3
$\begingroup$

I am not a chemist, in fact, I come from a computer science background. However, I am involved in a project related to artificial intelligence-based drug discovery. For this, I am trying to make a dataset that contains smiles notation of some chemical compounds along with their chemical properties. In the PubChem database, there are lots of available properties, which are:

MolecularFormula, MolecularWeight, CanonicalSMILES, IsomericSMILES, InChI, InChIKey, IUPACName, XLogP, ExactMass, MonoisotopicMass, TPSA, Complexity, Charge, HBondDonorCount, HBondAcceptorCount, RotatableBondCount, HeavyAtomCount, IsotopeAtomCount, AtomStereoCount, DefinedAtomStereoCount, UndefinedAtomStereoCount, BondStereoCount, DefinedBondStereoCount, UndefinedBondStereoCount, CovalentUnitCount, Volume3D, XStericQuadrupole3D, YStericQuadrupole3D, ZStericQuadrupole3D, FeatureCount3D, FeatureAcceptorCount3D, FeatureDonorCount3D, FeatureAnionCount3D, FeatureCationCount3D, FeatureRingCount3D, FeatureHydrophobeCount3D, ConformerModelRMSD3D, EffectiveRotorCount3D, ConformerCount3D.

Surely, I don't want to select all of them as it will result in an overly high-dimensional dataset. So my question is:

What are the physicochemical properties related to "medical" drugs I should keep?

$\endgroup$

1 Answer 1

7
$\begingroup$

There have been many attempts at defining the properties that make a molecule ‘drug-like’, often based on retrospective analysis of approved/marketed drugs.

The most famous of these is the Lipinski rule of 5 which define a set of parameters based on hydrogen bond donor count, hydrogen bond acceptor count, molecular mass, and LogP. Although many medicinal chemists use this to guide decisions, it’s worth noting that less than half of new drugs actually obey these guidelines.

The properties you use also depend on what kind of drug-like molecule you’re tying to find - if your target is in the brain, you want a very different set of properties to a drug that acts in the lungs.

LogD, Molecular weight, aromatic ring count, rotatable bond count, tPSA, hydrogen bond donors/acceptors are a fairly standard set of properties considered routinely (and all able to be easily computed).

The reality is that despite huge amounts of data, we’re really very bad at predicting whether any given molecule will end up being ‘drug-like’, let alone whether it has the potential to be taken forward to an actual clinical candidate.

(As an aside the issue is confounded by our inability to predict in silico some properties: xLogP is a calculation, but rarely correlates to an actual measured LogP value so if you use xLogP to make a prediction about drug-likeness, you’re starting off with a fairly large error)

$\endgroup$
3
  • 1
    $\begingroup$ Thank you @NotEvans for the amazing answer. I would like to ask about LogD and aromatic ring count, how can I compute them as they are not available in the provided properties? $\endgroup$
    – mac179
    Oct 31, 2021 at 7:39
  • 1
    $\begingroup$ You’re very welcome. LogD, like LogP is an experimental value - but there are various calculators (the good ones are commercial). Aromatic ring count could be worked out from the structure if you wanted, but there are libraries available to do so - RDKIT is highly useful for calculating these kinds of parameters. There is also an open source GUI called DaraWarrior that can calculate simple properties based on structure (inchi/smiles) $\endgroup$
    – NotEvans.
    Oct 31, 2021 at 8:15
  • 1
    $\begingroup$ @AmineChadi Have look in the chat room here. $\endgroup$
    – Buttonwood
    Oct 31, 2021 at 16:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.