Genomics

Genomics is concerned with understanding the genes that are within a genome (it is estimated that there are 20,000–25,000 genes in the human genome), it is also concerned with how those genes interact with each other and other environmental factors.¹⁵ Specifically, genomics is concerned with interactions between loci and alleles as well as considering other key interactions such as epistasis (effect of gene-gene interactions), pleiotrophy (effect of a gene on traits), and heterosis.

Libbrecht and Noble published an article on the applications of machine learning in genetics and genomics.¹⁶ The authors discuss the different uses of supervised, semisupervised, unsupervised, generative, and discriminative approaches to modelling as well as the uses of machine learning using genetic data.¹⁶ The authors explain that machine learning algorithms can use a wide variety of genomic data, as well as being able to learn to identifier particular elements and patterns in a genetic sequence.¹⁶ Furthermore, it can be used to annotate genes in terms of their functions and understand the mechanisms behind gene expression.¹⁶

Transcriptomics

The transcriptome is the set of RNA transcripts that the genome produces in certain situations.¹⁷ Transcriptomics signals aid in understanding drug target adverse effects.¹⁸

Methods such as RNA-Seq are used to profile the transcriptome. As a method it can detect transcripts from organisms where their genomic sequence is not currently profiled and has low background signal.¹⁹ It can be used to understand differential gene expression,²⁰ RNA-Seq is supported with next generation sequencing of which allows for large numbers of read outs.²⁰

Transcriptomic data have been used in machine learning algorithms in cases such as machine learning diagnostic pipeline for endometriosis where supervised learning approaches were used on RNA-seq as well as enrichment-based DNA methylation datasets.²¹ Another use has been the development of GERAS (Genetic Reference for Age of Single-cell), which is based on their transcriptomes, the authors Singh and co-authors explain that it can assess individual cells to chronological stages which can help in understanding premature aging.²² It has also been used alongside machine learning algorithms to aid in diagnostics and disease classification of growth hormone deficiency (random forest in this case).

Metabolomics and lipomics

Metabolomics and lipidomics are concerned with the metabolome and the lipidome, respectively. Metabolomics allows for the understanding of the metabolic status and biochemical events observed in a biological, or cellular, system.¹³ Approaches in metabolomics includes the identification and quantification of known metabolites, profiling, or quantification of larger lists of metabolites (either identified or unknown compounds) or a method known as metabolic fingerprinting, of which is used to compare samples to a sample population to observe differences.²³ Metabolomics has been combined with machine learning to identify weight gain markers (again Random Forest algorithms were used).²⁴ Sen and co-authors have shown that deep learning has been applied to metabolomics in various areas such as biomarker discovery and metabolite identification (amongst others).²⁵

Lipids are grouped into eight different categories including fatty acyls, glycerolipids, glycerophosolipids, sphingolipds, saccharolipds, polyketides, sterol, and prenol lipids.²⁶ They are important in cellular functions and are complex in nature, change under different conditions such as physiological, pathological and environmental.²⁷ Lipidomics has be used to show tissue-specific fingerprints in rat,²⁶ shown potential in risk prediction and therapeutic studies²⁸ and can be used through the drug discovery process.²⁷ Fan and co-authors used machine learning with lipidomics by developing SERRF (Systematic Error Removal using Random Forest) which aids in the normalization of large-scale untargeted lipidomics.²⁹

Proteomics

Proteomics is concerned with the study of proteins. Proteomes can refer to the proteins at any level, for example, on the species level, such as all the proteins in the human species, or within a system or organ. In addition, one of the major difficulties with proteomics is its nature to change between cells and across time.³⁰ Questions may include understanding the protein expression level in the cell or identifying the proteins being modulated by a drug. Key areas of proteomic study involve, protein identification, protein structure, analysis of posttranslational modifications.

Typically a proteomic experiment is broken down into three key steps; the proteomics separation from its source such as a tissue. The acquisition of the protein structural information and finally, database utilization.³¹ Experimental procedures to separate a protein from its source involve electrophoresis where the proteins appear as lines on a gel, separated by their molecular weight.³¹ They are visualized by staining the gel and then preceded by acquiring an image of the gel. The proteins can be removed from the gel to be digested and put through a mass spectrometer. Sequencing is often completed by mass spectrometry methods, of which involves ionization of the sample, analysis of the mass, peptide fragmentation, and detection ultimately leads to database utilization. A typical global proteomics experiment involves profiling of several compounds to determine changes in particular proteins. By analyzing the observed abundance of the proteins across different treatment channels it is possible to observe treatment effects.

Swan and co-authors published applications of machine learning using proteomic data. The authors note that MS-derived proteomic data can be used in machine learning either directly using the mass spectral peaks or the identified proteins and can be used to identify biomarkers of disease as well as classifying samples.³² Gessulat and co-authors developed Prosit, a deep neural network that predicts the chromatographic retention time as well as the fragment ion intensity of peptides.³³

SDF format

SDF formats (structure data files formats) were developed by Molecular Design Limited (MDL) and are used to contain chemical information such as structure. The first section contains general information about the compound, including its name, its source and any relevant comments. The counts line has 12 fields that are of fixed length. The first two give the number of atoms and bonds described in the compound. Often Hydrogens are left implicit and can be included based on valence information.³⁸ The second block is known as the atom block (atom information encoded) and the third is known as the bond block where bond information is encoded. In the atom block, each line corresponds to each individual atom. The first three fields of each line correspond to the atoms position with its x-y-z coordinates.³⁸ Typically the atom symbol will be represented and the rest of the line relates to specific information such as charge information.³⁸ The bond blocks also have one line per individual block, and the first two fields index the atoms and the third field indicated the type of bond. The fourth refers to the stereoscopy.³⁸

InChI and InChI Key format

InChI³⁹ is a nonproprietary line notation or 1D structural representation method of which aims to be canonical identifier for structures (and thus is suitable for cross database comparisons).⁴⁰ Owing to uniqueness of InChI, it has been used to derive canonical SMILES (described later) to create something called InChIfied SMILES.⁴¹ InChI key is a hashed and condensed version of the full InChI string.

It was developed by the International Union of Pure and Applied Chemistry (IUPAC) along with the National Institute of Standards and Technology (NIST). It is continually updated by the InChI Trust. InChI captures a wide variety of compound information, not limited to its stereochemistry, charge and bond connectivity information.

InChI keys were developed to allow for searching of compounds as the full InChI is too long for this. It contains 27 characters, the first 14 corresponding to the connectivity information. Separated by a hyphen is the next eight characters that include other chemical information of the structure. The following characters (each separated by a hyphen) give information about the type of InChI, the version of it and finally, the protonation information of the compound.

It has an almost zero chance of two separate molecules having the same key. It was estimated that if 75 databases each had 1 billion structures, there would be one instance of two molecules having the same InChI key. Despite this, an example of a “collision” was identified with two compounds with different formulae and no stereochemistry.^42–44 This estimated rarity of collisions was experimentally tested and suggested that if uniqueness was desired it would probably need a longer hash.⁴⁵

SMILES and SMARTS format

The simplified molecular-input line entry system also known as (SMILES) is one of the most commonly used.^46–48 SMILES are based on molecular graph theory where the nodes of a graph are the atoms and the edges are the bonds.^{46, 47} Generic SMILES do not give details on the chirality or the isotopic nature of the structure (of which are known as isomeric SMILES).⁴⁹

One problem with SMILES is that a single structure can be represented in multiple different SMILES strings and therefore, it is recommended to use canonicalized structures to prevent one compound being identified as multiple due to the different representations used. Daylight give an example of the ways that the SMILES string CCO can be written, including OCC, [Ch3][CH2][OH], CCO, and C(O)C.⁴⁹

Daylight gives an in depth explanation of the rules for generating and understanding SMILES strings⁴⁹ and the common rules are summarized here. SMILES follow encoding rules, namely, the use of atomic symbols for atoms with aliphatic carbons being represented with a capital C and aromatic carbons being written with a lower case c. Brackets are used to describe abnormal valences and must include any attached hydrogens, as well as a number of + or – to indicate valance count. Absence of these will result in it being assumed there are zero hydrogens or charge. To indicate isotopic rules, the atomic symbol is preceded by its atomic mass such as [12C] or [13C]. On a side note, hydrogens are often omitted when writing SMILES strings and can be highlighted by either implicit nature (normal assumptions), explicit nature by either count (within brackets) or as explicit atoms themselves [H]. Bonds are represented by –, =, #, or : to depict single, double, triple, or aromatic bonds, respectively. Alternatively, atoms may be placed next to each other with the assumption that either a single or an aromatic bond separates them. To include direction, and / are used. Branching is dealt with within parentheses (of which can be nested) and cyclic structures contain a digit to indicate the breaking of a bond in the ring such as C1CCCCC1. Any disconnected structures are separated by a period. Dealing with tetrahedral centers can be represented by @ (neighbors are anticlockwise) or @@ (neighbors are clockwise) after the chiral atom. Many specific natures of compounds, such as tautomerization, chirality and shape, need to be explicitly specified in SMILES notation.

Extending on from SMILES is the SMARTS⁴⁹ notation which is designed to aid with substructure searching. SMARTS, extend atoms and bonds by including special symbols to allow for generalized identification, for example, the use of * to denote the identification of any atom or ~ to denote any bond. Many of these rules follow the rules of logical rule matching in coding languages such as the use of an explanation mark to denote NOT this, as an example, [!C] tells us to find not aliphatic carbons.

Daylight describes the difference between SMARTS and SMILES as SMARTS describing patterns and SMILES describing molecules. In addition, SMILES are valid SMARTS.

Fingerprint format

A molecular descriptor’s role is to provide one and capture similarity and differences between compounds in a chosen dataset. There are multiple kinds of molecular descriptors that range in dimensionality (0D, 1D, 2D, 3D, and 4D). A molecular fingerprint is an example of a 1D-descriptor. It is a binary string with a list of substructures or other predefined patterns.⁵⁰ They are defined before a model is trained to avoid overfitting on sparse or small datasets. If a specified pattern is found in a molecule, the corresponding bit in the binary string is set to “1,” otherwise it is set to “0.”⁵¹

Example of fingerprints are ECFP4 (extended connectivity for high dimensional data, up to four bonds), FCFP4 (functional class-based, extended connectivity), MACCS (166 predefined MDL keys), MHFP6 (for circular structures) Bayes affinity fingerprints (bioactivity and similarity searching), PubChemFP (for existence of certain substructures), KRFP (from the 5-HT 5A dataset to classify between active or inactive compounds). Sometimes it is better to create custom fingerprints than rely on predefined ones.⁵²

Essentially the features of the molecules (such as the presence of a particular atom) are extracted, hashed, and then the bits are set.⁵³ There are a wide host of available fingerprints that can be used as discussed in Table 1.

Performance of a machine learning model and prediction accuracy depends on the quality of data and descriptors and fingerprints chosen. For instance, fingerprint-based descriptors, for example, ECFP or MACCS, are recommended for active substances with functional groups located in meta or para positions.⁶⁵ For genotoxicity prediction, Support Vector Machines (SVM) models perform best with PubChemFPs. However, the authors recommend combining Random Forest (RF) and MACCS fingerprints.⁶⁶

Extended-connectivity fingerprints (ECFPs) were designed for structure-activity modeling of which are topological and circular.⁵⁵ They are related to Morgan fingerprints, but differ in their algorithm. The ECFP algorithm is well documented⁵⁵ and summarized here. Each atom is assigned an identifier of which is updated to capture neighboring atom information. Finally, any duplicate identifiers are removed (so the same feature is only represented once). Rather than a bit vector, ECFC derive a count of features.⁶⁷

In comparison, to the ECFP algorithm of which has a predetermined set of iterations, Morgan fingerprints and their algorithm⁵⁶ continue to have iterative generations until uniqueness is achieved. This process is described by Rodgers and Hahn⁵⁵ in their extended-connectivity fingerprints paper where they explain that for Morgan fingerprints, their atom identifiers are not dependent on the atoms original numbering and uses identifiers from previous iterations after encoding invariant atom information into an initial identifier. Essentially the Morgan algorithm iterates through each atom and captures information about all possible paths through the atom, given a predetermined radius size.⁶⁸ Morgan fingerprints were designed to address molecular isomorphism⁵⁵ and are often used for comparing molecular similarity. These are hashed into a bit vector length (also predetermined). The iterative process involves each atom identifier in a compound and updating the information about it. For example, at iteration 0, only information about the atom is captured (as well as related bonds) whereas as the iterations increase, so does the information about the atom’s neighbors, and so on.

Two other popular fingerprints are MACCS keys and Daylight fingerprints. The Molecular ACCess System (MACCS) keys is a predefined set of 116 substructures.⁵⁴ A problem with the MACCS keys is that there is no publication that defines what each of the 116 substructures are. Generally, when citing, individuals refer to a paper discussing the re-optimization of MDL keys.^{54, 69} Daylight fingerprints are a form of path fingerprints which enumerate across the paths of a graph and translate them into a bit vector.⁷⁰ Signature fingerprints are not binary and are based on extended valence sequence.⁵⁸ They are topological descriptors that also describe the connectivity of the atoms within a compound.⁷¹

Other descriptors

A molecular descriptor can be derived from experimental data or calculated theoretically. Examples of nonfingerprint molecular descriptors include reactivity, shape, binding properties, atomic charges, molecular orbital energies, frontier orbital densities, molar refractivity, polarization, charge transfer, dipole moment, electrophilicity,⁷² molecular and quantum dynamics.⁷³

Molecular descriptors are generated with the use of tools, for example, PaDEL-Descriptor,⁷⁴ OpenBabel,⁷⁵ RDKit,^{36, 53} CDKit,⁷⁶ and E-Dragon.⁷⁷

Structural 2D descriptors perform well in models handling binary information such as classification and class probability estimation models⁷⁸ and in association rules learning.^79–81 There exists no universal descriptor that works best with every prediction model. However, various descriptor types can be combined as input data for a model to achieve higher performance.

There are various commercial and open-source software, databases, and servers that use molecular descriptors to predict toxic endpoints: OECD QSAR Toolbox,⁸² Derek Nexus,⁸³ FAF-Drugs4,⁸⁴ eTOXsys,⁸⁵ TOXAlerts,⁸⁶ Schrödinger’s CombiGlide^87–89 Predictor, Leadscope Hazard Expert,⁹⁰ VEGA,⁹¹ METEOR.⁸³ ChemBench,⁹² ChemSAR,⁹³ ToxTree,⁹⁴ Lazar,⁹⁵ admetSAR,⁹⁶ Discovery Studio⁹⁷ and Pipeline Pilot⁹⁸ are ML-based tools. For more detailed information, please refer to review on computational methods in HTC by Hevener, 2018.⁹⁹

Furthermore, descriptors can also be calculated for protein structures. Local descriptors have been shown to aid in the characterization of amino acid neighborhoods.¹⁰⁰ The tool ProtDCal calculates numerical sequence and structure based descriptors of proteins.¹⁰¹ Another publication had the authors develop a sequence descriptor (in matrix form) alongside a deep neural network that could be used for predicting protein-protein interactions.¹⁰²

Similarity measures

It is often a requested task to compare the similarity of two compounds. Different similarity metrics are summarized in Table 2. Similarity can be rephrased as comparing the distance between the compounds to evaluate how different two compounds are. For fingerprint-based similarity calculations, Tanimoto index is a popular method.¹¹⁰ A study compared several of these metrics comparing molecular fingerprints. They identified that the Tanimoto index, Dice index, Cosine coefficient and Soergel distance to be best and recommended that Euclidean and Manhattan distances not be used on their own.¹¹⁰

Table 2

Table of different similarity metrics.

Name	Equation	Equation information
Tanimoto/Jaccard103, 104	$T\left.a\text{,}b\right)=\frac{N_c}{N_a+N_b-N_c}$	N = number of attributes in objects a and b C = intersection set
Tversky105	$similarity\left.A\text{,}B\right)=\frac{AB}{aA+bB+AB}$	α = weighs the contribution of the first reference molecule The similarity measure is asymmetric106
Dice107	$similarity\left.A\text{,}B\right)=\frac{2\ast AB}{A+B+AB}$	AB is bits present in both A and B
Manhattan106	$similarity\left.A\text{,}B\right)=\frac{A+B}{A+B+AB+!A!B}$	The more similar the fingerprint the lower the similarity score (acting more like a distance measure)106
Euclidean distance106	$Dist\left.A\text{,}B\right)=\sqrt{\frac{AB+!A!B}{A+B+AB+!A!B}}$	!A!B represents the bits that are absent in both A and B
Cosine108, 109	$similarity\left.x\text{,}y\right)=cos\left.\theta \right)=\frac{x.y}{\left.\left.x\right|\right|\ast \left.\left.y\right|\right|}$	x = compound x y = compound y

The reason for comparing the similarity of compounds is that, in combinatorial library design, chemists may reject compounds that have a Tanimoto coefficient ≥  0.85 similar to another compound already chosen from the library.¹¹¹ This is for the purpose of ensuring structural diversity within the library. A study showed that by using Daylight fingerprints, and Tanimoto similarity, found that there was only a 30% chance that two compounds that were highly similar were both active, likely due to differences in target interactions.¹¹¹

$Similarity=\frac{1}{1+distance}$

Eq. (1) is used for calculating the similarity of two compounds.

Toxicity related databases

As a result of the application of high throughput screening (HTS) and development of novel chemical and biological research techniques in the 21st century, a number of publicly available repositories is rapidly growing. This enables integration of siloed information and prediction of less evident side effects resulting from synergistic effects, and complex drug-drug interactions can be discovered. In this section, we present an overview of existing data sources related to toxicogenomics, organ toxicity, binding affinity, biochemical pathways, bioactivity, molecular interactions, gene-disease linkage, histopathology, oxidative stress, protein-protein interactions, metabolomics, transcriptomics, proteomics, and epigenomics (Table 3).

TOXNET¹⁸⁷ is an aggregator of other toxicity-related databases on breastfeeding and drugs, developmental toxicology literature, drug-induced liver injury, household product safety, and animal testing alternatives. TOXNET is available via PubMed since December 2019. ToxCast¹⁵⁸ and ECOTOX¹⁶¹ are two databases created by the US Environmental Protection Agency. They contain high-throughput and high-level cell response data related to toxicity and environmental impact of over 1800 chemicals, consumer products, food and cosmetic additives. Tox21¹⁵⁶ is a collaborative database between some of the US Federal Agencies that aggregates toxicology data on commercial chemicals, pesticides, food additives, contaminants, and medical compounds? ToxBank Data Warehouse¹⁵⁷ stores systemic pharmacology information and additionally integrates into models predicting repeated-dose toxicity. PubChem¹⁵³ and DrugBank¹⁶² are not purely toxicology databases; however, they collect bioactivity and biomolecular interactions data as well as clinical and patent information, respectively. ChEMBL¹⁸⁸ and CTD¹⁴⁷ databases contain manually curated data on chemical molecule and gene or protein interactions, chemical molecule and disease as well as gene and disease relationships. There exist various online public resources devoted to drug side effects: SIDER,¹²⁰ OFF-SIDES,¹⁸⁹ and CEBS.⁶⁶ These data are integrated with pathway-focused sites, for example, KEGG,¹⁹⁰ PharmGKB,¹³⁸ and Reactome,¹⁴¹ which are curated and peer-reviewed pathway databases. The following table contains the main ones, however, it is not exhaustive.

A large number of molecular-omics data is present in the public domain and allow for reusing and exchange data from between experiments. High-dimensional and noisy biological signals used in, for example, differential gene expression, gene co-expression networks, compound protein-protein interaction networks, signature matching and organ toxicity analysis, often require a standardized ontology as well as manual data curation before they can be used to train a model.¹⁸ However, the following public databases offer relatively high-quality data. DrugMatrix¹⁹¹ contains in vivo rat liver, kidney, heart and thigh muscle from Affymetrix GeneChip Rat Genome 230 2.0 Array GE Codelink and Open TG-GATEs¹⁶⁷ contain rat liver and kidney data. The latter also contains human and rat in vitro hepatocytes histopathology, blood chemistry and clinical chemistry data. Toxicity data for five human cancer cell lines derived from the Affymetrix GeneChip Human Genome U133A Array are stored in the Connectivity Map.¹⁶⁴ Microscopy images of up to 77 cell lines treated with various chemical compounds and gene expression data can be found in the Library of Integrated Network-based signatures L1000 (LINCS dataset).¹⁶⁴

Many of the resources above have multiple applications. A wide variety of resources are available for proteomic studies from the EBI including UniProt KnowledgeBase (UniProtKB) and PRIDE.¹⁹² Uniprot provides freely accessible resources of protein data such as protein sequences and functional information. UniProtKB is included in these resources.^170–172 It is split into two sections namely, the manually annotated and reviewed section known as UniProtKB/Swiss-Prot. The second section, UniProtKB/TrEMBL refers to the computationally annotated and nonreviewed section of the data. Owing to be computationally annotated, EBI states that there is high annotation coverage of the proteome.¹⁷² These data can be used to find evidence for protein function or subcellular location.¹⁷² Finally, PRIDE incudes protein and peptide identifications (such as details of posttranslational modifications) alongside evidence from mass spectrometry.^175–177

This growth in the number of data repositories and databases has been fueled by the large amount of proteomic data generated.¹⁷⁵ The Protein DataBank is concerned with structural protein information such as the 3D shape if the protein and is maintained by the RCSB.^{173, 174}

To deal with this, The HUPO Proteomics Standards Initiative,¹⁹³ or HUPO-PSI for short, was developed to ensure the universal adoption of stable data formats that has resulted in aggregation of proteomic data.¹⁷⁵ The HUPO-PSI’s about section states that these standards were developed “to facilitate data comparison, exchange and verification”.¹⁹³ However, it does not deal with the quality of data and the issues that brings.

Other key resources include WITHDRAWN¹⁸⁰ of which contains information about withdrawn and discontinued drug, DISGeNET¹⁸¹ and Open Targets¹⁸² for target-disease relationships. The Clinical Pharmacology and British Pharmacology Society Guide to Pharmacology Database¹⁸³ also contains target information and information about a variety of ligands. GOSTAR¹⁸⁵ and SureChEMBL¹⁹⁴ both contain information on target-compound information from patents with GOSTAR also containing that information available from literature.

Drug safety databases

To monitor, systematically review, and enable data-driven decisions on drug safety, WHO Collaborating Monitoring Centre in Uppsala¹⁹⁵ and National Competent Authorities (NCAs) maintain several databases dedicated to safety signals collection (Fouretier et al., 2016). The largest and the oldest ones are WHO VigiBase (1968), EU Eudravigilance, FDA FAERS, and VAERS, but most countries have established their own databases supported by Geographical Information Systems (GIS).¹⁹⁶ Geolocalization allows using these databases to detect both global and local trends. Table 4 presents an overview of the largest publicly accessible databases related both to postmarketing surveillance, unsolicited reporting, and solicited reporting from clinical trials.

Majority of the unsolicited resources is unstructured, fragmentary, unstandardized and suffering from the presence of confounders. Although WHO, ICH, and NCAs have taken a considerable standardization effort, the quality of ADR, reports vary across countries.¹⁹⁷ Additional curation of the data in indispensable as databases contains duplicates, missing data points, and it has high sample variance.

Furthermore, cases when patients were administered drugs as intended and no ADR occurred, are naturally not reported. From the perspective of data analysis and developing machine learning models lack their presence in a dataset results in class imbalance, survivorship bias and high numbers of false positives in predictions.¹⁹⁸ Thus, one cannot calculate the rate of occurrence for the whole population basing on spontaneous resources only. Otherwise, the risk of false-positive reporting for certain medicines may be artificially elevated.¹⁹⁹ Finally, statistical significance in a model does not always mean clinical relevance. A majority of patients might be likely to respond better to certain medications statistically. However, some atypical side effects may occur that lower the quality of life of a small number of patients and hence outweigh the benefits.

Finally, longitudinal patient medical history may not always be easily retrieved, and thus it is challenging to verify reported information as well as establish causality understood as in ICH-E2A guideline.²⁰⁰ Reports submitted to SRS databases are subjective and often contain inconsistent records when compared with original medical documentation.

Key public data-resources for precision medicine

This section describes many completed and ongoing efforts to generate large-scale datasets from cell lines, patients and healthy volunteers. These datasets are a necessary asset that will be used to generate novel AI/ML-based models to guide precision medicine.

Resources for accessing metadata and analysis tools

Accessing and analyzing raw sequencing data can be quite cumbersome for most biologists. Resources that present analyzed or easy to grasp data on genetic alterations as well as pathway level analysis are very helpful. Several such resources that can be used directly for hypothesis generation/verification exist. Some of these are listed in Table 7.

Table 7

Resources for accessing metadata and analysis tools.

Database	Content	Omics readout	Weblink	Last update	Reference
COSMIC	Tumor samples and >  1000 cell lines	Expert curated database of somatic mutations	https://cancer.sanger.ac.uk/cosmic	V92, August 2020	233
Cancer DepMap	Data from CCLE, TCGA, GDSC, RNAi and CRISPR screens	Genomics, proteomics, RNAi/CRISPR screens and drug sensitivity	https://depmap.org/portal/	Every 90 days
Cell Model Passports	Cell lines and organoids	Mutations, expression, CNV, methylation, fusions, drug response, CRISPR score	https://cellmodelpassports.sanger.ac.uk/passports		234
cBioPortal	The portal hosts a total of 263 cancer studies including CCLE and TCGA data	Mutations, CNV, RNAseq, RPPA	http://www.cbioportal.org/		235
TCGA-CDR	Clinical data resource for high quality survival outcome analytics	Survival data	See reference		236
mSigDB	Annotated gene sets for use with GSEA	Gene sets	http://software.broadinstitute.org/gsea/msigdb/index.jsp		237
Enricher	Annotated gene sets for use with GSEA	Gene sets	https://amp.pharm.mssm.edu/Enrichr/		238
GeneMANIA	Hypothesis generation regarding function of a gene	Multiple omics-based data	https://genemania.org/		239
L1000CDS2	LINCS L1000 characteristic direction signature search engine	Finds consensus L1000 small molecule signatures that match user input signatures	https://amp.pharm.mssm.edu/L1000CDS2/#/index		240
Geneshot	Ranking genes based on text mining	Literature, expression data	https://amp.pharm.mssm.edu/geneshot/		241

Fig. 2 recapitulates progress on data generation frontier that include drug screening in cell lines, functional genomics (RNAi and CRISPR) screens, detailed characterization of cell lines and finally exome or whole genome sequencing of patients and healthy volunteers. Some of these data were already used employing AI/ML-based approaches to identify novel synthetic lethality pairs, predict drug IC50, or even clinical outcome prediction.^{207, 242, 243} By designing an AI algorithm to analyze CT scan images, researchers have created a radiomic signature that defines the level of lymphocyte infiltration of a tumor and provides a predictive score for the efficacy of immunotherapy in the patient.²⁴⁴ Gene expression profile analysis of needle biopsy specimens was performed from the livers of 216 patients with hepatitis C-related early-stage cirrhosis who were prospectively followed up for a median of 10 years. Evaluation of 186-gene signature used to predict outcomes of patients with hepatocellular carcinoma showed this signature is also associated with outcomes of patients with hepatitis C-related early-stage cirrhosis.²⁴⁵ Recently, whole-genome sequencing was used to accurately predict profiles of susceptibility to first-line antituberculosis drugs.²⁴⁶

Table 8 lists some of the examples of historical data sets, potential methods to analyze them, and their respective applications in biopharma. The recent innovation in the field of AI has been enabled primarily by the confluence of rapid advances in affordable computing power in the form of cloud computing, infrastructure to process and manage large-scale data sets and architectures and methodologies such as neural networks.

Table 8

Selected examples of historical data sets, potential methods to analyze them and their respective applications in biopharma.

Examples	Data type	Data and methods	Applications in biopharma
National Biomedical Imaging Archive (NBIA); GenomeRNAi	Imaging data	Image preprocessing and analyses, data annotation, data extraction, segmentation, deep learning, computer vision	Clinical or cellular phenotyping, patient stratification and disease subclassification
TCGA; dbGAP	Genomic data	Variant calling, annotation, structural variants differential expression	Diagnosis, disease subtyping, therapeutic matching, clinical trial matching
UK Biobank; BioMe Biobank	Biobanks and electronic health records	Clinical trajectory estimation, biomarker-based modeling	Predict risk of diseases, real world evidence modeling
ClinicalTrials.gov; AACT Database	Clinical trials databases	Clinical trial protocols, performance metrics, patient population summaries	Predictive modeling of clinical trial metrics