoxicity in that assay. It gives no information about selectivity for cancer cells, safety or activity in patients.
A stronger test is to predict complexes before they are measured. That has not yet been done.
References
References 1 to 10 are the sources of the copper activity data, and their numbers match those printed in the copper figures.
- 1A. Mishra et al., Chem. Sci. 2025, 16, 21104. doi.org/10.1039/D4SC06628F
- 2M. P. Kesavan et al., J. Photochem. Photobiol. B 2017, 167, 20. doi.org/10.1016/j.jphotobiol.2016.11.024
- 3S. Stocchetti et al., RSC Med. Chem. 2024, 15, 4180. doi.org/10.1039/D4MD00543K
- 4M. Salimi et al., BioMetals 2015, 28, 267. doi.org/10.1007/s10534-015-9820-5
- 5A. Das and M. Sankaralingam, Dalton Trans. 2024, 53, 14364. doi.org/10.1039/D4DT01075B
- 6L. M. Balsa et al., ChemMedChem 2022, 17, e202100520. doi.org/10.1002/cmdc.202100520
- 7M. Rocha et al., New J. Chem. 2019, 43, 18832. doi.org/10.1039/C9NJ04975D
- 8P. Levín et al., Inorg. Chem. Front. 2021, 8, 3238. doi.org/10.1039/D1QI00233C
- 9Y. Burgos-López et al., Inorg. Chim. Acta 2019, 487, 31. doi.org/10.1016/j.ica.2018.11.039
- 10Y. Burgos-López et al., Polyhedron 2022, 213, 115621. doi.org/10.1016/j.poly.2021.115621
- 11L. Breiman, Mach. Learn. 2001, 45, 5. doi.org/10.1023/A:1010933404324
- 12R. E. Aird et al., Br. J. Cancer 2002, 86, 1652. doi.org/10.1038/sj.bjc.6600290
- 13A. Habtemariam et al., J. Med. Chem. 2006, 49, 6858. doi.org/10.1021/jm060596m
- 14R. E. Morris et al., J. Med. Chem. 2001, 44, 3616. doi.org/10.1021/jm010051m
- 15C. Bannwarth, S. Ehlert and S. Grimme, J. Chem. Theory Comput. 2019, 15, 1652. doi.org/10.1021/acs.jctc.8b01176
- 16S. Ehlert, M. Stahn, S. Spicher and S. Grimme, J. Chem. Theory Comput. 2021, 17, 4250. doi.org/10.1021/acs.jctc.1c00471
- 17W. Kabsch, Acta Crystallogr. A 1976, 32, 922. doi.org/10.1107/S0567739476001873
- 18N. Mardirossian and M. Head-Gordon, J. Chem. Phys. 2016, 144, 214110. doi.org/10.1063/1.4952647
- 19F. Weigend and R. Ahlrichs, Phys. Chem. Chem. Phys. 2005, 7, 3297. doi.org/10.1039/B508541A
- 20D. Andrae, U. Häußermann, M. Dolg, H. Stoll and H. Preuß, Theor. Chim. Acta 1990, 77, 123. doi.org/10.1007/BF01114537
- 21A. V. Marenich, C. J. Cramer and D. G. Truhlar, J. Phys. Chem. B 2009, 113, 6378. doi.org/10.1021/jp810292n
- 22Q. Sun et al., J. Chem. Phys. 2020, 153, 024109. doi.org/10.1063/5.0006074
- 23X. Wu et al., WIREs Comput. Mol. Sci. 2025, 15, e70008. doi.org/10.1002/wcms.70008
- 24CCDC 170362, Experimental Crystal Structure Determination, Cambridge Crystallographic Data Centre, 2001. doi.org/10.5517/cc5q8kp
- 25Y. Bengio and Y. Grandvalet, J. Mach. Learn. Res. 2004, 5, 1089. jmlr.org/papers/v5/grandvalet04a.html