What is the correct regular expression for InChI?

Question

When I look for a regular expression for InChI, I succeed on EBI's MIRIAM Registery website:

^InChI\=1S?\/[A-Za-z0-9]+(\+[0-9]+)?(\/[cnpqbtmsih][A-Za-z0-9\-\+\(\)\,\/]+)*$

And indeed it works for some examples (using Python's re module):

import re

pattern = '^InChI\=1S?\/[A-Za-z0-9]+(\+[0-9]+)?(\/[cnpqbtmsih][A-Za-z0-9\-\+\(\)\,\/]+)*$'

s = 'InChI=1S/C2H6O/c1-2-3/h3H,2H2,1H3'

re.findall(pattern, s)  

will print

[('', '/c1-2-3/h3H,2H2,1H3')]

However, if I take other examples, as e.g. this one

s2 = 'InChI=1S/C6H12O6/c7-1-2-3(8)4(9)5(10)6(11)12-2/h2-11H,1H2/t2-,3-,4+,5+,6?/m1/s1'

it fails

re.findall(pattern, s2)

will return

[]

What is the correct and general regular expression for InChI?

EDIT

If I see this correctly, in the pattern from above, question marks are not taken into account which are used to mark both “unknown” and “undefined” chiral centres. If one adjusts the regex accordingly, it at least works for the example above; whether it then holds for all InChI is unclear, of course.

pattern = '^InChI\=1S?\/[A-Za-z0-9]+(\+[0-9]+)?(\/[cnpqbtmsih][A-Za-z0-9\-\+\(\)\,\/\?]+)*$'

s2 = 'InChI=1S/C6H12O6/c7-1-2-3(8)4(9)5(10)6(11)12-2/h2-11H,1H2/t2-,3-,4+,5+,6?/m1/s1'

re.findall(pattern, s2)

will return

[('', '/c7-1-2-3(8)4(9)5(10)6(11)12-2/h2-11H,1H2/t2-,3-,4+,5+,6?/m1/s1')]

Answer 1

As long as you are not looking for something very specific, and you are not performing a sanity check of what you are matching against, the provided regular expression will match more often wrong than falsely not matching. It will even match more nonsense InChI than total number of existing ones. It is therefore not much better than the trivial catch them all:

^InChI\=1S?\/[^\s]+(\s|$)

The space (and some few others) is a character that will not be part of a correct InChI. The above will therefore match all correct InChI strings (and many more), and certainly more correct ones than the one given in the question (or in your answer).

The principal problem with the whole approach is basically given in the technical FAQ in section 5. Depending on what we are looking at, not all the layers will be present. There is not even a guarantee that a sum formula is present, as represented by the proton: InChI=1S/p+1. This one (and its relatives) is basically responsible why regex is doomed from the start, and that there is no correct regex in the first place.

If we wanted to implement only a tiny fraction of sanity control to reduce the number of mismatches, we will soon see that the whole string becomes about a page long. Let's examine the given regex a little more to actually understand what it and why it's attempt is futile.

^InChI\=1S?\/[A-Za-z0-9\.]+(\+[0-9]+)?(\/[cnpqbtmsih][A-Za-z0-9\-\+\(\)\,\/\?\;\.]+)*$

Beginning with

^  

matches the start of the string, which is not really necessary

 InChI\=1S?

matches the prefix, while leaving it open if standard or not

           \/[A-Za-z0-9\.]+(\+[0-9]+)?

matches a sum formula (even the silly ones) or anything that doesn't have a main layer and goes straight to isotopes or charges, etc.. Better would here be to make the sum formula optional, i.e. (\/[A-Za-z0-9\.]+)?

                                      (\/[cnpqbtmsih][A-Za-z0-9\-\+\(\)\,\/\?\;\.]+)*$

this is pretty much a fancy way of saying match any kind of layer, as often as you want,( or not,) in what order is just convenient, with whatever content. However, it does not match *, which it should: InChI=1S/2C2H5.Zn/c2*1-2;/h2*1H2,2H3; as given in the FAQ.

A much better approach would be to match the layers separately, to have a little control, because they have a predefined order. For example start with the connectivity 1.1 layer (\/c[0-9\-\*\;\(\)]+)?, move to the connectivity 1.2 or hydrogen layer (\/h[H0-9,\(\)\;]+)?. I am not too sure whether I caught all special characters that might be inserted. Then match on with charge (\/q(\+|\-)+[1-9]+[0-9]*)? and the proton balance \/p, the various stereochemistry layers, fixed layers, etc. There are plenty. I probably got the charge already halfway wrong, as I am not sure whether it needs ; or not.

Coming up with a correct and not too greedy regex for each layer is already a nice task, and if you can do something useful with it, good. For everything else, there is a reason that a software exists to create and parse those strings. For every other purpose the greedy match-it-all from the start should work perfectly fine. Now, InChIKeys are something very different ...

Answer 2

This might not be the final version but covers all cases I have encountered so far (if you find an InChI which cannot be matched with this expression, please post it in the comments along with a source, so that one can adapt the regex):

pattern = '^InChI\=1S?\/[A-Za-z0-9\.]+(\+[0-9]+)?(\/[cnpqbtmsih][A-Za-z0-9\-\+\(\)\,\/\?\;\.]+)*$'

The following changes were made compared to the one posted in the question:

1. One also accounts for . in the chemical formula; check inchi-trust for an example
2. One takes ? into account; check inchi-trust for a detailed explanation
3. One also checks for ; ; check inchi-trust for an example
4. One also has to take . into account later on, not only in the chemical formula. That is explained in "The IUPIC International Chemical identifier":

   /m parity inverted to obtain relative stereo ( 1 = inverted, 0 = not inverted, . = inversion does not affect the parity)

I guess this post will be updated in the future but for my examples it works now and all changes can be backed up by trustworthy sources.

The regular expression is updated now to the one I proposed. Apparently, they are happy to change their expression, so if you find cases where a valid InChI cannot be mapped, please report it to them or add it here in the comments. Clearly, this regular expression might also match strings which are actually not chemicals but at least one should be able to identify each valid InChI as InChI.