Molecules Orthonym 1.0.7 does not name: PubChem snapshot (2026-10-14)

[Kohulan]

This post records which PubChem molecules Orthonym could not name, and why, as a fixed reference point. A data file lists every one of them by CID (available on request). Re-run the same CIDs with a later version to measure progress.

The run

• Orthonym 1.0.7, commit 4658fcd4f, best-effort tier (Orthonym(general_fallback=True, general_fallback_unverified=True, allow_aromatic_general=True).name_tiered(smiles)), 45 s per molecule.
• Input: PubChem CID-SMILES (2026-09-27), 124,717,116 rows. Kept 112,607,951 single-component molecules with carbon and only organic elements.
• Score: name -> OPSIN 2.9.0 -> InChIKey, compared with the input's InChIKey.
• Named: 112,088,617 (99.539%). Round trip exact: 112,086,816 (99.537% of all). Wrong structure: 0.
• Not named: 519,334 (0.461%). This post is about these.

Why they failed (code stored by the run)

Code	Rows	% of not named	Meaning
UNNAMEABLE	437,113	84.2%	the namer declined
NATIVE_HANG	38,396	7.4%	the run's watchdog stopped a worker that wrote nothing for 90 s
EXC	31,933	6.1%	RDKit refuses the SMILES: mostly halogens with an expanded valence (point 4 below)
STRUCTURE_TOO_LARGE	7,707	1.5%	declined as too large
TIMEOUT	4,160	0.8%	over the 45 s limit per molecule
ISOLATED_ATOM	23	0.0%	a single atom
abstain	2	0.0%	declined, no code

UNNAMEABLE does not say where the namer declined. A refusal census (a re-run that records each refusal site) is the next step.

Structure classes

Each molecule is put in the FIRST class that matches, in the order of this table, so the rows add up to 519,334. All matching traits are in the flags column of the file. PubChem name reads back = PubChem's own IUPAC name for the CID gives the exact input InChIKey through OPSIN 2.9.0; it shows that a provable name exists for that molecule.

Class (first match)	Rows	%	Main codes	PubChem name reads back	Example CIDs
invalid_input_rdkit	31,933	6.1%	EXC 31,933	0 (0.0%)*	20591001, 23546608, 87065314
radical	114,185	22.0%	UNNAMEABLE 112,810, NATIVE_HANG 874	0 (0.0%)	5462310, 175670874, 18991443
isotope_label	122,951	23.7%	UNNAMEABLE 109,771, NATIVE_HANG 8,584	112,950 (91.9%)	59925419, 59965145, 60127937
boron_silicon_selenium	36,404	7.0%	UNNAMEABLE 32,909, NATIVE_HANG 2,398	31,023 (85.2%)	10015164, 85987466, 139213653
over_100_heavy_atoms	25,469	4.9%	NATIVE_HANG 15,595, STRUCTURE_TOO_LARGE 5,652	19,020 (74.7%)	89829761, 89829762, 101016548
net_charge	74,228	14.3%	UNNAMEABLE 72,713, NATIVE_HANG 1,510	44,120 (59.4%)	14818482, 20111956, 20306693
zwitterion_or_ylide	31,688	6.1%	UNNAMEABLE 31,208, NATIVE_HANG 477	27,346 (86.3%)	144594019, 71328318, 85870860
hypervalent_halogen_or_n	15,811	3.0%	UNNAMEABLE 15,686, NATIVE_HANG 124	15,501 (98.0%)	155731853, 22951201, 57468495
phosphorus	2,751	0.5%	UNNAMEABLE 2,442, NATIVE_HANG 309	1,724 (62.7%)	14017, 15939506, 15939505
macrocycle	15,608	3.0%	UNNAMEABLE 11,900, NATIVE_HANG 3,706	12,482 (80.0%)	5364836, 101402580, 170698536
bridged	11,898	2.3%	UNNAMEABLE 11,487, NATIVE_HANG 409	7,823 (65.8%)	10345, 20152800, 163728282
spiro	13,761	2.6%	UNNAMEABLE 11,915, NATIVE_HANG 1,846	9,856 (71.6%)	164106464, 123738850, 153378349
fused_ring_system_5plus	10,763	2.1%	UNNAMEABLE 9,856, NATIVE_HANG 904	9,616 (89.3%)	123218, 9993608, 10430221
six_plus_ring_systems	1,309	0.3%	UNNAMEABLE 1,218, NATIVE_HANG 91	1,121 (85.6%)	134501929, 20583946, 3005095
peptide_like	1,199	0.2%	UNNAMEABLE 810, NATIVE_HANG 387	408 (34.0%)	6638418, 60119552, 6638392
sugar_like	45	0.0%	UNNAMEABLE 28, NATIVE_HANG 17	29 (64.4%)	4982, 657, 173882171
ten_plus_stereocentres	171	0.0%	NATIVE_HANG 121, UNNAMEABLE 50	166 (97.1%)	6743, 5224, 5066
partial_stereo	303	0.1%	UNNAMEABLE 248, NATIVE_HANG 55	274 (90.4%)	24005, 10887849, 163676569
small_le10_heavy	102	0.0%	NATIVE_HANG 86, UNNAMEABLE 16	95 (93.1%)	3024, 75, 7849
acyclic	887	0.2%	UNNAMEABLE 504, NATIVE_HANG 383	730 (82.3%)	14257, 27621, 28319
no_standout_trait	7,868	1.5%	UNNAMEABLE 7,348, NATIVE_HANG 520	6,898 (87.7%)	14310, 14311, 14863
all	519,334	100%		301,182 (58.0%)

What makes them special

Rates below compare the not-named rows with a sample of named rows (shard 0000, 224,180 named rows, scaled x500 to estimate all named rows). A trait counts every not-named row that has it, not only the first class.

Trait	Not named with trait	Share of not named	Share of named (sample)	Est. share of PubChem molecules with trait that are not named
radical (unpaired electron)	114,185	22.0%	0.23%	31.1%
isotope label	125,895	24.2%	0.32%	26.1%
hypervalent halogen or N	45,899	8.8%	0.02%	73.0%
boron, silicon or selenium	143,434	27.6%	2.08%	5.8%
over 100 heavy atoms	36,769	7.1%	0.41%	7.5%
net charge	99,672	19.2%	1.98%	4.3%
macrocycle (ring > 12 atoms)	34,411	6.6%	0.84%	3.5%
fused ring system with 5+ rings	116,557	22.4%	2.03%	4.9%
spiro	51,807	10.0%	2.92%	1.6%
bridged	58,938	11.3%	3.32%	1.6%
zwitterion or ylide (charges cancel)	45,514	8.8%	4.35%	0.9%

1. Unusual chemistry, not ring complexity, drives the gap. Radicals, isotope labels and hypervalent halogen/N are each rare among named molecules (0.02-0.32%) but common among not-named ones. Bridged and spiro molecules fail only about 1.6% of the time.
2. Boron is mostly radical boron. 104,411 not-named rows contain B, and 66,618 of them (63.8%) also carry a radical: many PubChem records draw boron as [B] with free valences.
3. Radicals: 114,185 rows, and PubChem's own names read back for 0 of them. The scorer read radical rows with OPSIN's radical mode (-r), so PubChem's names fail because they do not express the radical. Orthonym already has a radical name path and a radical identity check; these rows are open work for it.
4. The 31,933 SMILES that RDKit refuses are mostly real compounds, not bad data. RDKit's default valence model allows a neutral Cl or Br only one bond; PubChem keeps the structure as deposited. By what RDKit refuses: halogens with an expanded valence such as chlorosyl, chloryl, perchloryl, bromosyl and λ3/λ5/λ7-bromanes, 15,857 (e.g. CID 13044168, 11851293, 62352); halogen anions with two bonds, 15,686 (e.g. CID 6412562, 6398996); four-bonded boron cations, 276; probable drawing errors such as a halogen anion with one bond, 109 (e.g. CID 58422749); other, 5. *The 0 in the class table is a limit of our check, not of the names: OPSIN reads PubChem's name for 15,158 of these rows, but our RDKit-based check cannot make an InChI from the result; OPSIN cannot read the name for 13,450, and 3,325 have no PubChem name.
5. Size matters at the top end. Of 113,582 PubChem molecules over 150 heavy atoms, 13,852 (12.2%) are not named: 9,861 NATIVE_HANG, 3,933 STRUCTURE_TOO_LARGE, 28 TIMEOUT, 29 EXC, 1 other. None of them is UNNAMEABLE: at this size the limit is time, not a missing rule.
6. Some NATIVE_HANG rows are run artefacts. The watchdog marks the row a stopped worker held. 1,001 hang rows have 40 heavy atoms or fewer (372 have 20 or fewer). Seven of them, tested with the current main branch, got a PIN in under 2 s each (e.g. CID 3024, 14257, 12286). The other 37,395 hang rows have 41+ heavy atoms and are slow for real: in a re-test of 300 random hang rows with the current main branch (6 workers, 25 minutes), only 13 finished.
7. 7,868 rows show no stand-out trait (7,348 UNNAMEABLE). They look like ordinary drug-like molecules; only the refusal census can explain them.

The data file (available on request)

orthonym_1.0.7_pubchem_unnamed_2026-10-14.tsv.gz (13.2 MB, 519,334 rows + header, sorted by CID), tab-separated:

Column	Content
cid	PubChem CID
smiles	the input SMILES used by the run
heavy_atoms	heavy-atom count
failure	code from the table above
class	first-match class from the table above
flags	every matching trait, ;-separated (also element_B, element_Si..., cip_digraph_blowup, ring_exocyclic_ez_double_bond)
pubchem_name_roundtrip	outcome of PubChem's own IUPAC name: rt_exact, no_pubchem_name, opsin_fail, inchi_fail, rt_constitutional, mismatch, inchi_key_artifact

How to re-assess (e.g. in one month)

1. Name every smiles in the file with the new version, same call and 45 s limit as above.
2. Score each name: OPSIN 2.9.0 -> InChIKey, equal to the input's InChIKey = solved.
3. Report per class and per failure: solved, still not named, and any wrong structure (must stay 0). Give counts next to every percentage.
4. Rows with pubchem_name_roundtrip = rt_exact are the provable target (301,182 rows today).

[Kohulan]

Update 2026-11-02: latest numbers

We named the molecules that were not exact again with newer code from the main branch (latest commit e6229a6ed, 45 s per molecule; one earlier re-run gave slow rows up to 240 s). Rows that were already exact were not run again. Wrong structure: 0.

The 519,334 molecules of this post

Now exact: 461,081 (88.78%).

Still not exact: 58,253 (11.22%): no name 56,521, a name OPSIN cannot read 1,255, a name whose stereo part differs 477.

No name by code: UNNAMEABLE 40,614; TIMEOUT 8,716; FULLERENE 2,845; STRUCTURE_TOO_LARGE 2,319; EXC (RDKit refuses the SMILES) 1,604; NATIVE_HANG 407; ISOLATED_ATOM 16.

The 1,801 molecules that 1.0.7 named but did not round-trip exactly: 1,499 are now exact.

The 12,109,165 molecules this post left out

The first run kept only single-component organic molecules. We have now named the other 12,109,165 rows too (more than one component, other elements, no carbon): 11,596,371 exact (95.765%). The largest group still without a name is elements that Orthonym does not name yet: 330,890 rows.

All of PubChem

Part	Rows	Exact	%
single-component organic	112,607,951	112,549,396	99.948%
the rest	12,109,165	11,596,371	95.765%
all	124,717,116	124,145,767	99.542%

Changes in the method since this post:

• The names are read back with the development version of OPSIN (main branch, commit 42e87ac6), not 2.9.0 (one earlier re-run used 2.9.0). Some names read back only there, for example the halogen anions with two bonds (ethyl(4-fluorophenyl)chloranuide).
• The score now also works for the hypervalent halogen inputs of point 4. RDKit's default reader refuses them, so for these rows only, both InChIKeys come from RDKit without its valence check. On 15,000 rows where both readers work, the two keys were the same every time.