Hassan2007/EcoAgro3D

AgroBench-3D

A transparent, data-centric molecular resource for safety-aware agrochemical machine learning.

AgroBench-3D consolidates curated organic molecular structures, calculated physicochemical descriptors, non-target safety proxy targets, and agricultural-role indicators into one fixed, model-ready schema. The current release contains 11,129 unique canonical organic structures and 26 fields, with no missing cells and no duplicate canonical SMILES. It is intended to reduce the repeated data-engineering effort that otherwise separates molecular graph modeling, descriptor-based QSAR, and safety-aware agrochemical-method development.

Important release boundary. The continuous toxicity-related fields in this release are descriptor-derived proxy scores, not traceable experimental LD50/LC50 measurements. The repository includes a 3D conformer archive (conformers_3d_master.sdf), but users should verify molecule-level CSV--SDF alignment, conformer coverage, and optimization status before model training. This version is suitable for data-pipeline development, reproducible curation studies, 2D/3D methodological prototyping, and conformer-aware representation learning; it must not be used for regulatory, ecological, or experimental-toxicity claims.

Dataset Details

Dataset Description

Property	Value
Repository ID	Hassan2007/EcoAgro3D
Display name	AgroBench-3D
Task families	Tabular regression; tabular classification; molecular graph learning; multi-task learning
Records	11,129 unique organic molecules
Schema	26 fields
Primary representation	Canonical SMILES, calculated descriptors, labels, and weak role annotations
3D status	conformers_3d_master.sdf is included; the documented workflow uses ETKDGv3 and MMFF94s
Curator and contact	Hassan Ahmed Hassan Zaki — hassanahmed07.e9@gmail.com
ORCID	0009-0005-0306-0898
License	MIT for curator-authored release materials; see the licensing note below for upstream-source obligations

AgroBench-3D is a data resource for a practical gap in agrochemical machine learning: relevant molecular structures, calculated properties, ecological-safety concepts, and agricultural-use labels are often fragmented across incompatible sources. This release establishes a stable tabular contract so users can load the same molecular identity, descriptors, proxy endpoints, and labels without independently reconstructing the schema.

Included Files

File	Status	Description
agrochemical_master.csv	Included (1.55 MB)	Master table containing 11,129 rows and 26 columns
EXP_data_source_code.ipynb	Included (29.3 kB)	Curation, descriptor, proxy-target, and 3D-generation workflow
conformers_3d_master.sdf	Included (38.9 MB)	ETKDGv3/MMFF94s conformer archive for coordinate-aware workflows

The CSV, notebook, and SDF are published at the repository root. A future revision should add a machine-readable manifest that maps every SDF record unambiguously to a CSV row through SMILES and/or a persistent molecule identifier, together with conformer-generation status and energy information.

Data Sources

The curation workflow combines a verified local agrochemical master set with structural records drawn from EPA ToxCast, Tox21, ClinTox, SIDER, BBBP, BACE, and HIV source tables. The non-agrochemical sources expand structural diversity; their inclusion must not be interpreted as evidence of agricultural efficacy or measured ecological safety for every included compound. Source snapshots, source-specific licenses, acquisition dates, and record-level assay provenance are not bundled with the present release and are required for a future experimentally grounded benchmark.

Intended Uses

Direct Use

Appropriate uses of this release include:

• Building and validating CSV, RDKit, PyTorch Geometric, DGL, or DeepChem data loaders for agrochemical molecules.
• Converting canonical SMILES to 2D molecular graphs for representation-learning and multi-task-learning prototypes.
• Reproducing descriptor distributions, quality-control checks, and the published proxy-target construction.
• Developing scaffold-splitting, leakage-detection, task-masking, calibration, and class-imbalance workflows.
• Prototyping ranking and generative-model interfaces using proxy objectives, provided all results are described as methodological rather than toxicological findings.
• Loading the supplied SDF for 3D graph neural networks, conformer-aware representation learning, and coordinate-conditioned generative-model engineering after verifying CSV--SDF record alignment.
• Auditing, extending, and replacing the current proxy targets with experimentally traceable endpoints.

Recommended Evaluation Practice

The CSV does not contain predefined train/validation/test splits. For predictive experiments, create and publish fixed scaffold-disjoint splits before comparing models. When source metadata are available, supplement them with source-disjoint or time-disjoint external evaluation. Use balanced accuracy, AUROC, AUPRC, calibration, and confidence intervals as appropriate; raw accuracy is insufficient for the strongly imbalanced mammalian proxy-label task.

Out-of-Scope Use

This release must not be used to:

• Make regulatory decisions, pesticide-registration decisions, environmental-risk assessments, or human-health claims.
• Treat bee_pLD50, aquatic_pLC50, or mammal_pLD50 as experimentally measured molar toxicity values.
• Represent *_safe_label fields as validated safety determinations or thresholded assay outcomes.
• Claim that a molecule is safe for honeybees, aquatic organisms, mammals, food exposure, or any non-target species.
• Present a single ETKDG/MMFF conformer as a complete conformational ensemble, a quantum-mechanically validated geometry, or a binding pose.
• Treat the sparse agricultural-role fields as complete ground-truth use classifications.

Dataset Structure

Unit of Observation

Each row represents one deduplicated, standardized organic molecular structure. Deduplication is performed on canonical SMILES after largest-fragment selection.

Field Reference

Group	Fields	Description
Identity and structure	name, CID, CAS, SMILES	Compound identifiers and canonical molecular structure. CID and CAS should be independently checked against source registries for identity-critical applications.
Structural and physicochemical descriptors	num_atoms, MW, LogP, TPSA, HBD, HBA, RotBonds	RDKit-derived heavy-atom count, molecular weight, calculated lipophilicity, topological polar surface area, hydrogen-bond donor/acceptor counts, and rotatable-bond count.
Continuous proxy scores	bee_pLD50, aquatic_pLC50, mammal_pLD50	Descriptor-derived continuous scores. Despite their names, these are not assay-derived pLD50/pLC50 measurements in the current release.
Binary proxy labels	bee_safe_label, aquatic_safe_label, human_safe_label	Binary values obtained by thresholding the corresponding proxy scores. In the supplied code, 1 denotes a score above its cutoff; field names should not be interpreted as validated safety annotations.
Agricultural-role indicators	herbicide, fungicide, insecticide, acaricide, nematicide, rodenticide, plant_growth_regulator, bactericide, molluscicide	Non-exclusive weak labels assigned using keywords and limited SMARTS patterns.

Descriptive Statistics

Property	Mean	Median	SD	Minimum	Maximum
Molecular weight (Da)	304.26	284.34	148.37	60.01	796.02
LogP	2.37	2.49	2.25	-13.20	14.57
TPSA (Ų)	65.04	57.36	47.54	0.00	399.71

Binary-Label Composition

Field	Label 0	Label 1	Interpretation for this release
bee_safe_label	5,632 (50.607%)	5,497 (49.393%)	Near-balanced proxy task
aquatic_safe_label	4,908 (44.101%)	6,221 (55.899%)	Near-balanced proxy task
human_safe_label	10,639 (95.597%)	490 (4.403%)	Strongly imbalanced proxy task

Agricultural-Role Coverage

The current weak-label coverage is sparse: herbicide has 40 positive records, fungicide has 94, and insecticide has 49. The other six role fields have no positive records, and 10,946 molecules have no assigned agricultural role. Use task masking, weak-supervision methods, or a curated subset; do not report all nine fields as fully supervised classification tasks.

Dataset Creation

Curation Rationale

The resource was created to make agrochemical molecular data easier to use as shared infrastructure. The design goal is a fixed schema that joins structure, molecular properties, non-target proxy objectives, and agricultural-role indicators, reducing the effort required to begin reproducible cheminformatics and machine-learning studies.

Data Collection and Processing

The supplied implementation performs the following operations:

1. Parses input SMILES with RDKit.
2. Splits disconnected fragments and retains the largest heavy-atom organic fragment.
3. Removes missing or unparsable strings, strings containing a vertical-bar delimiter, compounds outside 4--65 heavy atoms, and elements outside C, H, N, O, S, P, F, Cl, Br, I, B, and Si.
4. Restricts molecular weight to 60--800 Da.
5. Serializes the retained structures as canonical SMILES and removes duplicate canonical structures.
6. Calculates molecular weight, LogP, TPSA, HBD, HBA, and rotatable-bond counts.
7. Assigns agricultural-role indicators using name keywords and limited SMARTS patterns.
8. Creates the current continuous and binary targets from explicit descriptor-based formulas.
9. Specifies a one-conformer ETKDGv3 embedding workflow with random seed 42 and MMFF94s minimization for up to 150 iterations.

The documented specification proposes a Tice-inspired LogP interval of -2 to 8; the supplied code calculates LogP but does not enforce this filter. The observed range is -13.20 to 14.57. Users reproducing a strict chemical-space release should implement the filter explicitly and publish a new, versioned artifact.

Current Proxy-Target Construction

Let M denote molecular weight, L LogP, A TPSA, R rotatable bonds, and D hydrogen-bond donors. The code computes the intermediate score:

$$h=0.45L+\frac{M}{350}-\frac{A}{120}\mathrm{.}h\; =\; 0.45L\; +\; \backslash frac\{M\}\{350\}\; -\; \backslash frac\{A\}\{120\}.$$

It then creates the clipped continuous proxy scores:

$$s_{\mathrm{b}\mathrm{e}\mathrm{e}}=\mathrm{clip}(3.5+h+0.5\mathbb{I}(R>4)-0.2,0.5,9.5),s\_\{\backslash mathrm\{bee\}\}\; =\; \backslash operatorname\{clip\}\backslash left(3.5\; +\; h\; +\; 0.5\backslash mathbb\{I\}(R\; >\; 4)\; -\; 0.2,\backslash ;\; 0.5,\backslash ;\; 9.5\backslash right),$$

$$s_{\mathrm{a}\mathrm{q}\mathrm{u}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{c}}=\mathrm{clip}(3.0+0.6L+\frac{M}{400},0.5,10.0),s\_\{\backslash mathrm\{aquatic\}\}\; =\; \backslash operatorname\{clip\}\backslash left(3.0\; +\; 0.6L\; +\; \backslash frac\{M\}\{400\},\backslash ;\; 0.5,\backslash ;\; 10.0\backslash right),$$

$$s_{\mathrm{m}\mathrm{a}\mathrm{m}\mathrm{m}\mathrm{a}\mathrm{l}}=\mathrm{clip}(2.5+0.3L+0.2D,0.5,8.5)\mathrm{.}s\_\{\backslash mathrm\{mammal\}\}\; =\; \backslash operatorname\{clip\}\backslash left(2.5\; +\; 0.3L\; +\; 0.2D,\backslash ;\; 0.5,\backslash ;\; 8.5\backslash right).$$

The binary fields are 1 when the respective score exceeds 5.0, 5.0, and 4.5. This explicit construction is useful for reproducibility and software testing, but it creates a direct leakage pathway: a model supplied with the same descriptors can recover the generating relationships. The current proxy fields are therefore not an independent predictive-toxicology benchmark.

3D-Conformer Procedure and Release Status

The repository includes conformers_3d_master.sdf (38.9 MB). The code specifies one ETKDGv3 conformer per standardized molecule, with explicit hydrogens, a fixed random seed of 42, and MMFF94s minimization for up to 150 iterations. The archive supports coordinate-aware engineering experiments. Nevertheless, a single force-field-minimized conformer is not a conformational ensemble, a quantum-mechanically validated geometry, or a protein-bound pose. Because a separate conformer manifest is not yet included, users should verify CSV--SDF identity mapping, conformer coverage, embedding failures, and optimization status before training or reporting 3D-model results.

Data Quality Checks

The released master table was audited for:

• Completeness: 0 null cells across 11,129 rows and 26 columns.
• Canonical-structure uniqueness: 0 duplicate SMILES entries after curation.
• Structural bounds: heavy-atom count from 4 to 57 and molecular weight from 60.01 to 796.02 Da.
• Label prevalence: reported above for every binary proxy task.
• Target dependence: the continuous proxy scores are strongly correlated because they share calculated descriptor inputs; this is expected and documented, not evidence of independent biological mechanisms.

Bias, Risks, and Limitations

Scientific and Technical Limitations

• No assay provenance for current toxicity fields. The score names resemble standard toxicological endpoints, but the current values are generated from descriptors and are not measured pLD50/pLC50 observations.
• Descriptor leakage. LogP, molecular weight, TPSA, HBD, and rotatable bonds directly enter the proxy formulas. Reported performance against these labels can overstate a model's ability to learn toxicological mechanisms.
• Sparse role annotations. Only three of nine agricultural-role fields currently contain positive examples; these are weak labels rather than a complete functional ontology.
• Single-conformer and mapping limitations. A 3D SDF is included, but a single ETKDG/MMFF conformer does not represent conformational diversity or a bound pose; a machine-readable CSV--SDF manifest and conformer quality report are still needed.
• Unenforced advertised LogP filter. Although the intended interval is -2 to 8, the source code does not apply it.
• Chemical identity scope. Zero duplicate canonical SMILES does not verify registry identity, stereochemical completeness, commercial availability, pesticide registration, biological activity, or environmental fate.

Responsible-Use Recommendations

Use this version as a transparent engineering resource. Clearly state that outcomes concern proxy-label prediction, data curation, conformer-aware method development, or other methodological evaluation. For ecological, toxicological, or decision-making use, first construct a new release with assay-level provenance, species and life-stage context, exposure route, endpoint definition, unit conversion, censoring status, quality flags, expert-curated agricultural roles, an audited CSV--SDF manifest and conformer quality report, and external validation.

Licensing and Redistribution Note

The mit metadata is intended for curator-authored documentation, code, and release materials. Before publishing a redistributed aggregate under this license, verify that each upstream data source permits redistribution and that its attribution, database-right, and license conditions are satisfied. If that audit identifies restrictions, change the repository license metadata and provide a source-specific licensing manifest before release.

Citation

If you use AgroBench-3D, cite the dataset release. The citation below uses @misc rather than @article because the Hugging Face release is a dataset artifact; replace the version field and add a DOI if a versioned archival release is minted.

BibTeX

@misc{zaki2026agrobench3d,
  author       = {Zaki, Hassan Ahmed Hassan},
  title        = {{AgroBench-3D}: A Data-Centric Resource for Safety-Aware Agrochemical Machine Learning},
  year         = {2026},
  month        = aug,
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Hassan2007/EcoAgro3D}},
  note         = {Hugging Face dataset repository}
}

APA 7th edition

Zaki, H. A. H. (2026). AgroBench-3D: A data-centric resource for safety-aware agrochemical machine learning [Data set]. Hugging Face. https://huggingface.co/datasets/Hassan2007/EcoAgro3D

If a manuscript is formally published, cite both the dataset release and the paper. Do not replace the dataset citation with an article citation unless the referenced article has actually been published.

Dataset Card Authors

Hassan Ahmed Hassan Zaki — ORCID 0009-0005-0306-0898

Dataset Card Contact

For questions, corrections, provenance contributions, or release updates, contact hassanahmed07.e9@gmail.com.