Skip to content

GitLab

Last edited by Miquel Duran-Frigola
Page history
This is an old version of this page. You can view the most recent version or browse the history.

datasets

Datasets

Datasets are the cornerstone of the CC.

Levels, coordinates and datasets

The CC is divided in five levels of increasing complexity:

Level Name Description
A Chemistry Chemical properties of the compounds.
B Targets Chemical-protein interactions.
C Networks Higher-order effects of small molecules.
D Cells Readouts of compound cell-based assays.
C Clinics Clinical data of drugs and environmental chemicals.

In turn, each level is divided in 5 sublevels or coordinates representing different aspects of the data. Each sublevel has an exemplary dataset, as described below:

Coordinate Name Description
A1 2D fingerprints Binary representation of the 2D structure of a molecule. The neighborhood of every atom is encoded using circular topology hashing.
A2 3D fingerprints Similar to A1, the 3D structures of the three best conformers after energy minimization are hashed into a binary representation without the need for structural alignment.
A3 Scaffolds Largest molecular scaffold (usually a ring system) remaining after applying Murcko’s pruning rules. Additionally, we keep the corresponding framework, i.e. a version of the scaffold where all atoms are carbons and all bonds are single. The scaffold and the framework are encoded with path-based 1024-bit fingerprints, suitable for capturing substructures in similarity searches.
A4 Structural keys 166 functional groups and substructures widely accepted by medicinal chemists (MACCS keys).
A5 Physicochemistry Physicochemical properties such as molecular weight, logP and refractivity. Number of hydrogen-bond donors and acceptors, rings, etc. Drug-likeness measurements e.g. number of structural alerts, Lipinski’s rule-of-5 violations or chemical beauty (QED).
B1 Mechanism of action Drug targets with known pharmacological action and modes (agonist, antagonist, etc.).
B2 Metabolic genes Drug metabolizing enzymes, transporters and carriers.
B3 Crystals Small molecules co-crystalized with protein chains. Data is organized according to the structural families of the protein chains.
B4 Binding Compound--protein binding data available in major public chemogenomics databases. Data mainly comes from academic publications and patents. Only binding affinities below a class-specific threshold are kept (kinases ≤ 30 nM, GPCRs ≤ 100 nM, nuclear receptors ≤ 100 nM, ion channels ≤ 10 uM and others ≤ 1 uM).
B5 HTS bioassays Hits from screening campaigns against protein targets (mainly confirmatory functional assays below 10 uM).
C1 Biological roles Ontology terms associated to small molecules with recognized biological roles, such as known drugs, metabolites and other natural products.
C2 Metabolic network Curated reconstruction of human metabolism, containing metabolites and reactions. Data is represented as a network where nodes are metabolites and edges connect substrates and products of reactions.
C3 Canonical pathways Canonical pathways related to the known receptors of compounds (as recorded in B4). Pathways are assigned via a guilt-by-association approach, i.e. a molecule is related to a pathway if at least one of the molecule targets is a member of it.
C4 Biological processes Similar to C3, biological processes from the gene ontology are associated to compounds via a guilt-by-association approach from B4 data. All parent terms are kept, from the leaves of the ontology to its root.
C5 Interactomes Neighborhoods of B4 targets are collected by inspecting several large protein-protein interaction networks. A random-walk algorithm is used to obtain a robust measure of 'proximity' in the network.
D1 Gene expression Transcriptional response of cell lines upon exposure to small molecules. A well-documented reference dataset of gene expression profiles is used to map all compound profiles using a two-sided gene set enrichment analysis.
D2 Cancer cell lines Small molecule sensitivity data (GI50) of a panel of 60 cancer cell lines.
D3 Chemical genetics Growth inhibition profiles in a panel of ~300 yeast mutants. Data are combined with yeast genetic interaction data, so that compounds can be assimilated to genetic alterations when they have similar profiles.
D4 Morphology Changes in U-2 OS cell morphology measured after compound treatment using a multiplexed-cytological cell painting assay. 812 morphology features are recorded via automated microscopy and image analysis.
D5 Cell bioassays Small molecule cell bioassays reported in ChEMBL, mainly growth and proliferation measurements found in the literature.
E1 Therapeutic areas Anatomical Therapeutic Chemical (ATC) codes of drugs. All ATC levels are considered.
E2 Indications Indications of approved drugs and drugs in clinical trials. A controlled medical vocabulary is used.
E3 Side effects Side effects extracted from drug package inserts via text-mining techniques.
E4 Disease phenotypes Manually curated relationships between chemicals and diseases. Chemicals include drug molecules and environmental substances, among others.
E5 Drug-drug interactions Changes in the effect of a drug when it is taken together with a second drug. Drug-drug interactions may alter pharmacokinetics and/or cause side effects.

Each of the coordinates can contain an arbitrary number of datasets. All datasets are fully described in the PostGreSQL database, and searchable at http://chemicalchecker.org/datasets/. They receive a numbered coding (e.g. A1.001).

Dataset characteristics

This is how we define a dataset:

Column Values Description
Code e.g.A1.001 Identifier of the dataset.
Level e.g. A The CC level.
Coordinate e.g.A1 Coordinates in the CC organization.
Name 2D fingerprints Display, short-name of the dataset.
Technical name 1024-bit Morgan fingerprints A more technical name for the dataset, suitable for chemo-/bio-informaticians.
Description 2D fingerprints are... This field contains a long description of the dataset. It is important that the curator outlines here the importance of the dataset, why did he/she make the decision to include it, and what are the scenarios where this dataset may be useful.
Unknowns True/False Does the dataset contain known/unknown data? Binding data from chemogenomics datasets, for example, are positive-unlabeled, so they do contain unknowns. Conversely, chemical fingreprints or gene expression data do not contain unknowns.
Permanent True/False Are measurements for each entry permanent? 2D fingerprints, for example, are permanent. However, most of the biological data may change/evolve with the different versions of the CC. This field, in essence, dictates whether the dataset should be completely updated in every update of the CC, or whether new entries can be simply appended.
Finished True/False Is the dataset considered to be finished? For examples, datasets coming from supplementary data of scientific papers are immutable, and they consequently need no updates in posterior versions of the CC.
Data type Discrete/Continuous The type of data that ultimately expresses de dataset, after the pre-processing. Categorical variables are not allowed; they must be converted to one-hot encoding or binarized. Mixed variables are not allowed, either.
Predicted True/False Is the dataset a result of a prediction (by us or by others?). Prediction results are perfectly valid CC datasets, in principle.
Connectivity True/False Is there a way to connect this dataset to other biological entities? We understand connectivity as a generalization of the cMap idea of matching gene expression signatures.
Connectivity comments Free text commenting on the connectivity strategy (e.g. type of distance) This field needs to be self-explanatory.
Keys e.g. CPD (we use @afernandez Bioteque nomenclature). May be NULL. In the core CC database, most of the times this field will correspond to CPD, as the CC is centered on small molecules. It only makes sense to have keys of different types when we do connectivity attempts, that is, for example, when mapping disease gene expression signatures.
Number of keys e.g. 800000 Number of samples in the dataset.
Features e.g. GEN (we use Bioteque nomenclature). May be NULL. When features correspond to explicit knowledge, such as proteins, gene ontology processes, or indications, we express with this field the type of biological entities. It is not allowed to mix different feature types. Features can, however, have no type, typically when they come from a heavily-processed dataset, such as gene-expression data. Even if we use Bioteque nomenclature to the define the type of biological data, it is not mandatory that the vocabularies are the ones used by the Bioteque; for example, I can use non-human Uniprot ACs, if I deem it necessary.
Number of features e.g. 1000 Number of features in the dataset.
Exemplary True/False Is the dataset exemplary of the coordinate. Only one exemplary dataset is valid for each coordinate. Exemplary datasets should have good coverage (both in keys space and feature space) and acceptable quality of the data.
Source Free text defining the source of data. More than one source is allowed. We have mild constraints in the nomenclature, here.
Version CC version The CC is updated every 6 months.
Public True/False Some datasets are public, and some are not, especially those that come from collaborations with the pharma industry.

The information above can be stored in a postgresql table named datasets.

It is important that we decide how to store and organize datasets in the file-system, though. For this I need advice and help from all of you, @mbertoni , @oguitart and @afernandez.

I suggest the following structure in e.g. aloy/web_checker/ (or somewhere else):

  • Each dataset is stored correspondingly, e.g. ./datasets/A/A1/001/.
  • A ./data.h5 file:
  • V: the matrix of values (np.int8, np.float32, ...)
  • keys: sorted alphabetically
  • features: sorted alphabetically
  • etc (de
  • A ./processing/ folder where a mini-pipeline is devoted to processing the data until obtaining the final dataset (data.h5). Downloads are not here!, since many downloads are shared between datasets. Downloads and download scripts are wherever @oguitart decides.
  • A ./connectivity/ folder where connectivity scripts are stored. I don't know how to organized this, yet. This folder will obviously be empty if no connectivity is possible for this dataset.
  • A ./models/ folder where persistent models are stored. This folder may be empty many times.

Dataset processing