M3L-20M: Multi-Modal Molecule Dataset

• M3L-20M is a large-scale multi-modal molecule dataset that integrates over 20 million molecules with 1D SMILES, 2D graphs, 3D structures, physicochemical properties, and textual descriptions.
• It enhances AI-driven drug discovery by aligning diverse modalities to improve molecule generation, property prediction, and downstream fine-tuning performance.
• Constructed from PubChem, ZINC, and QM9, the dataset employs rigorous quality control, including human expert scoring and prompt-tuned GPT outputs, ensuring reliable multi-modal data.

Searching arXiv for the exact term and likely disambiguations to ground the article. Searching arXiv for "M3L-20M" and related titles. M$^{3}$-20M, referred to in some query contexts as M3L-20M, is a large-scale multi-modal molecule dataset built for AI-driven drug design and discovery. It contains over 20 million molecules and aligns, for each molecule, a 1D SMILES string, a 2D molecular graph, a 3D molecular structure, physicochemical properties, and a textual description. The dataset is positioned as a resource for training, fine-tuning, and prompting large models for molecular science, with intended use extending beyond molecule generation and molecular property prediction to lead optimization, virtual screening, pharmacokinetics modeling, drug-target interaction prediction, retrosynthesis, reaction prediction, name prediction, and molecule captioning (Guo et al., 2024).

1. Nomenclature and referential ambiguity

The string “M3L-20M” is not a stable arXiv title-form and is potentially ambiguous. In the molecular-data context, it refers to M$^{3}$-20M, the large-scale multi-modal molecule corpus introduced for AI-driven drug design and discovery. Closely neighboring names in the literature denote distinct objects: iLab-20M is a controlled turntable image dataset with 704 object instances and about 22 million images for object recognition, and M$^{3}$L denotes a Multi-Modal Multi-Level Transformer for language-based video editing rather than a molecular dataset (Zhao et al., 2016, Fu et al., 2021, Guo et al., 2024).

A second source of confusion is scale-based rather than lexical. ANI-1 is also a 20M-scale chemistry resource, but it is a quantum-chemically labeled dataset containing more than 20 million DFT-calculated conformations for 57,462 small organic molecules and is intended for machine-learning interatomic potentials, especially off-equilibrium potential-energy-surface modeling, not for multimodal alignment across sequence, graph, 3D geometry, properties, and language (Smith et al., 2017). In that sense, M$^{3}$-20M occupies a different position in the dataset landscape: it is organized around cross-modal molecular representation rather than conformer energetics.

2. Scale, composition, and comparative position

The defining property of M$^{3}$-20M is scale. The paper reports over 20 million molecules and states that the dataset is 71 times larger in the number of molecules than the largest existing dataset at the time of writing. It is also described as the largest open-access multi-modal molecular dataset in the paper’s comparison. The scale differential is especially visible when it is set against prior multi-modal resources, which remain in the range of thousands or hundreds of thousands of molecules rather than tens of millions (Guo et al., 2024).

Dataset	Molecules	Position in comparison
QM9	134K	prior multi-modal dataset
PCdes	1.5K	prior multi-modal dataset
PubChemSTM	280K	prior multi-modal dataset
igcdata	220K	prior multi-modal dataset
M$^{3}$-20M	20M	multi-modal integrated molecule dataset

The textual layer is unusually large. The paper reports 360,133 PubChem descriptions versus 20,249,090 M$^{3}$-20M descriptions, indicating that the corpus extends textual coverage far beyond what is directly available from PubChem. The physicochemical layer includes 26 key properties. Among the largest reported counts are Molecular Weight at 19,174,000, Hydrogen Bond Donor Count at 19,173,687, Hydrogen Bond Acceptor Count at 19,173,559, Rotatable Bond Count at 19,173,464, Exact Mass at 19,173,398, and Monoisotopic Mass at 19,173,327. By contrast, several experimental properties are sparse: Physical Description 2031, Color/Form 246, Odor 100, Boiling Point 130, Melting Point 633, Flash Point 35, Solubility 10,813, and Density 170. This unevenness is central to how the dataset should be interpreted: it is massive and broad, but not uniformly dense across all annotation types.

3. Modalities and construction pipeline

M$^{3}$-20M is explicitly built as a multi-modal integrated molecule dataset. Each molecule is associated with a 1D SMILES string, a 2D molecular graph, a 3D molecular structure, physicochemical properties, and a textual description. The paper argues that these modalities are complementary: SMILES provides sequence-level representation, 2D graphs encode bonded topology, 3D structures provide spatial arrangement, properties provide structured descriptors, and text contributes semantic and functional context. The authors also construct seven multi-modal downstream sub-datasets—QM9-MM, MOSES-MM, BACE-MM, BBBP-MM, HIV-MM, ClinTox-MM, and Tox21-MM—to adapt the full corpus to common generation and prediction tasks (Guo et al., 2024).

The primary data sources are PubChem, ZINC, and QM9. The construction workflow is presented as a sequence of steps: collect molecules from PubChem, ZINC, and QM9; use RDKit to derive 2D molecular graphs; use the PubChem API to obtain 3D SDF structures; extract physicochemical properties from PubChem and related databases; construct text from PubChem descriptions or from property templates; generate missing descriptions using GPT-3.5; apply human expert scoring to filter GPT-generated texts; and assemble the final multi-modal dataset and linked task-specific subdatasets.

The per-modality acquisition procedures are concrete. SMILES are collected from the source databases and act as the anchor representation. For 2D graphs, the authors use RDKit’s Chem functions to extract atomic features and chemical bond characteristics. For 3D structures, they batch download SDF files via the PubChem API and use RDKit’s GetAtomPosition function to calculate 3D coordinates of central atoms. For physicochemical properties, the dataset uses 26 properties, of which eighteen are computed and eight are experimental, with the experimental fields drawn from CAMEO Chemicals and the Hazardous Substances Data Bank.

The textual channel is assembled in three ways. First, descriptions are directly extracted from PubChem when they already exist. Second, physicochemical properties are converted into structured natural-language descriptions using a template of the form “property name is a specific value.” Third, GPT-3.5 is used for molecules lacking both textual descriptions and physicochemical property values or for molecules not present in PubChem. The paper notes that the initial prompt was revised because the basic version caused wrong names and incorrect functional-group identification; the revised prompt instructs GPT-3.5 to describe the molecule using only what can be inferred from the SMILES, without inventing names or unsupported claims.

4. Curation, synthetic text, and quality control

The paper presents M$^{3}$-20M as a curated rather than purely aggregated corpus. A key element is the treatment of GPT-generated descriptions. The dataset contains 1,073,845 GPT-3.5-generated descriptions, which the paper states is only 0.934% of the total descriptions. This directly rebuts the common assumption that the textual modality is predominantly synthetic: the generated portion is small relative to the extracted and property-derived text (Guo et al., 2024).

Quality control is organized around human expert scoring. Six chemistry students from top Chinese universities score generated descriptions on four dimensions: Accuracy with maximum 5, Effectiveness with maximum 2, Comprehensiveness with maximum 2, and Simplicity with maximum 1, for a total score out of 10. Descriptions scoring above 5 are kept, and those below 5 are regenerated. Any description with major scientific errors is also regenerated. This procedure is paired with prompt revision intended to suppress unsupported molecular naming and functional-group hallucination.

The paper also describes a maintenance model rather than a one-time release. A six-person maintenance team, consisting of three PhD students and three master’s students, is stated to update and correct the dataset twice a week. A plausible implication is that the authors treat the resource as a living corpus whose text and metadata may continue to be refined, rather than as a static benchmark snapshot.

5. Benchmarking tasks and reported performance

The empirical validation focuses on two major tasks: molecule generation and molecular property prediction. For generation, the paper evaluates both a generic prompting setting and a benchmark setting based on MOSES-MM. In the generic setting, GLM4, GPT-3.5, and GPT-4 receive 10 in-context examples and are asked to generate 100 molecules; the reported metrics are Validity, Uniqueness, and Novelty. Multi-modal examples consistently outperform single-modal SMILES-only examples. For GLM4, Validity improves from 72.73% to 85.37%, Uniqueness from 68.17% to 82.25%, and Novelty from 97.90% to 98.10%. For GPT-3.5, the corresponding changes are 76.01% to 84.8%, 46.46% to 93.86%, and 85.72% to 96.51%. For GPT-4, they are 92.3% to 97.99%, 58.64% to 70.17%, and 90.66% to 98.95%. On MOSES-MM, the paper reports gains such as GPT-3.5 SNN/Test 0.53 to 0.56, Scaf/Test 6.90e-4 to 0.01, Filters 0.8 to 1.0, logP Wasserstein distance 1.63 to 0.99, SA 0.63 to 0.57, and QED 0.15 to 0.12. In a cross-dataset comparison using GPT-3.5, M$^{3}$-20M reaches 84.8% Validity, 93.86% Uniqueness, and 96.51% Novelty, compared with QM9 at 75.17 / 64.13 / 89.43, PCdes at 84.31 / 82.64 / 93.253, PubChemSTM at 82.24 / 80.89 / 95.91, and igcdata at 78.34 / 61.21 / 95.07 (Guo et al., 2024).

For molecular property prediction, the paper separates regression and classification. Regression is evaluated on QM9-MM with targets dipole moment, $^{3}$0, and isotropic polarizability, $^{3}$1, using the MAE definition

$^{3}$2

The paper notes that the notation appears reversed relative to the usual convention, but the intended meaning is the average absolute difference between prediction and ground truth. Three input settings are compared: SMILES only, SMILES + 3D, and SMILES + 3D + text. The reported trend is that adding modalities usually improves MAE, especially when text is added. For GPT-3.5, $^{3}$3 changes from 1.78 to 1.82 to 1.56 and $^{3}$4 from 39.50 to 35.37 to 34.74. For GLM-4, $^{3}$5 changes from 1.27 to 1.13 to 1.01 and $^{3}$6 from 12.82 to 14.53 to 12.05. For GPT-4, $^{3}$7 changes from 1.17 to 1.28 to 1.27 and $^{3}$8 from 13.71 to 28.9 to 9.50.

Classification is evaluated on BACE-MM, BBBP-MM, ClinTox-MM, HIV-MM, and Tox21-MM with ACC Mean, ACC Variance, and ACC Standard Deviation. The reported ACC Mean values show that multi-modal data often improves accuracy: BACE-MM 0.5680 to 0.5780, BBBP-MM 0.2280 to 0.2720, ClinTox-MM 0.9260 to 0.9280, and HIV-MM 0.9740 to 0.9680, the latter being a slight decline. On Tox21-MM, examples of gains include NR-AR 0.9620 to 0.9680, NR-AhR 0.8740 to 0.9020, NR-Aromatase 0.9620 to 0.9720, and NR-ER 0.9140 to 0.9240, while SR-HSE standard deviation decreases from 0.0295 to 0.0114. The paper also reports fine-tuning results for Llama3-8b: BBBP 0.959 to 0.959, BACE 0.844 to 0.887, Tox21 0.817 to 0.849, HIV 0.780 to 0.803, and ClinTox 0.985 to 0.998. Taken together, these results support the paper’s claim that the dataset benefits both prompting of closed-source models and adaptation of open-source models.

6. Research role, analytic signals, and limitations

The paper frames M$^{3}$9-20M as a broad resource for molecular representation learning rather than a task-specific benchmark only. Its supporting analyses include a word-frequency analysis of generated descriptions, a correlation matrix among physicochemical properties, and a PCA visualization of sampled molecules. Molecular Weight and XLogP3 are singled out as particularly informative because they correlate strongly with other properties and are used as the first two PCA components for visualization. This suggests that the dataset is intended not only as a training corpus but also as an object of descriptive analysis in its own right (Guo et al., 2024).

Several caveats are explicit. The corpus depends heavily on PubChem, with additional contributions from ZINC and QM9. Only a small fraction of descriptions are GPT-generated, but those descriptions still require prompt engineering and human review. Experimental property annotations are sparse, so not all molecules have all 26 properties. The benchmark evaluations are mostly on prompting and small-scale fine-tuning rather than exhaustive training of large domain-specific foundation models. The paper also positions the dataset as openly accessible for non-commercial use. These points matter because they delimit what kind of resource M$^{3}$0-20M is: massive in breadth and modality alignment, but heterogeneous in annotation density and still centered on foundational representation and prompting workflows.

In the contemporary arXiv ecosystem, M$^{3}$1-20M is therefore best understood as a multimodal molecular corpus at unprecedented open scale. Its central contribution is not merely a larger molecule count, but the explicit alignment of sequence, graph, 3D geometry, properties, and language within one dataset. The reported experimental behavior indicates that this alignment can improve molecule generation, molecular property prediction, and downstream fine-tuning, while the dataset’s construction details and limitations indicate that its main research value lies in large-scale cross-modal supervision for AI-driven drug design and discovery.