MolScreen — drug-likeness¶
MolScreen answers whether a molecule looks like a drug: it parses SMILES, computes physico-chemical properties, runs six medicinal-chemistry filters, and compares against known drugs for novelty. It is pure RDKit — no network, no model weights.
Properties¶
>>> from neorx.molscreen import calculate_properties
>>> props = calculate_properties("CC(=O)Oc1ccccc1C(O)=O")
>>> props.smiles
'CC(=O)Oc1ccccc1C(=O)O'
>>> props.molecular_weight, props.logp
(180.04, 1.31)
>>> props.hbd, props.hba, props.tpsa
(1, 3, 63.6)
>>> props.rotatable_bonds, props.num_rings, props.num_aromatic_rings
(2, 1, 1)
>>> props.fsp3, props.molar_refractivity
(0.111, 44.71)
Note that props.smiles is the canonical form, not what was passed in:
the input C(O)=O came back as C(=O)O. Each field's docstring in the
model explains why the property matters for absorption or permeability —
see the API reference.
Parsing¶
>>> from neorx.molscreen import validate_smiles, canonicalise
>>> validate_smiles("CCO")
True
>>> validate_smiles("xyz!")
False
>>> canonicalise("C(C)O")
'CCO'
parse_smiles returns an RDKit Mol or None. name_to_smiles
resolves a common drug name, and smart_parse accepts either a name or a
SMILES string.
The six filters¶
Each filter returns a result carrying both a verdict and the specific
violations, so a failure is actionable rather than a bare False.
>>> from neorx.molscreen import parse_smiles, run_all_filters
>>> aspirin = parse_smiles("CC(=O)Oc1ccccc1C(O)=O")
>>> for result in run_all_filters(aspirin):
... print(f"{result.name:9} {result.passed!s:6} {result.violations}")
Lipinski True []
Veber True []
Ghose False ['Atom count 13 outside [20, 70]']
Egan True []
PAINS True []
Brenk False ['phenol_ester']
| Filter | Asks |
|---|---|
| Lipinski | Rule of Five: MW, LogP, HBD, HBA — oral bioavailability |
| Veber | Rotatable bonds and TPSA — oral bioavailability from a different angle |
| Ghose | MW, LogP, atom count, molar refractivity within drug-like ranges |
| Egan | LogP and TPSA — absorption |
| PAINS | Pan-assay interference substructures: compounds that hit everything |
| Brenk | Unwanted functional groups: reactive, toxic, metabolically unstable |
Aspirin failing two of six is the right answer, not a bug. It is too small for Ghose's 20-atom floor, and its phenol ester is on Brenk's list. Aspirin is a real drug that would not survive a modern lead-like filter set — which is a useful calibration for how to read these verdicts.
Classification¶
>>> from neorx.molscreen import classify_drug_likeness, DrugLikelihoodCategory
>>> [c.value for c in DrugLikelihoodCategory]
['Drug-like', 'Lead-like', 'Not drug-like']
>>> classify_drug_likeness(run_all_filters(aspirin))
'Lead-like'
classify_drug_likeness takes the filter results, not the molecule.
Drug-likeness and synthetic accessibility¶
>>> from neorx.molscreen import qed_score, sa_score
>>> round(qed_score(aspirin), 3)
0.55
>>> round(sa_score(aspirin), 3)
1.58
QED is RDKit's quantitative estimate of drug-likeness, 0 to 1, combining MW, LogP, HBD, HBA, PSA, rotatable bonds, structural alerts and aromaticity into one desirability score.
SA score is synthetic accessibility, 1 (trivial) to 10 (very hard). Aspirin at 1.58 is about as easy as organic synthesis gets. The composite scorer inverts and normalises this, since lower is better.
Similarity and novelty¶
>>> from neorx.molscreen import tanimoto_similarity
>>> round(tanimoto_similarity("CCO", "CCCO"), 4)
0.5556
find_similar_drugs searches a cache of known drugs and is what supplies
the novelty dimension of the composite score — a candidate highly similar
to an approved drug is not novel. build_cache and clear_cache manage
that cache.
From the command line¶
Reports¶
ScreeningReport and ComparisonReport bundle a molecule's properties,
filter results and similar drugs for reporting. See the
API reference.