The same compound drawn as a different tautomer looks like a different compound to software. Calculated cLogP and TPSA also change with the form you draw. When you put several compounds on one screen and ask the AI about them, you also need a way to tell it which compound you mean. And an SDF taken from PubChem or a similar source can open with its structures piled on top of each other.
This time I introduce three features I added to ChemOmni (canonical tautomers, compound numbers, and SDF import and export). In all of them the AI writes no structures and works with results calculated by RDKit. I sent four requests six times each to OpenAI’s gpt-6.1-terra and checked the results separately with RDKit.
Source code: https://github.com/aurio9923/chemomni (MIT license)
Showing the canonical tautomer
Drawing the same compound as a different tautomer changes its SMILES, its fingerprint and its calculated values. One way to make the drawings uniform before comparing them is the canonical tautomer. RDKit’s TautomerEnumerator enumerates the tautomers and picks one by a fixed scoring scheme. However a compound is drawn, the same compound gets the same canonical tautomer.
ChemOmni’s tool tautomers finds this canonical tautomer with RDKit. The AI only chooses the molecule and how to show the result; it does not write the SMILES of any tautomer. When you ask something like “Show me the canonical tautomer,” the canonical tautomer opens in a new tab only if it differs from the drawn form, with the atoms that changed from the drawing marked in red. If the drawn form is already the canonical tautomer, no tab opens and the AI says so.
Example 1: Piroxicam
Piroxicam is often drawn as the enol, as its name (4-hydroxy-…) suggests. I drew it as the enol and asked:
Show me the canonical tautomer of this compound
The AI called tautomers with the setting that opens only the canonical tautomer (show: "canonical"), and the form with a ketone at position 4 opened in a new tab (Figure 1). The result was the same in all six runs.
Two drawings of the same compound give different calculated values (Table 1). cLogP is 1.58 for the drawn enol and 0.91 for the canonical tautomer. The Tanimoto similarity of their Morgan fingerprints is 0.56, which is the “sim 0.56” on the tile in Figure 1. When you compare cLogP across a list of compounds or search by similarity, you see differences in drawing rather than in structure unless you make the drawings uniform before calculating.
| Form | cLogP | TPSA (Ų) |
|---|---|---|
| Drawn enol | 1.58 | 99.6 |
| Canonical tautomer (ketone at position 4) | 0.91 | 96.4 |
Table 1. Two drawings of piroxicam and their calculated values (recalculated with RDKit’s Crippen cLogP and TPSA).
Note that the canonical tautomer is not a prediction of the form that predominates in water or in a protein. It is a representative form that RDKit picks by fixed rules, used to compare structures and to merge duplicates. The tool description says so, and in all six answers the AI added that the canonical form is not a prediction of the predominant tautomer.
Example 2: Imatinib
Show me the canonical tautomer of this compound
For imatinib, the usual drawing is already the canonical tautomer. The AI called the tool with the same setting and, without opening a tab, answered that the drawn structure is already the canonical tautomer (Figure 2). It was the same in all six runs.
To see the tautomers as a list, ask something like “Show me the tautomers.” RDKit’s enumeration also includes forms no chemist would draw, so ChemOmni leaves the following off the list and returns the number left off and the reasons to the AI: forms in which an aromatic ring is no longer aromatic (a phenol as a cyclohexadienone, for example), amides and lactams written as C(OH)=N, enols of amides, acids and esters, and amino groups on an aromatic ring written as an exocyclic C=N (the imine forms of aminopyridines and aminopyrimidines).
Numbering compounds
When one tab holds several compounds, numbers tell the AI which one you mean. Pressing the [#] button above the tab (Number compounds) writes 1, 2, 3, … under the structures (Figure 3). The numbers run from the top row down and from left to right. They stay in the drawing as editor text, so you can also change them by hand.
When you send a chat message, ChemOmni works out which number belongs to which structure and passes them to the AI as [1] SMILES. The AI passes numbers (ref) to the tools and does not write the SMILES of the structures. ChemOmni decides which number belongs to which structure from the positions of the texts and the structures.
As an example, I drew three structures with the same molecular formula (C5H4N4O) on one tab: 1 is allopurinol drawn as the 4-hydroxy form, 2 is allopurinol drawn as the oxo form, and 3 is hypoxanthine.
Of compounds 1–3, which are the same compound drawn only as different tautomers?
The AI called tautomers three times, once per number, and compared only the canonical tautomers it received, without opening tabs. In all six runs the answer was “1 and 2 are the same compound and 3 is a different one,” which is correct. 1 and 2 have the same canonical tautomer. 3 has the same molecular formula but its five-membered ring has the nitrogens in different places (an imidazole, not a pyrazole), so it is not a tautomer. In all six answers the AI explained that 3 has a different connectivity; twice it called 3 a different constitutional isomer.
Importing and exporting SDF
I took the 2D SDFs of five BCR-ABL inhibitors (imatinib, nilotinib, dasatinib, bosutinib, ponatinib) from PubChem and put them in one file. Each record of a PubChem 2D SDF has its own coordinates and holds every hydrogen as an atom. Loaded into the Ketcher editor as it is, every record is placed at its own coordinates, and the five structures and their hydrogens are drawn on top of each other (Figure 4, top).
Opening the same file with the [Import file] button above ChemOmni’s tabs puts the five structures on a new tab in file order, numbered 1–5 (Figure 4, bottom). Hydrogens that can be left implicit are not drawn. Stereochemistry shown by wedges, including wedges on hydrogens, is read before the hydrogens are removed.
On the imported tab, I asked a question using the numbers.
Compare the molecular weights of compounds 1–5 and tell me which exceed 500
The AI passed numbers 1–5 to compute_properties and received a table of properties calculated by RDKit (Figure 5). In all six runs the answer was “compounds 2, 4 and 5,” which agreed with the molecular weights recalculated from the file (nilotinib 529.53, bosutinib 530.46, ponatinib 532.57).
To export, choose SDF, CSV or SMILES from the [Export] menu above the tab. In the SDF each structure is one record with fields for its number (Number), SMILES and properties calculated by RDKit (formula, molecular weight, cLogP, TPSA, HBD, HBA, rotatable bonds, QED, SA score, number of Lipinski violations, InChIKey). Fields for structural alerts (PAINS, Brenk) are added only to the structures that hit one (one Brenk alert each for bosutinib and ponatinib among the five). The CSV holds the same content with one structure per row; the alert columns are empty for structures without a hit. The five exported records matched the imported file in order and structure, and molecular weight, cLogP and TPSA matched the recalculated values within rounding.
On the other hand, the fields of the original SDF (PubChem CID, IUPAC name and so on) and the record names are not kept in the export. To match the export against the original data, use the InChIKey in the export.
Design decisions
- The AI does not write tautomer structures. RDKit enumerates and picks the canonical tautomer; the AI only chooses the molecule and how to show the result.
- The canonical tautomer opens in a tab only when it differs from the drawn form. If it is the same, the AI just says so.
- Tautomers that are generally unnatural are left off the list, with the number left off and the reasons returned. If you ask to see every form, the ones left off are shown too.
Summary
The three features this time give the AI and the chemist a common footing when they talk about compounds. Canonical tautomers make differences in drawing uniform, numbers point to the compounds on the screen, and SDF import and export exchange compounds with files. In every case RDKit decides the structures, and the AI only chooses the molecule and how to show the result. In the 24 runs (four requests, six times each), every answer matched the results recalculated with RDKit.
Notes
- Model: OpenAI’s gpt-6.1-terra (Responses API, reasoning effort low, the app’s default). Figures 1–3 and 5 also show this model’s results.
- Repetition: the four requests were sent six times each in a new chat (24 in total). Example 2 was sent after resetting the drawing of imatinib each time; Examples 1, 3 and 4 were each sent on the same tab.
- Language: the requests were sent in English, as quoted. The six-run results in this article come from the English runs (24 in total); the figures come from one further run of each request, made to take the screenshots, with the same answers. The Japanese version of this article reports the same requests sent in Japanese, with the same results.
- Structures: piroxicam (enol form)
CN1C(C(=O)Nc2ccccn2)=C(O)c2ccccc2S1(=O)=O, imatinibCc1ccc(NC(=O)c2ccc(CN3CCN(C)CC3)cc2)cc1Nc1nccc(-c2cccnc2)n1, allopurinolO=c1[nH]cnc2[nH]ncc12andOc1ncnc2[nH]ncc12, hypoxanthineO=c1[nH]cnc2nc[nH]c12. - SDF: the 2D SDFs from PubChem PUG REST (CID 5291, 644241, 3062316, 5328940, 24826799), joined into one file in this order.
- Checks: RDKit 2026.09.2. The canonical tautomer was recomputed with TautomerEnumerator (set to keep sp3 and double-bond stereochemistry) and also compared, without stereochemistry, with the result of the default settings. Molecular weight, cLogP and TPSA were recalculated with Descriptors.MolWt, Crippen and CalcTPSA. Similarity is the Tanimoto coefficient of Morgan fingerprints (radius 2, 2048 bits).