A curated diverse molecular database of blood-brain barrier permeability with chemical descriptors

The highly-selective blood-brain barrier (BBB) prevents neurotoxic substances in blood from crossing into the extracellular fluid of the central nervous system (CNS). As such, the BBB has a close relationship with CNS disease development and treatment, so predicting whether a substance crosses the BBB is a key task in lead discovery for CNS drugs. Machine learning (ML) is a promising strategy for predicting the BBB permeability, but existing studies have been limited by small datasets with limited chemical diversity. To mitigate this issue, we present a large benchmark dataset, B3DB, complied from 50 published resources and categorized based on experimental uncertainty. A subset of the molecules in B3DB has numerical log BB values (1058 compounds), while the whole dataset has categorical (BBB+ or BBB−) BBB permeability labels (7807). The dataset is freely available at https://github.com/theochem/B3DB and 10.6084/m9.figshare.15634230.v3 (version 3). We also provide some physicochemical properties of the molecules. By analyzing these properties, we can demonstrate some physiochemical similarities and differences between BBB+ and BBB− compounds.


Background & Summary
The blood-brain barrier (BBB) denotes a regulatory and protective mechanism of microvasculature in the central nervous system (CNS) that is central to regulating the homeostatis of the CNS 1,2 and protecting the CNS from toxins, pathogens, and inflammations 3 . However, it is estimated that 98% of small molecule drugs are not BBB permeable 4 . Therefore, predicting BBB permeability for small molecules is a vital but challenging task in drug discovery and development [4][5][6][7] . However, existing computational models for a molecule's BBB permeability are inadequate. In particular, they are restricted by the limited size and chemical diversity of existing sets of training data 8 . Moreover, although many different machine-learning (ML) models for predicting BBB permeability have been proposed, these models are not directly comparable because they use widely varying training data, ranging from as few as 45 molecules 9,10 to as many as 7236 molecules 11 . The purpose of this paper is to curate an accessible, clean, well-documented, and reasonably comprehensive dataset of BBB permeability data and present it in a way that is convenient for those building new BBB predictive models. While our database, B3DB, is not the first attempt to curate data from the literature to construct a molecular BBB database, B3DB contains more molecules, and categorizes the molecules based on experimental uncertainty. Both features are very helpful when developing and validating ML models for BBB.
There are two types of data for BBB, numerical and categorical data. Numerical data is usually reported as log BB, the logarithm of brain-plasma concentration ratio, brain blood

Methods
The next three sections describe how raw data was collected from various sources, cleaned, and curated. We then describe how the dataset was extended with chemical descriptors (beyond the reference BBB value). This workflow is summarized in Fig. 1 Table 1. For each data source, a standard Excel workbook is formatted for further processing. If the original data is in portable document format (PDF), it is converted to a pandas 18 . DataFrame and then stored in XLSX format with tabula-py 19 . For files in DOCX or DOC extension, as well as CSV, TXT and other Excel compatible formats, they are converted to Excel XLSX format directly, using Microsoft Office. We performed several automated consistency checks (e.g., numerical data should be reported as floating-point numbers) and manually verified a subset of the data to ensure that the data was faithfully transferred to *.xlsx format. In total, 33825 raw data records were collected. The 50 datasets have various formats and include a wide range of information, so we constructed a template that contained only the most essential data, compound name, simplified molecular-input line-entry system (SMILES) string, PubChem compound identifier (CID), log BB, BBB+/BBB− (whether a compound is BBB permeable or not), the IUPAC International Chemical Identifier (InChI), the threshold value used to determine categorical type of a compound, and the literature source for that data value. Data cleaning. In the data cleaning stage, an initial molecule specification (a SMILES string, PubChem CID, and/or compound name) is input; the output is also a SMILES string, but with transcription and typographical www.nature.com/scientificdata www.nature.com/scientificdata/ errors fixed, and with salts/solvents removed. In addition, molecules containing heavy metal atoms are removed from the database. A followed up standardization of molecular reorientation is performed which include updating valences, kekulizing and normalizing molecules, and neutralizing molecular charges. The basic procedure is shown in Fig. 2(a). www.nature.com/scientificdata www.nature.com/scientificdata/ The first step is to fix invalid SMILES strings. For example, white spaces and line breaks in SMILES were removed. Some other issues (e.g., where a dash was used in lieu of a negative sign for the molecular charge) were manually remedied. Our data is drawn from 50 distinct sources, and a full molecule specification is not always provided. For example, some sources list only the compound names (and not the SMILES strings or PubChem CIDs); other sources list only PubChem CIDs. In these cases, PubChemPy 20 was used to access the PubChem 21 database to retrieve information about missing compound names, SMILES strings and PubChem CIDs. When only the compound name was available, there can be multiple PubChem instances. If this were to happen, the first Pubchem instance is selected and a note is added to the database flagging the potential ambiguity. Fortunately this does not seem to occur in this specific database. There are also a few molecules for which only molecular structures, and not SMILES or compound names, are provided. In these cases we built the molecules manually and searched for the Pubchem CID and SMILES string with the PubChem web interface. All the SMILES strings were loaded into RdKit 22 (version 2019.03.4) to build molecule objects. If the object is None, the SMILES is considered to be invalid. This leads to 33771 measured BBB instances.
Stereochemistry can play a significant role in a molecule's BBB permeability because of transporters' specific stereoselectivity 14,15 . However, there is no stereochemical information in SMILES strings. To add stereochemical information to SMILES, and to deal with generic SMILES strings that were technically valid but not in canonical form, the original SMILES were upgraded to isomeric SMILES by using PUG-REST API 23 wherever possible. Otherwise, the canonical SMILES were retrieved from PubChem database with PUG-REST API 23 . The inclusion of stereochemical data about the molecules is an important, and (we believe) unique feature of B3DB.
Once the SMILES representations are fixed, ChEMBL_Structure_Pipeline 24 was used to strip the salts and neutralize the charge. Molecules containing metal atoms or heavy atom with atomic number greater than 20 (e.g., Zinc, Bromine, Krypton, Iodine, and Xenon) were removed. Molecules with more than 7 boron atoms are also excluded due to problems of depicting borane compounds. Implicit valence and ring information were recomputed followed by kekulizing, normalization of molecules and molecular charges were neutralized. These revisions change the molecular structure, so the Pubchem CIDs were updated from the revised SMILES strings. Data curation. The curation procedures for numerical and categorical data are summarized in Fig. 3.
To curate the data, a unique chemical identifier is required. Although InChI is unique in principle, it cannot resolve tautomeric forms, which is a common source of ambiguity and error in chemical structure representation. Therefore, we examined the unique InChI generated with RdKit and the isomeric SMILES (and canonical SMILES where isomeric SMILES is unavailable). The number of unique SMILES is greater than the number of unique InChI values, but the redundancy is merely because each SMILES represents a specific resonance structure.
Curation of numerical data. To curate the 8841 numerical BBB data values, log BB values for each molecule were merged into a list. The 20 instances with log BB <= −9 were regarded as outliers because, based on the distribution of log BB values, they seemed suspicious. Next, we identified molecules where there are multiple reported log BB values and eliminated those molecules from the database if the reported values differed significantly. Specifically, we eliminated 16 molecules where max (log BB) − min(log BB) > 1. The values that remain after curation are merged into 1065 molecular records. The molecular records are augmented, as necessary, to ensure that they are complete, including compound name, IUPAC name, isomeric (canonical) SMILES, etc..
Here is the detailed curation procedure for numeric data.  www.nature.com/scientificdata www.nature.com/scientificdata/ by less than 5% from the mean value. In these cases, the mean value is used as the log BB value for the molecule.
3. Group C (3 molecules). Other molecules with two distinct log BB values. The (weighted) mean value is used as the curated value for group C (just as for group B). 4. Group D (149 molecules). Other molecules with more than two distinct values; whichever value occurs with greatest frequency is used. In three case, two distinct values were reported with maximum frequency; we discarded those molecules from the dataset.
The 7 molecules which failed to be categorized as group A, B, C or D, they are discarded. The final dataset therefore contains 1058 molecules; for most of these molecules (815 molecules) multiple, mutually consistent, values of log BB are reported in the literature.
Curation of categorical data. The 33689 data values were divided into two categories, numerical data and (binary) categorical data. all sources agree on the categorical label. The unambiguous label is used. 3. Group C (3077 molecules). Molecules where all sources agree on the categorical label, but the sources that do not report their threshold value. 4. Group D (51 molecules). Molecules with two different BBB permeability labels. The most prevalent label is used. In the 45 cases where the two labels occurred with equal frequency, the molecule was discarded.
The 7807 remaining molecular records are augmented to ensure that they are complete, including compound name, IUPAC name, isomeric (canonical) SMILES, etc..

Data extension with chemical descriptors.
To better facilitate building BBB predictive models, the curated datasets were extended with chemical descriptors. Then 1613 chemical descriptors were calculated with mordred version 1.1.1 16 . The purpose of providing this extended data is to facilitate easy use of the B3DB, without requiring precomputation of cheminformatics descriptors.

Data records
There are two datasets provided in this study, one with numeric log BB values (1058 molecules) and the other with categorical labels (7807 molecules with 4956 BBB+ and 2851 BBB−). B3DB data is stored in the comma-separated values (CSV) format and contains SMILES representations, compound name, IUPAC name, log BB value, threshold, BBB+/BBB− and the corresponding references along with 1613 molecular descriptors. This is summarized in Table 2. The data are openly accessible at GitHub (https://github.com/theochem/B3DB) as well as figshare platform 33 . available from PubChem, also isomeric SMILES. We then attempt to load each SMILES string into OEChem Toolkit 17 as an OEGraphMol object; if this is successful then this SMILES is regarded as valid. (See Fig. 2(b)).
analysis of curated datasets. The BBB data comes from 50 sources, and was acquired in different laboratories, under different conditions, and using different protocols. To characterize the experimental uncertainty, we examine the agreement between reported values, Fig. 4. For 92.82% of the numerical data, there at most two unique log BB values are reported as shown in Fig. 4(a,c). Similarly, for 99.34% of the molecules, only a single categorical label is reported (Fig. 4(d)); this is true even though the same molecule may appear in as many as 23 distinct sources (Fig. 4(b)). More detailed data can be found in Tables 3-6. Figure 5 reveals some features of the B3DB dataset. Presuming that the molecules in the dataset are relatively representative of (bio)organic molecules in general, the log BB for most of organic compound lie within the interval [−2, 2] (see Fig. 5(a)). The distribution of log BB values indicates that the numerical dataset is relatively balanced, though skewed towards BBB+ compounds.
Lipinski's Rule of 5 https://www.sciencedirect.com/science/article/abs/pii/S0169409X00001290?via%3Dihub is a simple rule-of-thumb for evaluating a molecule's drug-likeness. Specifically, Lipinski's Rule of 5 states that good absorption or permeation is more likely if a molecule has less than: 5 hydrogen-bond donors, 10  www.nature.com/scientificdata www.nature.com/scientificdata/ hydrogen-bond acceptors, 500 Dalton molecular weight, and a predicted log P value less than 5. It is observed that the molecule weight of most BBB+ compounds (93.10%) is less than 500 Dalton. In contrast, there are many molecules with molecular weight greater than 500 Dalton (31.22%) that are BBB− compounds. Nonetheless, aside from the a long tail of heavy BBB− compounds, the distribution of molecular weights for BBB+ and BBB− molecules is not dissimilar (see Fig. 5(b,f)). 98.8% of BBB+ compounds and 23.4% of BBB− compounds have fewer than 5 hydrogen-bond donors; 97.6% of BBB+ compounds and 66.0% of BBB− compounds have fewer than 10 hydrogen-bond acceptors. This supports the idea that hydrophilic compounds find it difficult to cross      www.nature.com/scientificdata www.nature.com/scientificdata/ have log P < 5. Taken together, the analysis of the selected physiochemical descriptors suggest that no single parameter can determine the BBB-permeability of a compound. This confirms that predicting BBB permeability computationally is challenging, and emphasizes the value of the B3DB dataset.

Usage Notes
None of the original data sources contain any quantification of uncertainty (e.g., the standard derivation), so it is recommended to incorporate the group categories when using the datasets. If one decides to use a different threshold to determine BBB+ and BBB− for a molecules, log BB can be used directly from the data reported in this study. The 1613 2D chemical descriptors, computed with mordred can facilitate building predictive models. Any further molecular preprocessing can be done with RdKit.

Code availability
The codes used in this study have been deposited to https://github.com/theochem/B3DB and https://doi. org/10.6084/m9.figshare.15634230.v3 (version 3) 33 . All the calculation were done with Python 3.7.9 under a virtual environment created with Anaconda on Linux.  Table 6. Occurrences of unique source BBB permeability labels for different groups in categorical dataset.