Drug Discovery

QM9 provides quantum chemical properties (at DFT level) for a relevant, consistent, and comprehensive chemical space of small organic molecules. This database may serve the benchmarking of existing methods, development of new methods, such as hybrid quantum mechanics/machine learning, and systematic identification of structure-property relationships. Source: QM9 Dataset Image Source: https://pubs.acs.org/doi/pdf/10.1021/ci300415d

Dataset contains 33,010 molecule-description pairs split into 80%/10%/10% train/val/test splits. The goal of the task is to retrieve the relevant molecule for a natural language description. It is defined as follows: To push the boundaries of multimodal models, we present a new IR task: \textbf{Text2Mol}. Given a text query and list of molecules without any reference textual information (represented, for example, as SMILES strings, graphs, or other equivalent representations) retrieve the molecule corresponding to the query. From a text description of a molecule, the model must incorporate the information in the description into a semantic representation which can be used to directly retrieve the molecule. This requires the integration of two very different types of information: the structured knowledge represented by text and the chemical properties present in molecular graphs. We assume there is only one correct (relevant) molecule for each description, so we consider two measures for this task: Hits@1 and mean reciprocal rank (MRR). 80% of the data is used for training. Retrieval is done against the entire corpus of molecules (train, val, test).

The Tox21 data set comprises 12,060 training samples and 647 test samples that represent chemical compounds. There are 801 "dense features" that represent chemical descriptors, such as molecular weight, solubility or surface area, and 272,776 "sparse features" that represent chemical substructures (ECFP10, DFS6, DFS8; stored in Matrix Market Format ). Machine learning methods can either use sparse or dense data or combine them. For each sample there are 12 binary labels that represent the outcome (active/inactive) of 12 different toxicological experiments. Note that the label matrix contains many missing values (NAs). The original data source and Tox21 challenge site is https://tripod.nih.gov/tox21/challenge/. Source: Tox21 Machine Learning Data Set Image Source: https://www.frontiersin.org/articles/10.3389/fenvs.2015.00080/full

The BBBP dataset comes from a study focused on modeling and predicting the permeability of the blood-brain barrier. The BBBP dataset contains binary labels indicating whether a compound can penetrate the blood-brain barrier (BBB) or not. Researchers use this dataset to develop and evaluate machine learning methods for predicting BBB permeability. It’s a critical task because understanding which compounds can cross the BBB is essential for drug discovery and designing therapeutics for neurological conditions. Check https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8708321/

ESOL is a water solubility prediction dataset consisting of 1128 samples.

The FreeSolv database offers a curated collection of experimental and calculated hydration-free energies for small molecules in water. It includes both experimental values obtained from prior literature and calculated values based on simulations. The goal is to provide accurate hydration-free energy data, which is essential for understanding solvation properties and interactions of molecules in aqueous environments.

The BACE dataset focuses on inhibitors of human beta-secretase 1 (BACE-1). It includes both quantitative (IC50 values) and qualitative (binary labels) binding results. The dataset comprises small molecule inhibitors across a wide range of affinities, spanning three orders of magnitude (from nanomolar to micromolar IC50 values). Specifically, it provides: 154 BACE inhibitors for affinity prediction. 20 BACE inhibitors for pose prediction. 34 BACE inhibitors for free energy prediction.

The HIV dataset was introduced by the Drug Therapeutics Program (DTP) AIDS Antiviral Screen, which tested the ability to inhibit HIV replication for over 40,000 compounds. Screening results were evaluated and placed into three categories: confirmed inactive (CI), confirmed active (CA), and confirmed moderately active (CM). Check https://link.springer.com/chapter/10.1007/978-3-540-89689-0_33

The ClinTox dataset compares drugs approved by the FDA and drugs that have failed clinical trials for toxicity reasons. The dataset includes two classification tasks for 1491 drug compounds with known chemical structures: (1) clinical trial toxicity (or absence of toxicity) and (2) FDA approval status.

SIDER contains information on marketed medicines and their recorded adverse drug reactions. The information is extracted from public documents and package inserts. The available information include side effect frequency, drug and side effect classifications as well as links to further information, for example drug–target relations. Source: Side Effect Resource Image Source: http://sideeffects.embl.de/drugs/2756/

QED is a linguistically principled framework for explanations in question answering. Given a question and a passage, QED represents an explanation of the answer as a combination of discrete, human-interpretable steps: sentence selection := identification of a sentence implying an answer to the question referential equality := identification of noun phrases in the question and the answer sentence that refer to the same thing predicate entailment := confirmation that the predicate in the sentence entails the predicate in the question once referential equalities are abstracted away. The QED dataset is an expert-annotated dataset of QED explanations build upon a subset of the Google Natural Questions dataset. Source: https://github.com/google-research-datasets/QED Image Source: https://github.com/google-research-datasets/QED

ToxCast is an initiative by the U.S. Environmental Protection Agency (EPA) aimed at predicting the potential toxicity of various chemical compounds. It involves high-throughput screening assays that evaluate thousands of chemicals across multiple biological endpoints. These endpoints cover a wide range of effects, including cell cycle disruptions, interactions with steroid receptors, and cytotoxicity.

The Maximum Unbiased Validation (MUV) dataset is a benchmark dataset selected from PubChem BioAssay. It was created by applying a refined nearest-neighbor analysis. The MUV dataset is specifically designed for the validation of virtual screening techniques.

BindingDB is a public, web-accessible database of measured binding affinities, focusing chiefly on the interactions of protein considered to be drug-targets with small, drug-like molecules. As of May 27, 2022, BindingDB contains 41,296 Entries, each with a DOI, containing 2,519,702 binding data for 8,810 protein targets and 1,080,101 small molecules. There are 5,988 protein-ligand crystal structures with BindingDB affinity measurements for proteins with 100% sequence identity, and 11,442 crystal structures allowing proteins to 85% sequence identity.You can also use BindingDB data through the Registry of Open Data on AWS: https://registry.opendata.aws/binding-db. This dataset using the split by TransformerCPI(doi.org/10.1093/bioinformatics/btaa524)

Language-molecule models have emerged as an exciting direction for molecular discovery and understanding. However, training these models is challenging due to the scarcity of molecule-language pair datasets. At this point, datasets have been released which are 1) small and scraped from existing databases, 2) large but noisy and constructed by performing entity linking on the scientific literature, and 3) built by converting property prediction datasets to natural language using templates. In this document, we detail the L+M-24 dataset, which has been created for the Language + Molecules Workshop shared task at ACL 2024. In particular, L+M-24 is designed to focus on three key benefits of natural language in molecule design: compositionality, functionality, and abstraction

Dataset Description: The interaction of 72 kinase inhibitors with 442 kinases covering >80% of the human catalytic protein kinome. Task Description: Regression. Given the target amino acid sequence/compound SMILES string, predict their binding affinity. Dataset Statistics: 0.3.2 Update: 25,772 DTI pairs, 68 drugs, 379 proteins. Before: 27,621 DTI pairs, 68 drugs, 379 proteins. [1] Davis, M., Hunt, J., Herrgard, S. et al. Comprehensive analysis of kinase inhibitor selectivity. Nat Biotechnol 29, 1046–1051 (2011). [2] Huang, Kexin, et al. “DeepPurpose: a Deep Learning Library for Drug-Target Interaction Prediction” Bioinformatics.

Therapeutics Data Commons is an open-science initiative with AI/ML-ready datasets and AI/ML tasks for therapeutics, spanning the discovery and development of safe and effective medicines. TDC provides an ecosystem of tools, libraries, leaderboards, and community resources, including data functions, strategies for systematic model evaluation, meaningful data splits, data processors, and molecule generation oracles. All resources are integrated via an open Python library.

Benchmark for de novo molecular design

Dataset Description: Toward making use of the complementary information captured by the various bioactivity types, including IC50, K(i), and K(d), Tang et al. introduces a model-based integration approach, termed KIBA to generate an integrated drug-target bioactivity matrix. Task Description: Regression. Given the target amino acid sequence/compound SMILES string, predict their binding affinity. Dataset Statistics: 0.3.2 Update: 117,657 DTI pairs, 2,068 drugs, 229 proteins. Before: 118,036 DTI pairs, 2,068 drugs, 229 proteins. References: [1] Tang J, Szwajda A, Shakyawar S, et al. Making sense of large-scale kinase inhibitor bioactivity data sets: a comparative and integrative analysis. J Chem Inf Model. 2014;54(3):735-743. [2] Huang, Kexin, et al. “DeepPurpose: a Deep Learning Library for Drug-Target Interaction Prediction” Bioinformatics.

CausalBench is a comprehensive benchmark suite for evaluating network inference methods on large-scale perturbational single-cell gene expression data. CausalBench introduces several biologically meaningful performance metrics and operates on two large, curated and openly available benchmark data sets for evaluating methods on the inference of gene regulatory networks from single-cell data generated under perturbations. The datasets consists of over 200000 training samples under interventions.

We have characterized 1000 human cancer cell lines and screened them with 100s of compounds. On this website, you will find drug response data and genomic markers of sensitivity. The Genomics of Drug Sensitivity in Cancer Project - http://www.cancerrxgene.org/ - was part of a Wellcome Trust-funded collaboration between The Cancer Genome Project at the Wellcome Sanger Institute (UK) and the Center for Molecular Therapeutics, Massachusetts General Hospital Cancer Center (USA). This collaboration integrated the expertise at both sites toward the goal of identifying cancer biomarkers that can be used to identify genetically defined subsets of patients most likely to respond to cancer therapies. We screened >1000 genetically characterized human cancer cell lines with a wide range of anti-cancer therapeutics. These compounds included cytotoxic chemotherapeutics as well as targeted therapeutics from commercial sources, academic collaborators, and the biotech and pharmaceutical industries. The sensitivity patterns of the cell lines were correlated with extensive genomic and expression data to identify genetic features that are predictive of sensitivity. This large collection of cell lines enabled us to capture much of the genomic heterogeneity that underlies human cancer, and which appears to play a critical role in determining the variable response of patients to treatment with specific agents.

The lipophilicity database refers to a collection of information related to the lipophilic properties of various molecules. Lipophilicity, also known as hydrophobicity, is a measure of how readily a substance dissolves in nonpolar solvents (such as oil) compared to polar solvents (such as water). In the context of drug discovery and pharmacology, understanding the lipophilicity of compounds is crucial because it affects their absorption, distribution, metabolism, and excretion (ADME) within the body.

Comparative evaluation of virtual screening methods requires a rigorous benchmarking procedure on diverse, realistic, and unbiased data sets. Recent investigations from numerous research groups unambiguously demonstrate that artificially constructed ligand sets classically used by the community (e.g., DUD, DUD-E, MUV) are unfortunately biased by both obvious and hidden chemical biases, therefore overestimating the true accuracy of virtual screening methods. We herewith present a novel data set (LIT-PCBA) specifically designed for virtual screening and machine learning. LIT-PCBA relies on 149 dose–response PubChem bioassays that were additionally processed to remove false positives and assay artifacts and keep active and inactive compounds within similar molecular property ranges. To ascertain that the data set is suited to both ligand-based and structure-based virtual screening, target sets were restricted to single protein targets for which at least one X-ray structure is available in complex with ligands of the same phenotype (e.g., inhibitor, inverse agonist) as that of the PubChem active compounds. Preliminary virtual screening on the 21 remaining target sets with state-of-the-art orthogonal methods (2D fingerprint similarity, 3D shape similarity, molecular docking) enabled us to select 15 target sets for which at least one of the three screening methods is able to enrich the top 1%-ranked compounds in true actives by at least a factor of 2. The corresponding ligand sets (training, validation) were finally unbiased by the recently described asymmetric validation embedding (AVE) procedure to afford the LIT-PCBA data set, consisting of 15 targets and 7844 confirmed active and 407,381 confirmed inactive compounds. The data set mimics experimental screening decks in terms of hit rate (ratio of active to inactive compounds) and potency distribution. It is available online at http://drugdesign.unistra.fr/LIT-PCBA for download and for benchmarking novel virtual screening methods, notably those relying on machine learning.

Comparative evaluation of virtual screening methods requires a rigorous benchmarking procedure on diverse, realistic, and unbiased data sets. Recent investigations from numerous research groups unambiguously demonstrate that artificially constructed ligand sets classically used by the community (e.g., DUD, DUD-E, MUV) are unfortunately biased by both obvious and hidden chemical biases, therefore overestimating the true accuracy of virtual screening methods. We herewith present a novel data set (LIT-PCBA) specifically designed for virtual screening and machine learning. LIT-PCBA relies on 149 dose–response PubChem bioassays that were additionally processed to remove false positives and assay artifacts and keep active and inactive compounds within similar molecular property ranges. To ascertain that the data set is suited to both ligand-based and structure-based virtual screening, target sets were restricted to single protein targets for which at least one X-ray structure is available in complex with ligands of the same phenotype (e.g., inhibitor, inverse agonist) as that of the PubChem active compounds. Preliminary virtual screening on the 21 remaining target sets with state-of-the-art orthogonal methods (2D fingerprint similarity, 3D shape similarity, molecular docking) enabled us to select 15 target sets for which at least one of the three screening methods is able to enrich the top 1%-ranked compounds in true actives by at least a factor of 2. The corresponding ligand sets (training, validation) were finally unbiased by the recently described asymmetric validation embedding (AVE) procedure to afford the LIT-PCBA data set, consisting of 15 targets and 7844 confirmed active and 407,381 confirmed inactive compounds. The data set mimics experimental screening decks in terms of hit rate (ratio of active to inactive compounds) and potency distribution. It is available online at http://drugdesign.unistra.fr/LIT-PCBA for download and for benchmarking novel virtual screening methods, notably those relying on machine learning.

Comparative evaluation of virtual screening methods requires a rigorous benchmarking procedure on diverse, realistic, and unbiased data sets. Recent investigations from numerous research groups unambiguously demonstrate that artificially constructed ligand sets classically used by the community (e.g., DUD, DUD-E, MUV) are unfortunately biased by both obvious and hidden chemical biases, therefore overestimating the true accuracy of virtual screening methods. We herewith present a novel data set (LIT-PCBA) specifically designed for virtual screening and machine learning. LIT-PCBA relies on 149 dose–response PubChem bioassays that were additionally processed to remove false positives and assay artifacts and keep active and inactive compounds within similar molecular property ranges. To ascertain that the data set is suited to both ligand-based and structure-based virtual screening, target sets were restricted to single protein targets for which at least one X-ray structure is available in complex with ligands of the same phenotype (e.g., inhibitor, inverse agonist) as that of the PubChem active compounds. Preliminary virtual screening on the 21 remaining target sets with state-of-the-art orthogonal methods (2D fingerprint similarity, 3D shape similarity, molecular docking) enabled us to select 15 target sets for which at least one of the three screening methods is able to enrich the top 1%-ranked compounds in true actives by at least a factor of 2. The corresponding ligand sets (training, validation) were finally unbiased by the recently described asymmetric validation embedding (AVE) procedure to afford the LIT-PCBA data set, consisting of 15 targets and 7844 confirmed active and 407,381 confirmed inactive compounds. The data set mimics experimental screening decks in terms of hit rate (ratio of active to inactive compounds) and potency distribution. It is available online at http://drugdesign.unistra.fr/LIT-PCBA for download and for benchmarking novel virtual screening methods, notably those relying on machine learning.

PCBA dataset 11 is a collection of high-quality dose-response data, formulated as a multitask learning benchmark from 128 high-throughput screening (HTS) assays.

Abstract: First released in 2006, DrugBank (https://go.drugbank.com) has grown to become the 'gold standard' knowledge resource for drug, drug-target and related pharmaceutical information. DrugBank is widely used across many diverse biomedical research and clinical applications, and averages more than 30 million views/year. Since its last update in 2018, we have been actively enhancing the quantity and quality of the drug data in this knowledgebase. In this latest release (DrugBank 6.0), the number of FDA approved drugs has grown from 2646 to 4563 (a 72% increase), the number of investigational drugs has grown from 3394 to 6231 (a 38% increase), the number of drug-drug interactions increased from 365 984 to 1 413 413 (a 300% increase), and the number of drug-food interactions expanded from 1195 to 2475 (a 200% increase). In addition to this notable expansion in database size, we have added thousands of new, colorful, richly annotated pathways depicting drug mechanisms and drug metabolism. Likewise, existing datasets have been significantly improved and expanded, by adding more information on drug indications, drug-drug interactions, drug-food interactions and many other relevant data types for 11 891 drugs. We have also added experimental and predicted MS/MS spectra, 1D/2D-NMR spectra, CCS (collision cross section), RT (retention time) and RI (retention index) data for 9464 of DrugBank's 11 710 small molecule drugs. These and other improvements should make DrugBank 6.0 even more useful to a much wider research audience ranging from medicinal chemists to metabolomics specialists to pharmacologists. #### More Information: - Paper: DrugBank 6.0: the DrugBank Knowledgebase for 2024 - PMID: 37953279 - PMCID: PMC10767804 - DOI: 10.1093/nar/gkad976 Knox, C., Wilson, M., Klinger, C. M., Franklin, M., Oler, E., Wilson, A., Pon, A., Cox, J., Chin, N. E. L., Strawbridge, S. A., Garcia-Patino, M., Kruger, R., Sivakumaran, A., Sanford, S., Doshi, R., Khetarpal, N., Fatokun, O., Doucet, D., Zubkowski, A., Rayat, D. Y., … Wishart, D. S. (2024). DrugBank 6.0: the DrugBank Knowledgebase for 2024. Nucleic acids research, 52(D1), D1265–D1275. https://doi.org/10.1093/nar/gkad976

Comparative evaluation of virtual screening methods requires a rigorous benchmarking procedure on diverse, realistic, and unbiased data sets. Recent investigations from numerous research groups unambiguously demonstrate that artificially constructed ligand sets classically used by the community (e.g., DUD, DUD-E, MUV) are unfortunately biased by both obvious and hidden chemical biases, therefore overestimating the true accuracy of virtual screening methods. We herewith present a novel data set (LIT-PCBA) specifically designed for virtual screening and machine learning. LIT-PCBA relies on 149 dose–response PubChem bioassays that were additionally processed to remove false positives and assay artifacts and keep active and inactive compounds within similar molecular property ranges. To ascertain that the data set is suited to both ligand-based and structure-based virtual screening, target sets were restricted to single protein targets for which at least one X-ray structure is available in complex with ligands of the same phenotype (e.g., inhibitor, inverse agonist) as that of the PubChem active compounds. Preliminary virtual screening on the 21 remaining target sets with state-of-the-art orthogonal methods (2D fingerprint similarity, 3D shape similarity, molecular docking) enabled us to select 15 target sets for which at least one of the three screening methods is able to enrich the top 1%-ranked compounds in true actives by at least a factor of 2. The corresponding ligand sets (training, validation) were finally unbiased by the recently described asymmetric validation embedding (AVE) procedure to afford the LIT-PCBA data set, consisting of 15 targets and 7844 confirmed active and 407,381 confirmed inactive compounds. The data set mimics experimental screening decks in terms of hit rate (ratio of active to inactive compounds) and potency distribution. It is available online at http://drugdesign.unistra.fr/LIT-PCBA for download and for benchmarking novel virtual screening methods, notably those relying on machine learning.

ImDrug is a comprehensive benchmark with an open-source Python library which consists of 4 imbalance settings, 11 AI-ready datasets, 54 learning tasks and 16 baseline algorithms tailored for imbalanced learning. It features modularized components including formulation of learning setting and tasks, dataset curation, standardized evaluation, and baseline algorithms. It also provides an accessible and customizable testbed for problems and solutions spanning a broad spectrum of the drug discovery pipeline such as molecular modeling, drug-target interaction and retrosynthesis.

In this paper, we propose Text-based Open Molecule Generation Benchmark (TOMG-Bench), the first benchmark to evaluate the open-domain molecule generation capability of LLMs. TOMG-Bench encompasses a dataset of three major tasks: molecule editing (MolEdit), molecule optimization (MolOpt), and customized molecule generation (MolCustom). Each task further contains three subtasks, with each subtask comprising 5,000 test samples. Given the inherent complexity of open molecule generation, we have also developed an automated evaluation system that helps measure both the quality and the accuracy of the generated molecules. Our comprehensive benchmarking of 25 LLMs reveals the current limitations and potential areas for improvement in text-guided molecule discovery. Furthermore, with the assistance of OpenMolIns, a specialized instruction tuning dataset proposed for solving challenges raised by TOMG-Bench, Llama3.1-8B could outperform all the open-source general LLMs, even surpassing GPT-3.5-turbo by 46.5% on TOMG-Bench.

This dataset provides a curated collection of approved drug Simplified Molecular Input Line Entry System (SMILES) strings and their associated protein sequences. Each small molecule has been approved by at least one regulatory body, ensuring the safety and relevance of the data for computational applications. The dataset includes 1,660 approved small molecules and their 2,093 related protein targets.