Evidence

Validation

The primary evidence is 240 structures of 33 metals, drawn at random from the Crystallography Open Database (COD) with a fixed seed. None of them was used in development.

The unseen set has 50 structures of Cu, Ni, Co, Zn and Mn, 40 of Fe, Cr, V and Ti, 72 of second- and third-row transition metals, 42 of lanthanides and 36 of main-group metals. No entry was rejected after it had been run; 1 entry has no coordinates. These sets are not blind in the strict sense: defects found on them in earlier measurements were corrected by general rules. Every figure on this page was measured with version 0.6.0, default settings and one seed. The depositors' own charges, oxidation states and bond lists were not read.

Composition

A pass compares composition, not connectivity. COD 1551217 passes the formula check, yet its drawing is wrong (see Limitations).

Metal–ligand bonds

On the unseen set, 94.8 % of the 3283 metal–ligand bonds listed by the depositors are drawn, and 99.4 % of the bonds drawn are listed. The program does not read these lists.

Charges and oxidation states

The structure as read is neutral for 195 of 238 unseen structures (82 %), but for only 40 of 62 polymers. Among the 16 finite structures that are not balanced, the causes are:

• a wrong ligand charge (5);
• a site shared by two metals (2);
• disorder wrongly kept or lost (4);
• cases, mostly metal–metal bonded clusters, in which the authors could not decide the correct charges (5).

Where the compound name states an oxidation state, it agrees with the state assigned for 28 of 28 unseen structures. The program cell2mol completes 54 of the 240 unseen structures. Where both programs give a result, the metal oxidation states agree in 40 of 45; that is agreement between two programs, not with ground truth.

Statuses returned

The status is deliberately cautious. Every polymer is needs_review, because an excerpt is drawn.

Drawings, and other layout programs

Of the 238 unseen structures drawn, 218 (92 %) are clean: no line through a label, no overlapping labels, no hidden bond, no coincident atoms, and no intrusion into or distortion of an aromatic ring. Of the held-out structures 41 of 50 are clean, and of the development structures 44 of 46.

A drawing that breaks no rule is not always readable. 16 unseen drawings have fifteen or more crossings at 0.12 or more per atom, or break three or more rules. Most of them are polyhedral oxo, alkoxide or sulfido cluster cores, which have no planar drawing. The median time was 3.4 s per structure on a laptop computer.

Seven freely available layout programs were each given exactly the molecular graph of the drawing CIFlat had made, including abbreviations and ghost atoms, with metal–ligand bonds as single bonds. Each was allowed 300 s per structure and scored by the same validator. This compares layout alone, and it favours the other programs by sparing them the perception step.

Structure by structure, against the best of the seven, CIFlat was better for 235 unseen structures, equal for 3 and worse for 0. Three things qualify this result:

• the scoring rules are CIFlat's own and measure legibility, not a chemist's preference;
• the other programs are much faster;
• a user starting from a CIF would not obtain these graphs unaided. Given the development CIFs, Open Babel found the same metal–ligand bonds as CIFlat in 2 of 46.

No assessment by chemists has been made, and the comparison covers free programs only.