ChEMBL small-molecule dataset

by Spencer Trinh

Dataset

ChEMBL Data is a manually curated database of small molecules used in drug discovery, including information about existing patented drugs

The Google BigQuery query is shown below to obtain the three datasets. The three csv files were combined using pandas. There are three csv files because the free trial version of Goolge BigQuery has a limit to the memory usage for querying public datasets, therefore three separate queries were run and combined in the end.

• First query:

SELECT *
  FROM (
  SELECT
  molregno,
  cmpd_rec.doc_id,
  comp.company,
  comp.country,
  prod.dosage_form,
  prod.route,
  prod.trade_name,
  prod.approval_date,
  prod.ad_type,
  prod.oral,
  prod.topical,
  prod.parenteral,
  ROW_NUMBER() OVER(PARTITION BY molregno ORDER BY molregno DESC) rn
  FROM `patents-public-data.ebi_chembl.compound_records_23` AS cmpd_rec
  JOIN `patents-public-data.ebi_chembl.molecule_synonyms_23` AS ms USING (molregno)
  JOIN `patents-public-data.ebi_chembl.research_companies_23` AS comp USING (res_stem_id)
  JOIN `patents-public-data.ebi_chembl.formulations` AS form USING (molregno)
  JOIN `patents-public-data.ebi_chembl.products_23` AS prod USING (product_id)

  ) as subq
 WHERE rn =1
 LIMIT 1000;

• Second query:

SELECT *
  FROM (
  SELECT
  molregno,
  cmpd_rec.doc_id,
  cmpd_struc.canonical_smiles,
  cmpd_rec.compound_name,
  cmpd_prop.mw_freebase,
  cmpd_prop.alogp,
  cmpd_prop.hba,
  cmpd_prop.hbd,
  cmpd_prop.psa,
  cmpd_prop.rtb,
  cmpd_prop.ro3_pass,
  cmpd_prop.num_ro5_violations,
  cmpd_prop.acd_most_bpka,
  cmpd_prop.acd_logp,
  cmpd_prop.acd_logd,
  cmpd_prop.molecular_species,
  cmpd_prop.full_mwt,
  cmpd_prop.aromatic_rings,
  cmpd_prop.heavy_atoms,
  cmpd_prop.num_alerts,
  cmpd_prop.qed_weighted,
  cmpd_prop.mw_monoisotopic,
  cmpd_prop.full_molformula,
  cmpd_prop.hba_lipinski,
  cmpd_prop.hbd_lipinski,
  cmpd_prop.num_lipinski_ro5_violations,
  md.max_phase,
  md.therapeutic_flag,
  md.structure_type,
  md.molecule_type,
  md.first_approval,
  md.natural_product,
  md.availability_type,
  md.withdrawn_flag,
  ROW_NUMBER() OVER(PARTITION BY molregno ORDER BY molregno DESC) rn
  FROM `patents-public-data.ebi_chembl.compound_records_23` AS cmpd_rec
  JOIN `patents-public-data.ebi_chembl.compound_properties_23` AS cmpd_prop USING (molregno)
  JOIN `patents-public-data.ebi_chembl.compound_structures_23` AS cmpd_struc USING (molregno)
  JOIN `patents-public-data.ebi_chembl.molecule_dictionary_23` AS md USING (molregno)
  ) as subq
  WHERE rn = 1 AND
  molregno in ({LIST OF MOLREGNO});

• Third query:

SELECT *
  FROM (
  SELECT
  cmpd_rec.molregno,
  cmpd_rec.doc_id,
  md.black_box_warning,
  md.inorganic_flag,
  md.indication_class,
  ROW_NUMBER() OVER(PARTITION BY cmpd_rec.molregno ORDER BY cmpd_rec.molregno DESC) rn
  FROM `patents-public-data.ebi_chembl.compound_records_23` AS cmpd_rec
  JOIN `patents-public-data.ebi_chembl.molecule_dictionary_23` AS md USING (molregno)
  ) as subq
  WHERE rn = 1 AND
  molregno in ({LIST OF MOLREGNO});

where the LIST OF MOLREGNO is a list that was obtained from the first query

Summary of Findings

The mean numerical features (molecular descriptors) of this small dataset are as follows:

Feature	Mean Value
mw_freebase	428.973270
alogp	2.370098
psa	93.431967
acd_most_bpka	6.661588
acd_logp	2.357922
acd_logd	1.053922
full_mwt	463.672533
qed_weighted	0.550195
hba	4.993333
hbd	1.873333
rtb	5.558667
num_ro5_violations	0.360000
aromatic_rings	1.484000
heavy_atoms	25.386667
num_alerts	0.928000
hba_lipinski	5.901333
hbd_lipinski	2.117333
num_lipinski_ro5_violations	0.254667

Furthermore, the most common chemical subgroup is benzene; the US has developed the most drug compounds and more specifically Pfizer has the most drugs on the market.

Key Insights for Presentation

In this dataset, neutral compounds are the most common at 349 in count. The least are zwitterions which are molecules or ions having separate positively and negatively charged groups. The nitrile group is acidic and seems to have the highest qed scores. Since benzene chemical subgroups are so ubiquitous they have very high density in the inner region of the violin plots. Molecules or compounds that have peptide bonds and are considered a zwitterion have very low qed scores, likewise with terpenes. Nitrile groups seem to have the highes qed scores. This seems to be correct since nitriles are very polar and act as a good hydrogen bond acceptor.

Benzene is very ubiquitous in all organic compounds, thus it was observed in most of the drug compounds in this dataset. Next in line are the ester groups and then double bond functional groups. Because this dataset is quite small, perhaps we aren't getting the full picture of how these chemical subgroups really count up. The selection of these chemical groups was arbitrary. There may be other chemical groups that could have been better to choose.