Software for building transition metal complexes in 3D and predicting their activity in cancer cell assays

Snowdrop Science Ltd develops software for building transition metal complexes atom by atom, preparing their 3D structures with quantum chemical calculations and predicting their activity in cancer cell assays from those structures. The work so far covers ruthenium(II) arene complexes [12, 14] and copper complexes [1], two classes of compounds studied as anticancer agents. The software is in development, and every result on this site is a retrospective test against published measurements.

0:00 / 0:38
Video. Building the copper(II) complex of 3-methoxypyrithione, [Cu(3-OMe-PT)2], in the Snowdrop builder. The recording has been edited for length. The atoms are placed and bonded, and formal charges are set on the nitrogen, oxygen and sulfur atoms of each ligand. Clean up then arranges the two ligands around copper, and Minimise optimises the structure. The final part shows the atom inspector for the copper atom, which lists a square planar centre with two Cu–O bonds of 1.94 Å and two Cu–S bonds of 2.25 Å, and then for one pyrithione nitrogen atom. The complex was reported by Mishra et al. [1].

Software

Building

Atoms and bonds are placed by hand in a 3D view. Each structure records chelating ligands, dative metal–ligand bonds and formal charges, together with the hapticity of any ligand bound to a metal through several atoms, such as the η6 arene in a ruthenium half-sandwich complex.

Clean up

The Clean up command positions the atoms using structural chemistry rules. These set valence shell electron pair repulsion (VSEPR) geometries, close aromatic rings, place chelating ligands and arrange the donor atoms around the metal centre. Two examples are the square planar copper centre in the video and an η6 arene seated on one face of an octahedral ruthenium centre. The result is a starting geometry for the Minimise command.

Minimise

The Minimise command optimises the structure with the semi-empirical method GFN2-xTB in implicit water, using the ALPB solvation model [15, 16]. A result is accepted only if the optimisation converges and no bond is broken or formed.

Export

Each exported structure records, for every atom, whether its coordinates came from Clean up or from Minimise. Before a file is saved under an existing compound name, the software checks that the molecule’s bond graph matches the graph recorded for that name.

Activity prediction

The prediction models take the 3D structure of a complex as input and estimate its IC50 in a cell assay. Each model is trained on published measurements for one metal and one cell line, and it is tested on complexes that were held out of training. Two datasets have been used so far. One holds 26 copper complexes measured in MCF-7 breast cancer cells and the other holds 49 ruthenium complexes measured in A2780 ovarian cancer cells.

Results so far

Copper activity prediction

Twenty copper complexes with published MCF-7 IC50 values were each predicted by models that had not been trained on them. The best 3D model had a mean absolute error (MAE) of 0.351 in pIC50, which corresponds to a factor of about 2.25 in IC50. A random forest scored 0.426, and a baseline that always predicts the median of the training compounds scored 0.602. Copper results

Ruthenium activity prediction

The same kind of test on 31 ruthenium complexes with published A2780 IC50 values gave an MAE of 0.330 for the best 3D model, 0.405 for a random forest and 0.422 for a baseline that predicts the mean of the training compounds. These are averages over three ways of dividing the data into training and held-out groups. On a fourth, the random forest did as well as the 3D models. Ruthenium results

Geometry optimisation

Thirteen ruthenium complexes were optimised with GFN2-xTB starting from their crystal structures. The median heavy-atom root-mean-square deviation (RMSD) between the optimised and crystal structures was 0.12 Å, and the median differences in Ru–N and Ru–Cl bond lengths were 0.023 and 0.022 Å. Geometry results

The copper and ruthenium errors come from different cell lines and assays, so they cannot be compared with each other. No prediction has yet been checked against a new measurement. The Methods page describes how the tests were run and lists the sources of all the data.

Company

Snowdrop Science Ltd is a private limited company registered in England and Wales, company number 16848046, incorporated in November 2025.

Jed Walsh founded the company and is its director. In his MChem research at the University of Leeds, he developed an inline method for determining the encapsulation efficiency of RNA lipid nanoparticles. In 2024 he took second place in the Reckitt Sustainability Challenge. The name Snowdrop Science is a dedication to his grandmother. The chemotherapy drug she received had poor selectivity for cancer cells, and that experience led him to work on metal-based anticancer compounds beyond platinum that could act more selectively.

Jed Walsh on LinkedIn

Contact

For questions about the software or the results, or to discuss working together, email jed.walsh@snowdropscience.co.uk.