My compound overlap is almost empty

What actually happened

An InChIKey is a hash of a molecular structure. Change the structure and the key changes, and “standardisation” is a set of operations that change the structure on purpose:

• Salt stripping removes counterions and solvents, so aspirin sodium becomes aspirin.
• Charge neutralisation removes formal charges where a neutral form exists, so a deprotonated carboxylate becomes a carboxylic acid.
• Tautomer normalisation picks one tautomer per tautomeric form, so warfarin’s keto and enol forms become one compound.

Every one of those is correct, standard, recommended practice. None of them is recorded in the file that comes out. So a column of InChIKeys is the answer to a question — which substances are these, and in what state? — of which the file states only the first half.

How big the difference actually is

Measured, across thirty ordinary drugs: twelve of them get a different InChIKey under a different standardisation state. Forty per cent, and the thirty were marketed drugs, salts and tautomer pairs, not curiosities.

Aspirin, arriving as its sodium salt the way a vendor ships it, has three keys:

Read the last two again. They differ in the final character — the charge block — and the third is the key plain aspirin gets. So a library that had its salts stripped and a library that had its salts stripped and was neutralised do not match on aspirin, and nothing about either column of keys suggests why.

Warfarin is the other shape of the same problem. Its keto and enol forms are two keys raw and one key after tautomer normalisation, so a normalised set and an unnormalised set overlap in nothing for warfarin.

Why the answer is plausible rather than absurd

Because the unaffected compounds still match. Eighteen of those thirty are the same key either way, so an overlap of two libraries that differ in standardisation is not empty: it is roughly the fraction of compounds that happen to have no salt, no charge and no tautomer. That is a number in the tens of per cent, which is exactly the range people attribute to assay noise, library composition, a different cut-off, or a real biological finding.

A zero would have been a gift. A plausible number gets written up.

Why no glance at the data catches it

This is the part that makes it worse than the gene-set version of the same failure. Two gene lists in different identifier spaces look different: TP53 is visibly not ENSG00000141510, and anyone eyeballing two columns sees it in a second.

Two compound sets standardised differently look identical. Both are columns of twenty-seven-character strings in the same format, with the same shape, from the same software. BSYNRYMUTXBXSQ-UHFFFAOYSA-M and BSYNRYMUTXBXSQ-UHFFFAOYSA-N are one character apart, in the block nobody reads, and no amount of looking at the file will tell you which standardiser produced it.

Where it bites hardest

A virtual screen scored against its own library. The hit list was standardised by the docking pipeline; the library was downloaded. The enrichment you compute is partly a measurement of your standardiser.

Merging two vendor catalogues. The same substance arrives twice under two keys, so a “unique compounds” count is wrong, and anything that averages over compounds double-counts the ones that were duplicated.

A benchmark against a public set. ChEMBL standardises its compounds and says so; a vendor file does not. Comparing your hits to ChEMBL measures the difference between your preparation and theirs, and comes out plausible.

Stereochemistry, which is the one with a clinical edge and has a page of its own, because its symptom is the opposite of this one’s: RDKit’s tautomer canonicaliser discards stereochemistry, so two enantiomers become one key and an overlap comes back too large rather than too small. See my inactive enantiomer came back as a hit.

What Ligase says instead

A compound table carries a line saying what standardised it, and the engine reads it:

Both states are printed in full, the operations that make them differ are named on their own line, and the closing question is the one to act on — were the two sets standardised differently? — rather than a restatement that they are not equal.

Where the two differ by more than one operation, that line is the whole of what is worth reading:

And two tables that each say nothing are refused rather than allowed to agree by silence, because agreeing by silence is this failure in its purest form:

The version of this that two honest declarations produce

The worst shape of it is not an undeclared file. It is two files that are both telling the truth.

salt-stripped does not pin a charge state and an InChIKey’s final block does. So a reference set that was salt-stripped and left anionic, and one that was salt-stripped and then neutralised, can both carry the honest declaration inchikey; salt-stripped — the values compare equal, the join goes ahead, and the same generated set scores 0% novel against one of them and 50% against the other. A consumer measured exactly that, with ligase check reporting no problems found and exiting 0 for both.

The vocabulary was coarser than the identifier it keys to, which is the same defect as an inferred value that compares equal to another inferred value, one level down. Two things changed:

• Charge state is now checked against the keys. A table declared charge-neutralised whose keys end in anything but N is refused at load, because that is a claim the file refutes.
• salt-stripped alone is documented as the trap it is. Public collections ship neutral parents, so the one-term form produces keys that match nothing in the wild. Comparing compound sets says so where someone is choosing a state.

What is still not checked is the part no key records: salt stripping and tautomer normalisation leave no trace, so those remain recorded on your word (D65).

How it is caught

import ligase as lg
 
library = lg.molecule_set("vendor/catalogue.tsv")
hits = lg.molecule_set("screen/hits.tsv")
 
# Nothing has run. This read both files' representation lines.
lg.check_step("overlap", {"library": library, "hits": hits})

The fix is yours to choose, and the engine does not choose it for you: standardise both to the same state, or stop comparing them and say in the write-up that they are not comparable. What it will not do is let the comparison through and let you read the result as chemistry.

Where you do standardise, lg.chem.standardize writes the state it arrived in onto its own output, so the next step is checked against a true statement rather than a claim nobody verified. See comparing compound sets.