Back

MoleculeNet: a benchmark for molecular machine learning

Zhenqin WuBharath RamsundarEvan N. FeinbergJoseph GomesCaleb GeniesseAneesh S. PappuKarl LeswingVijay Pande

article2017Chemical Science2,715 citations

Presents MoleculeNet, a standardized open-source benchmark integrated into DeepChem that provides common datasets, evaluation metrics, and baseline implementations to systematically compare machine learning algorithms across diverse molecular, biophysical, and quantum mechanical tasks.

Listen

The article addresses the challenge that molecular machine learning lacks standardized benchmarks, making it difficult to compare new algorithms fairly. Molecular datasets are typically small, heterogeneous, and expensive to obtain, while input molecules vary widely in structure and the properties to predict range from quantum mechanics to physiological effects on the human body. Without common datasets, metrics, and evaluation protocols, progress has been slow despite growing interest in data-driven methods for chemistry.

The work introduces MoleculeNet, a benchmark that curates 17 public datasets covering over 700,000 compounds across four categories—quantum mechanics, physical chemistry, biophysics, and physiology—along with defined metrics, data splits, and open-source implementations of featurization and learning methods integrated into the DeepChem library.

The authors evaluated conventional models such as random forests, kernel SVMs, and multitask networks alongside graph-based models that use learnable featurizations. They tested multiple splitting strategies including random, scaffold, stratified, and time-based splits, and reported results using appropriate regression and classification metrics such as MAE, RMSE, ROC-AUC, and PRC-AUC.

Graph-based models with learnable representations delivered the strongest results on 11 of the 17 datasets, often outperforming conventional approaches by clear margins when sufficient data were available. However, these models showed reduced robustness on smaller datasets and under high class imbalance. Physics-informed featurizations proved more critical than the choice of learning algorithm for quantum and biophysical tasks. Data-driven methods also matched or exceeded traditional ab-initio calculations on physical chemistry properties once moderate amounts of training data were provided.

These findings indicate that learnable featurizations offer a scalable path forward for molecular property prediction, yet additional data, improved handling of imbalance, and task-specific features remain necessary for biophysical and physiological applications. The benchmark establishes a common platform that can accelerate method development in both chemistry and machine learning, similar to the role ImageNet played in computer vision.

The authors recommend extending MoleculeNet to cover protein structure prediction and DNA modeling, encouraging community contributions of new datasets and algorithms to DeepChem, and collecting more high-quality data for physiology tasks. Limitations include the computational cost of some graph models, potential overfitting on small datasets, variability across data splits, and the fact that reported numbers reflect fixed training-time constraints rather than fully optimized runs from the original papers.

Schrödinger, Inc.Stanford Universityclassificationlibrariesmachine learningmolecular machine learningMoleculeNet datasetscomputer-aided molecular design
arXiv: 1703.00564deepchem/deepchem
  • Paper: Convolutional Networks on Graphs for Learning Molecular Fingerprints, David Duvenaud et al. (2015). Introduces differentiable neural molecular fingerprints directly from graph structures, establishing the learnable molecular representation paradigm benchmarked in MoleculeNet.
  • Paper: Neural Message Passing for Quantum Chemistry, Justin Gilmer et al. (2017). Formulates the Message Passing Neural Network (MPNN) framework for quantum chemistry on molecular graphs, providing one of the core graph-based model architectures evaluated across MoleculeNet.
  • Paper: The Graph Neural Network Model, Franco Scarselli et al. (2009). Establishes the foundational graph neural network architecture that operates directly on irregular molecular graph structures without fixed-size vector flattening.
  • Paper: Weisfeiler-Lehman Graph Kernels, N. Shervashidze et al. (2011). Presents the Weisfeiler-Lehman subtree graph kernel framework, which forms the theoretical and methodological basis for discrete structural comparisons and baseline graph featurizations.
  • Paper: Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules, R. Gómez-Bombarelli et al. (2016). Demonstrates data-driven continuous latent representations of molecules using variational autoencoders on standard datasets like QM9 and ZINC, framing the representation learning goals MoleculeNet evaluates.
  • Paper: Learning Convolutional Neural Networks for Graphs, Mathias Niepert et al. (2016). Develops PATCHY-SAN for convolutional learning over arbitrary graphs via neighborhood normalization, providing key context for non-spectral spatial graph deep learning baselines.
  • Paper: An Overview of Multi-Task Learning in Deep Neural Networks, Sebastian Ruder (2017). Synthesizes hard and soft multi-task deep learning mechanisms, explaining the multi-task molecular property prediction formulations evaluated throughout MoleculeNet.
  • Paper: XGBoost: A Scalable Tree Boosting System, Tianqi Chen et al. (2016). Details the scalable tree boosting system XGBoost, which serves as a standard non-deep learning baseline for tabular and fingerprint-based molecular property prediction.
Cover for MoleculeNet: a benchmark for molecular machine learning

Abstract

Molecular machine learning has been maturing rapidly over the last few years. Improved methods and the presence of larger datasets have enabled machine learning algorithms to make increasingly accurate predictions about molecular properties. However, algorithmic progress has been limited due to the lack of a standard benchmark to compare the efficacy of proposed methods; most new algorithms are benchmarked on different datasets making it challenging to gauge the quality of proposed methods. This work introduces MoleculeNet, a large scale benchmark for molecular machine learning. MoleculeNet curates multiple public datasets, establishes metrics for evaluation, and offers high quality open-source implementations of multiple previously proposed molecular featurization and learning algorithms (released as part of the DeepChem open source library). MoleculeNet benchmarks demonstrate that learnable representations are powerful tools for molecular machine learning and broadly offer the best performance. However, this result comes with caveats. Learnable representations still struggle to deal with complex tasks under data scarcity and highly imbalanced classification. For quantum mechanical and biophysical datasets, the use of physics-aware featurizations can be more important than choice of particular learning algorithm.

Topics

classificationlibrariesmachine learning

Keywords

molecular machine learningMoleculeNet datasetscomputer-aided molecular designmolecular graphsopen source DeepChemTensor Neural NetworksFreeSolv datasetlearnable featurizationshyperparameter optimization

Table of Contents

  • 1 Introduction
  • 2 Related Work
  • 3 Methods
  • 3.1 Datasets
  • 3.1.1 QM7/QM7b
  • 3.1.2 QM8
  • 3.1.3 QM9
  • 3.1.4 ESOL
  • 3.1.5 FreeSolv
  • 3.1.6 Lipophilicity
  • 3.1.7 PCBA
  • 3.1.8 MUV
  • 3.1.9 HIV
  • 3.1.10 PDBbind
  • 3.1.11 BACE
  • 3.1.12 BBBP
  • 3.1.13 Tox21
  • 3.1.14 ToxCast
  • 3.1.15 SIDER
  • 3.1.16 ClinTox
  • 3.2 Dataset splitting
  • 3.3 Metrics
  • 3.4 Featurization
  • 3.4.1 ECFP
  • 3.4.2 Coulomb Matrix
  • 3.4.3 Grid Featurizer
  • 3.4.4 Symmetry Function
  • 3.4.5 Graph Convolutions
  • 3.4.6 Weave
  • 3.5 Models - Conventional Models
  • 3.5.1 Logistic Regression
  • 3.5.2 Support Vector Classification
  • 3.5.3 Kernel Ridge Regression
  • 3.5.4 Random Forests
  • 3.5.5 Gradient Boosting
  • 3.5.6 Multitask/Singletask Network
  • 3.5.7 Bypass Multitask Networks
  • 3.5.8 Influence Relevance Voting
  • 3.6 Models - Graph Based Models
  • 3.6.1 Graph Convolutional models
  • 3.6.2 Weave models
  • 3.6.3 Directed Acyclic Graph models
  • 3.6.4 Deep Tensor Neural Networks
  • 3.6.5 ANI-1
  • 3.6.6 Message Passing Neural Networks
  • 4 Results and Discussion
  • 4.1 Physiology and Biophysics Tasks
  • 4.2 Biophysics Task - PDBbind
  • 4.3 Physical Chemistry Tasks
  • 4.4 Quantum Mechanics Tasks
  • 5 Conclusion
  • 6 Appendix
  • 6.1 Model Training and Hyperparameter Optimization
  • 6.1.1 Logistic Regression (Logreg)
  • 6.1.2 Support Vector Classification (KernelSVM)
  • 6.1.3 Kernel Ridge Regression (KRR)
  • 6.1.4 Random Forest (RF)
  • 6.1.5 Gradient Boosting (XGBoost)
  • 6.1.6 Multitask/Singletask Networks
  • 6.1.7 Bypass Networks
  • 6.1.8 Influence Relevance Voting (IRV)
  • 6.1.9 Graph Convolutional models (GC)
  • 6.1.10 Weave models
  • 6.1.11 Deep Tensor Neural Networks (DTNN)
  • 6.1.12 Directed Acyclic Graph models (DAG)
  • 6.1.13 Message Passing Neural Networks (MPNN)
  • 6.1.14 ANI-1
  • 6.2 Performances
  • 6.3 Grid Featurizer
  • 6.4 ClinTox
  • 6.5 Dataset and model access
  • 6.6 Model validation
  • 6.6.1 Graph Convolutional models
  • 6.6.2 Directed Acyclic Graph models
  • 6.6.3 Weave models
  • 6.6.4 Deep Tensor Neural Network
  • 6.6.5 Message Passing Neural Network
  • 6.6.6 Influence Relevance Voting
  • References

Knowls

  1. Knowl 1 — MoleculeNet Benchmark Dataset Suite

    data/table

    MoleculeNet is a curated benchmark suite for molecular machine learning comprising 17 datasets across four functional tiers: quantum mechanics, physical chemistry, biophysics, and physiology, encompassing over 700,000 compounds and over 800 learning tasks.

    Category Dataset Data Type # Tasks Task Type # Compounds Rec. Split Rec. Metric
    Quantum Mechanics QM7 SMILES, 3D 1 Regression 7,160 Stratified MAE
    QM7b 3D coordinates 14 Regression 7,210 Random MAE
    QM8 SMILES, 3D 12 Regression 21,786 Random MAE
    QM9 SMILES, 3D 12 Regression 133,885 Random MAE
    Physical Chemistry ESOL SMILES 1 Regression 1,128 Random RMSE
    FreeSolv SMILES 1 Regression 642 Random RMSE
    Lipophilicity SMILES 1 Regression 4,200 Random RMSE
    Biophysics PCBA SMILES 128 Classification 437,929 Random PRC-AUC
    MUV SMILES 17 Classification 93,087 Random PRC-AUC
    HIV SMILES 1 Classification 41,127 Scaffold ROC-AUC
    PDBbind SMILES, 3D 1 Regression 11,908 Time RMSE
    BACE SMILES 1 Classification 1,513 Scaffold ROC-AUC
    Physiology BBBP SMILES 1 Classification 2,039 Scaffold ROC-AUC
    Tox21 SMILES 12 Classification 7,831 Random ROC-AUC
    ToxCast SMILES 617 Classification 8,575 Random ROC-AUC
    SIDER SMILES 27 Classification 1,427 Random ROC-AUC
    ClinTox SMILES 2 Classification 1,478 Random ROC-AUC

    Each benchmark task specifies recommended dataset splitting strategies and primary evaluation metrics tailored to its data characteristics and operational use cases.

  2. Knowl 2 — Molecular Dataset Splitting Mechanisms

    model/method

    MoleculeNet establishes four standardized dataset splitting strategies to evaluate molecular machine learning models under realistic chemical and operational constraints, partitioning all datasets into training, validation, and test subsets according to an 80/10/10 ratio:

    • Random Splitting: Allocates molecular samples uniformly at random across the training, validation, and test partitions.
    • Scaffold Splitting: Groups molecules based on their two-dimensional Bemis-Murcko core structural frameworks generated by RDKit, placing structurally distinct chemotypes into separate partitions. This provides a rigorous test of out-of-distribution generalization compared to random partitioning.
    • Stratified Splitting: Applied to real-valued regression datasets (such as QM7). Molecular samples are ordered by ascending continuous label value and evenly assigned across training, validation, and test partitions so that every subset covers the full numeric range of the target property.
    • Time Splitting: Partitions datasets chronologically based on experimental deposition date (applied to protein-ligand complexes in PDBbind), training models on older structures and evaluating them on newer structures to simulate prospective real-world discovery workflows.

  3. Knowl 3 — Metric Selection Protocol for Molecular Classification Under Class Imbalance

    model/method

    In molecular virtual screening and bioactivity datasets, target properties often display severe class imbalance, with active compounds representing a tiny fraction of the total dataset. MoleculeNet establishes a decision threshold for evaluation metrics based on positive class prevalence:

    • If the positive label rate is strictly less than 2%2\%, the Area Under the Precision-Recall Curve (PRC-AUC) is the recommended metric.
    • If the positive label rate is greater than or equal to 2%2\%, the Area Under the Receiver Operating Characteristic Curve (ROC-AUC) is the recommended metric.

    This decision rule arises from the differing denominators of the False Positive Rate (FPR\text{FPR}) and Precision\text{Precision}:

    FPR=FPFP+TN\text{FPR} = \frac{\text{FP}}{\text{FP} + \text{TN}}

    Precision=TPTP+FP\text{Precision} = \frac{\text{TP}}{\text{TP} + \text{FP}}

    where TP\text{TP} is true positives, FP\text{FP} is false positives, and TN\text{TN} is true negatives. In large-scale screening datasets with extreme negative dominance, the massive TN\text{TN} count suppresses FPR\text{FPR}, masking significant false positive burdens in ROC curves. Precision explicitly balances TP\text{TP} against FP\text{FP}, penalizing high false positive rates in severe imbalance settings.

  4. Knowl 4 — Grid Featurizer for Protein-Ligand Complex Binding Interactions

    model/method

    The Grid Featurizer computes a fixed-length 2052-dimensional structural fingerprint for co-crystallized protein-ligand complexes (such as those in PDBbind) by characterizing binding site topology, intra-molecular features, and inter-molecular contact geometries:

    1. Binding Pocket Extraction: All protein atoms residing within a 4.5 A˚4.5\text{ \AA} Euclidean distance threshold from any atom in the ligand are extracted as active binding pocket atoms.
    2. Intra-molecular Circular Fingerprints (512 dimensions): Standard circular fingerprints of radius 2 are generated independently for ligand atoms and protein pocket atoms, and then hashed together into a joint 512-dimensional vector.
    3. Inter-molecular Contact Fingerprints (1536 dimensions): Contacting ligand-protein atom pairs are partitioned into three pairwise Euclidean distance intervals: 0–2 A˚0\text{--}2\text{ \AA}, 2–3 A˚2\text{--}3\text{ \AA}, and 3–4.5 A˚3\text{--}4.5\text{ \AA}. Interacting atom pairs within each distance bin are hashed into a 512-dimensional fingerprint, producing 3×512=15363 \times 512 = 1536 features.
    4. Chemical Interaction Descriptors (4 dimensions): The total number of salt bridges formed between protein and ligand is recorded as 1 feature, and hydrogen bonds are counted across three discrete distance bins (3 features).

    Concatenating these feature sets yields a descriptor of length 512+1536+4=2052512 + 1536 + 4 = 2052.

  5. Knowl 5 — Coulomb Matrix Representation of Molecular Geometry

    equation

    The Coulomb Matrix M∈RN×NM \in \mathbb{R}^{N \times N} encodes the 3D conformation of an NN-atom molecule using nuclear charges ZIZ_I and Cartesian atomic coordinates RI∈R3\mathbf{R}_I \in \mathbb{R}^3 via the formula:

    MIJ={0.5ZI2.4for I=JZIZJ∥RI−RJ∥for I≠JM_{IJ} = \begin{cases} 0.5 Z_I^{2.4} & \text{for } I = J \\ \frac{Z_I Z_J}{\|\mathbf{R}_I - \mathbf{R}_J\|} & \text{for } I \ne J \end{cases}

    where diagonal elements approximate the self-energy of each isolated atom II through a power-law fit of its nuclear charge ZIZ_I, and off-diagonal elements represent the electrostatic Coulomb repulsion between distinct nuclei II and JJ.

    The Coulomb Matrix is invariant under 3D spatial translations and rotations of the molecular coordinate frame. Full invariance to atom indexing permutations is achieved during model training either by sorting rows and columns by their Euclidean norm or by applying random row/column permutations paired with binary expansion transforms.

  6. Knowl 6 — ClinTox Benchmark Dataset

    definition

    The ClinTox dataset contains 1,491 chemical compounds with verified chemical structures, designed as a benchmark for predicting clinical trial outcomes and regulatory success. The dataset defines two binary classification tasks:

    1. Clinical Trial Toxicity Failure: Identifies whether a compound failed during human clinical trials due to adverse safety or toxicity endpoints. Compound names and outcomes are derived from the Aggregate Analysis of ClinicalTrials.gov (AACT) database by selecting trials marked as terminated, suspended, or withdrawn where the documented rationale included "adverse", "toxic", or "death".
    2. FDA Approval Status: Identifies whether a compound has achieved regulatory approval by the US Food and Drug Administration (FDA), compiled from annotations in the SWEETLEAD database.

  7. Knowl 7 — Comparative Benchmark Performance: Graph-Based versus Conventional Models

    data/table

    MoleculeNet benchmarks conventional machine learning models using static descriptors (Extended-Connectivity Fingerprints ECFP4, Coulomb Matrix CM, and Grid Featurizer) against graph-based deep architectures (Graph Convolution GC, Weave, DAG, DTNN, ANI-1, MPNN) across all 17 datasets:

    Category Dataset Metric Best Conventional Best Graph-Based
    Quantum Mechanics QM7 MAE KRR(CM): 10.22 ANI-1: 2.86
    QM7b MAE KRR(CM): 1.05 DTNN: 1.77*
    QM8 MAE Multitask: 0.0150 MPNN: 0.0143
    QM9 MAE Multitask(CM): 4.35 DTNN: 2.35
    Physical Chemistry ESOL RMSE XGBoost: 0.99 MPNN: 0.58
    FreeSolv RMSE XGBoost: 1.74 MPNN: 1.15
    Lipophilicity RMSE XGBoost: 0.799 GC: 0.655
    Biophysics PCBA AUC-PRC Logreg: 0.129 GC: 0.136
    MUV AUC-PRC Multitask: 0.184 Weave: 0.109
    HIV AUC-ROC KernelSVM: 0.792 GC: 0.763
    BACE AUC-ROC RF: 0.867 Weave: 0.806
    PDBbind (full) RMSE RF(grid): 1.25 GC: 1.44
    Physiology BBBP AUC-ROC KernelSVM: 0.729 GC: 0.690
    Tox21 AUC-ROC KernelSVM: 0.822 GC: 0.829
    ToxCast AUC-ROC Multitask: 0.702 Weave: 0.742
    SIDER AUC-ROC RF: 0.684 GC: 0.638
    ClinTox AUC-ROC Bypass: 0.827 Weave: 0.832

    *Note: In QM7b, DTNN outperformed KRR(CM) on 14 of the 16 individual property tasks, but its aggregate mean MAE across all tasks was skewed due to varying label magnitudes.

    Graph-based models with learnable representations outperform conventional methods on 11 of the 17 benchmark datasets. Conventional methods (such as Random Forest, XGBoost, and Kernel SVM) maintain an advantage on smaller datasets (under 3,000 samples such as BACE, BBBP, and SIDER) and under severe class imbalance (MUV).

  8. Knowl 8 — Physics-Aware Featurizations in Quantum Mechanical Property Prediction

    empirical result

    On quantum mechanical property benchmarks (QM7, QM7b, QM8, and QM9), incorporating 3D Cartesian coordinates, interatomic distances, and nuclear charges is critical for prediction accuracy:

    • On QM7 atomization energy prediction, 3D physics-aware architectures achieve strong results (ANI-1 test MAE 2.86±0.25 kcal/mol2.86 \pm 0.25\text{ kcal/mol}; DTNN test MAE 8.8±3.5 kcal/mol8.8 \pm 3.5\text{ kcal/mol}; KRR with Coulomb Matrix test MAE 10.2±0.3 kcal/mol10.2 \pm 0.3\text{ kcal/mol}), whereas 2D graph convolutions (test MAE 77.9±2.1 kcal/mol77.9 \pm 2.1\text{ kcal/mol}) and ECFP-based neural networks (test MAE 123.7±15.6 kcal/mol123.7 \pm 15.6\text{ kcal/mol}) fail to model the target energy surface.
    • On QM9, distance-aware deep models (DTNN test MAE 2.4±1.12.4 \pm 1.1; MPNN test MAE 3.2±1.53.2 \pm 1.5) substantially outperform standard ECFP-based multitask networks (test MAE 16.0±0.216.0 \pm 0.2).
    • Across all 39 individual quantum mechanical tasks evaluated across QM7, QM7b, QM8, and QM9, physics-aware graph models (ANI-1, DTNN, and MPNN) achieve the top performance on 28 tasks.

  9. Knowl 9 — Data Efficiency and Accuracy of Graph Models on Physical Chemistry Benchmarks

    empirical result

    On physical chemistry datasets (ESOL for water solubility, FreeSolv for hydration free energy, and Lipophilicity), graph neural networks match or exceed the accuracy of ab-initio physical computations:

    • On ESOL, MPNN achieves a test RMSE of 0.58±0.03 logS0.58 \pm 0.03\text{ logS}, approaching experimental measurement error (approximately ±0.5 logS\pm 0.5\text{ logS}) and outperforming gradient boosted trees (XGBoost test RMSE 0.99±0.140.99 \pm 0.14).
    • On FreeSolv, MPNN achieves a test RMSE of 1.15±0.12 kcal/mol1.15 \pm 0.12\text{ kcal/mol}, outperforming alchemical free energy molecular dynamics calculations (RMSE 1.5 kcal/mol1.5\text{ kcal/mol}) and conventional ECFP models (Singletask test RMSE 1.87±0.07 kcal/mol1.87 \pm 0.07\text{ kcal/mol}; XGBoost test RMSE 1.74±0.15 kcal/mol1.74 \pm 0.15\text{ kcal/mol}).
    • Training size scaling experiments on FreeSolv demonstrate that Weave models achieve state-of-the-art free energy prediction (RMSE ≤1.5 kcal/mol\le 1.5\text{ kcal/mol}) with 200 training compounds, requiring approximately one-third of the training sample volume needed by fixed-descriptor feedforward neural networks to attain equivalent accuracy.

  10. Knowl 10 — Bypass Multitask Neural Network Architecture

    model/method

    Bypass Multitask Networks mitigate negative transfer and representation interference in multi-task molecular property prediction by augmenting a shared network trunk with direct task-specific bypass connections.

    For a dataset with ntasksn_{\text{tasks}} target labels, the network comprises ntasks+1n_{\text{tasks}} + 1 independent components:

    1. One shared multitask layer sequence mapping input features to a shared latent representation across all tasks.
    2. ntasksn_{\text{tasks}} independent bypass layer sets mapping raw input features directly to each individual task output without passing through the shared trunk.

    Predictions for task ii integrate activations from both the shared multitask representation and the ii-th bypass pathway. This structure allows shared representations to exploit common chemical features across tasks while preventing task-specific variations from corrupting the joint representation. On the ClinTox benchmark, the bypass multitask network achieves a test ROC-AUC of 0.827±0.0510.827 \pm 0.051, outperforming standard multitask networks (0.778±0.0550.778 \pm 0.055).

Coverage note — Omitted specific software package command invocation listings, Gaussian process optimization iteration configurations, and individual per-model training runtimes on specific GPU hardware nodes as secondary implementation details.

References

  1. 1.Gasteiger, J.; Zupan, J. Angewandte Chemie International Edition 1993, 32, 503–527.
  2. 2.Zupan, J.; Gasteiger, J. Neural networks in chemistry and drug design; John Wiley & Sons, Inc., 1999.
  3. 3.Varnek, A.; Baskin, I. Journal of chemical information and modeling 2012, 52, 1413–1437.
  4. 4.Mitchell, J. B. Wiley Interdisciplinary Reviews: Computational Molecular Science 2014, 4, 468–481.
  5. 5.Devillers, J. Neural networks in QSAR and drug design; Academic Press, 1996.
  6. 6.Schneider, G.; Wrede, P. Progress in biophysics and molecular biology 1998, 70, 175–222.
  7. 7.LeCun, Y.; Bengio, Y.; Hinton, G. Nature 2015, 521, 436–444.
  8. 8.Schmidhuber, J. Neural networks 2015, 61, 85–117.
  9. 9.Ma, J.; Sheridan, R. P.; Liaw, A.; Dahl, G. E.; Svetnik, V. Journal of chemical information and modeling 2015, 55, 263–274.
  10. 10.Ramsundar, B.; Kearnes, S.; Riley, P.; Webster, D.; Konerding, D.; Pande, V. arXiv preprint arXiv:1502.02072 2015,
  11. 11.Unterthiner, T.; Mayr, A.; Klambauer, G.; Steijaert, M.; Wenger, J.; Ceulemans, H.; Hochreiter, S. Deep Learning as an Opportunity in Virtual Screening. Deep Learning and Representation Learning Workshop (NIPS 2014). 2014.
  12. 12.Wallach, I.; Dzamba, M.; Heifets, A. arXiv preprint arXiv:1510.02855 2015,
  13. 13.Delaney, J. S. Journal of Chemical Information and Modeling 2004, 44, 1000–1005.
  14. 14.Lusci, A.; Pollastri, G.; Baldi, P. Journal of chemical information and modeling 2013, 53, 1563–1575.
  15. 15.Mobley, D. L.; Wymer, K. L.; Lim, N. M.; Guthrie, J. P. Journal of Computer-Aided Molecular Design 2014, 28, 135–150.
  16. 16.Mobley, D. L.; Guthrie, J. P. Journal of Computer-Aided Molecular Design 2014, 28, 711–720.
  17. 17.Rupp, M.; Tkatchenko, A.; Müller, K.-R.; Lilienfeld, O. A. v. Physical Review Letters 2012, 108, 058301.
  18. 18.Montavon, G.; Rupp, M.; Gobre, V.; Vazquez-Mayagoitia, A.; Hansen, K.; Tkatchenko, A.; Müller, K.-R.; Lilienfeld, O. A. v. New Journal of Physics 2013, 15, 095003.
  19. 19.Schütt, K. T.; Arbabzadah, F.; Chmiela, S.; Müller, K. R.; Tkatchenko, A. arXiv preprint arXiv:1609.08259 2016,
  20. 20.McGibbon, R. T.; Taube, A. G.; Donchev, A. G.; Siva, K.; Hernández, F.; Hargus, C.; Law, K.-H.; Klepeis, J. L.; Shaw, D. E. The Journal of Chemical Physics 2017, 147, 161725.
  21. 21.Rogers, D.; Hahn, M. Journal of Chemical Information and Modeling 2010, 50, 742–754.
  22. 22.Duvenaud, D.; Maclaurin, D.; Aguilera-Iparraguirre, J.; Gómez-Bombarelli, R.; Hirzel, T.; Aspuru-Guzik, A.; Adams, R. P. arXiv preprint arXiv:1509.09292 2015,
  23. 23.Kearnes, S.; McCloskey, K.; Berndl, M.; Pande, V.; Riley, P. arXiv preprint arXiv:1603.00856 2016,
  24. 24.Miller, G. A. Communications of the ACM 1995, 38, 39–41.
  25. 25.Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; Fei-Fei, L. ImageNet: A Large-Scale Hierarchical Image Database. CVPR09. 2009.
  26. 26.Russakovsky, O.; Deng, J.; Su, H.; Krause, J.; Satheesh, S.; Ma, S.; Huang, Z.; Karpathy, A.; Khosla, A.; Bernstein, M.; Berg, A. C.; Fei-Fei, L. International Journal of Computer Vision (IJCV) 2015, 115, 211–252.
  27. 27.Krizhevsky, A.; Sutskever, I.; Hinton, G. E. ImageNet Classification with Deep Convolutional Neural Networks. NIPS Proceedings. 2012.
  28. 28.Szegedy, C.; Liu, W.; Jia, Y.; Sermanet, P.; Reed, S.; Anguelov, D.; Erhan, D.; Vanhoucke, V.; Rabinovich, A. arXiv preprint arXiv:1409.4842 2014,
  29. 29.He, K.; Zhang, X.; Ren, S.; Sun, J. arXiv preprint arXiv:1512.03385 2015,
  30. 30.DeepChem: Deep-learning models for Drug Discovery and Quantum Chemistry. https://github.com/deepchem/deepchem, Accessed: 2017-09-27.
  31. 31.others,, et al. Journal of Machine Learning Research 2011, 12, 2825–2830.
  32. 32.others,, et al. arXiv preprint arXiv:1603.04467 2016,
  33. 33.Sheridan, R. P. Journal of chemical information and modeling 2013, 53, 783–790.
  34. 34.Bolton, E. E.; Wang, Y.; Thiessen, P. A.; Bryant, S. H. Annual reports in computational chemistry 2008, 4, 217–241.
  35. 35.Wang, T.; Xiao, J.; Suzek, T. O.; Zhang, J.; Wang, J.; Zhou, Z.; Han, L.; Karapetyan, K.; Dracheva, S.; Shoemaker, B. A.; Bolton, E.; Gindulyte, A.; Bryant, S. H. Nucleic Acids Research 2012, 40, D400–D412.
  36. 36.Gražulis, S.; Chateigner, D.; Downs, R. T.; Yokochi, A.; Quirós, M.; Lutterotti, L.; Manakova, E.; Butkus, J.; Moeck, P.; Le Bail, A. Journal of Applied Crystallography 2009, 42, 726–729.
  37. 37.Groom, C. R.; Bruno, I. J.; Lightfoot, M. P.; Ward, S. C. Acta Crystallographica Section B: Structural Science, Crystal Engineering and Materials 2016, 72, 171–179.
  38. 38.Berman, H.; Henrick, K.; Nakamura, H. Nature Structural & Molecular Biology 2003, 10, 980–980.
  39. 39.Quantum Machine. http://quantum-machine.org/datasets/, Accessed: 2017-09-27.
  40. 40.Weininger, D. Journal of chemical information and computer sciences 1988, 28, 31–36.
  41. 41.Blum, L. C.; Reymond, J.-L. Journal of the American Chemical Society 2009, 131, 8732–8733.
  42. 42.Ramakrishnan, R.; Hartmann, M.; Tapavicza, E.; Lilienfeld, O. A. v. The Journal of Chemical Physics 2015, 143, 084111.
  43. 43.Ruddigkeit, L.; Deursen, R. v.; Blum, L. C.; Reymond, J.-L. Journal of Chemical Information and Modeling 2012, 52, 2864–2875.
  44. 44.Ramakrishnan, R.; Dral, P. O.; Rupp, M.; Lilienfeld, O. A. v. Scientific Data 2014, 1, 140022.
  45. 45.Hersey, A. ChEMBL Deposited Data Set - AZ dataset; 2015.
  46. 46.Rohrer, S. G.; Baumann, K. Journal of Chemical Information and Modeling 2009, 49, 169–184.
  47. 47.AIDS Antiviral Screen Data. https://wiki.nci.nih.gov/display/NCIDTPdata/AIDS+Antiviral+Screen+Data, Accessed: 2017-09-27.
  48. 48.Wang, R.; Fang, X.; Lu, Y.; Wang, S. Journal of Medicinal Chemistry 2004, 47, 2977–2980.
  49. 49.Wang, R.; Fang, X.; Lu, Y.; Yang, C.-Y.; Wang, S. Journal of Medicinal Chemistry 2005, 48, 4111–4119.
  50. 50.Liu, Z.; Li, Y.; Han, L.; Li, J.; Liu, J.; Zhao, Z.; Nie, W.; Liu, Y.; Wang, R. Bioinformatics 2014, 31, 405–412.
  51. 51.Subramanian, G.; Ramsundar, B.; Pande, V.; Denny, R. A. Journal of Chemical Information and Modeling 2016, 56, 1936–1949.
  52. 52.Martins, I. F.; Teixeira, A. L.; Pinheiro, L.; Falcao, A. O. Journal of Chemical Information and Modeling 2012, 52, 1686–1697.
  53. 53.Tox21 Challenge. https://tripod.nih.gov/tox21/challenge/, Accessed: 2017-09-27.
  54. 54.Richard, A. M. et al. Chemical Research in Toxicology 2016, 29, 1225–1251.
  55. 55.Kuhn, M.; Letunic, I.; Jensen, L. J.; Bork, P. Nucleic Acids Research 2016, 44, D1075–D41079.
  56. 56.Altae-Tran, H.; Ramsundar, B.; Pappu, A. S.; Pande, V. arXiv preprint arXiv:1611.03199 2016,
  57. 57.Medical Dictionary for Regulatory Activities. http://www.meddra.org/, Accessed: 2017-09-27.
  58. 58.Gayvert, K. M.; Madhukar, N. S.; Elemento, O. Cell Chemical Biology 2016, 23, 1294–1301.
  59. 59.Artemov, A. V.; Putin, E.; Vanhaelen, Q.; Aliper, A.; Ozerov, I. V.; Zhavoronkov, A. bioRxiv 2016, 095653.
  60. 60.Novick, P. A.; Ortiz, O. F.; Poelman, J.; Abdulhay, A. Y.; Pande, V. S. PLOS ONE 2013, 8.
  61. 61.Aggregate Analysis of ClincalTrials.gov (AACT) Database. https://www.ctti-clinicaltrials.org/aact-database, Accessed: 2017-09-27.
  62. 62.Bemis, G. W.; Murcko, M. A. Journal of Medicinal Chemistry 1996, 39, 2887–2893.
  63. 63.Landrum, G. RDKit: Open-Source Cheminformatics Software. https://www.rdkit.org/.
  64. 64.Jain, A. N.; Nicholls, A. Journal of Computer-Aided Molecular Design 2008, 22, 133–139.
  65. 65.Hastie, T.; Tibshirani, R.; Friedman, J. The Elements of Statistical Learning: Data Mining, Inference, and Prediction; Springer, 2009.
  66. 66.Davis, J.; Goadrich, M. The Relationship Between Precision-Recall and ROC Curves. Proceedings of the 23rd International Conference on Machine Learning. 2006.
  67. 67.Gómez-Bombarelli, R.; Duvenaud, D.; Hernández-Lobato, J. M.; Aguilera-Iparraguirre, J.; Hirzel, T. D.; Adams, R. P.; Aspuru-Guzik, A. arXiv preprint arXiv:1610.02415 2016,
  68. 68.Durrant, J. D.; McCammon, J. A. Journal of Chemical Information and Modeling 2011, 51, 2897–2903.
  69. 69.Da, C.; Kireev, D. Journal of chemical information and modeling 2014, 54, 2555–2561.
  70. 70.Behler, J.; Parrinello, M. Physical Review Letters 2007, 98, 146101.
  71. 71.Smith, J. S.; Isayev, O.; Roitberg, A. E. arXiv preprint arXiv:1610.08935 2016,
  72. 72.Breiman, L. Machine learning 2001, 45, 5–32.
  73. 73.Friedman, J. H. Annals of statistics 2001, 1189–1232.
  74. 74.Ramsundar, B.; Liu, B.; Wu, Z.; Verras, A.; Tudor, M.; Sheridan, R. P.; Pande, V. Manuscript in preparation
  75. 75.Swamidass, S. J.; Azencott, C.-A.; Lin, T.-W.; Gramajo, H.; Tsai, S.-C.; Baldi, P. Journal of chemical information and modeling 2009, 49, 756–766.
  76. 76.Gilmer, J.; Schoenholz, S. S.; Riley, P. F.; Vinyals, O.; Dahl, G. E. arXiv preprint arXiv:1704.01212 2017,
  77. 77.others,, et al. The annals of statistics 2000, 28, 337–407.
  78. 78.Cortes, C.; Vapnik, V. Machine learning 1995, 20, 273–297.
  79. 79.Chen, T.; Guestrin, C. arXiv preprint arXiv:1603.02754 2016,
  80. 80.Kearnes, S.; Goldman, B.; Pande, V. arXiv preprint arXiv:1606.08793 2016,
  81. 81.Baskin, I. I.; Palyulin, V. A.; Zefirov, N. S. Journal of chemical information and computer sciences 1997, 37, 715–721.
  82. 82.Kireev, D. B. Journal of chemical information and computer sciences 1995, 35, 175–180.
  83. 83.Vinyals, O.; Bengio, S.; Kudlur, M. arXiv preprint arXiv:1511.06391 2015,
  84. 84.scikit-learn: Machine Learning in Python. http://scikit-learn.org/stable/, Accessed: 2017-10-18.
  85. 85.eXtreme Gradient Boosting. https://github.com/dmlc/xgboost, Accessed: 2017-10-18.

Citation

MLA
Wu, Z., et al. “MoleculeNet: A Benchmark for Molecular Machine Learning”. Chemical Science, vol. 9, no. 2, 2018, pp. 513–30, https://doi.org/10.1039/c7sc02664a.
APA
Wu, Z., Ramsundar, B., Feinberg, E. N., Gomes, J., Geniesse, C., Pappu, A. S., Leswing, K., & Pande, V. (2018). MoleculeNet: a benchmark for molecular machine learning. Chemical Science, 9(2), 513–530. https://doi.org/10.1039/c7sc02664a
Chicago
Wu, Z., B. Ramsundar, E. N. Feinberg, et al. 2018. “MoleculeNet: A Benchmark for Molecular Machine Learning”. Chemical Science 9 (2): 513–30. https://doi.org/10.1039/c7sc02664a.
Harvard
Wu, Z. et al. (2018) “MoleculeNet: a benchmark for molecular machine learning”, Chemical Science, 9(2), pp. 513–530. Available at: https://doi.org/10.1039/c7sc02664a.
Vancouver
1. Wu Z, Ramsundar B, Feinberg EN, Gomes J, Geniesse C, Pappu AS, Leswing K, Pande V (2018) MoleculeNet: a benchmark for molecular machine learning. Chemical Science 9:513–530

BibTeX

@article{Wu_2018, title={MoleculeNet: a benchmark for molecular machine learning}, volume={9}, ISSN={2041-6539}, url={http://dx.doi.org/10.1039/c7sc02664a}, DOI={10.1039/c7sc02664a}, number={2}, journal={Chemical Science}, publisher={Royal Society of Chemistry (RSC)}, author={Wu, Zhenqin and Ramsundar, Bharath and Feinberg, Evan N. and Gomes, Joseph and Geniesse, Caleb and Pappu, Aneesh S. and Leswing, Karl and Pande, Vijay}, year={2018}, pages={513–530} }
Metadata:Crossref

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF