ZINC-Curated: Chemical Spaces & Fragment Models

• ZINC-curated resources are chemically constrained subsets derived from the extensive ZINC database using formal inclusion criteria for enhanced transferability and modularity.
• Examples like AGZ7, the MYH9 library, and FragAtlas-62M illustrate diverse curation pipelines, from graph-based fragmentation to descriptor harmonization and language model training.
• Distinct curation protocols establish admissible chemical subspaces, enabling robust analyses in quantum machine learning, network analysis, and fragment-based generative modeling.

Searching arXiv for papers related to ZINC-curated datasets and ZINC-based chemical-space resources. “ZINC-Curated” denotes datasets, libraries, and analytical workflows derived from the ZINC chemical database and then systematically filtered, standardized, fragmented, or re-embedded for downstream computation. In the cited literature, ZINC is described as a public database of commercially accessible compounds curated for virtual screening and drug discovery, and ZINC-derived curation is used in at least three distinct ways: to build transferable quantum-chemical fragment dictionaries such as AGZ7, to construct descriptor-complete screening libraries for network analysis, and to train large fragment LLMs on the full ZINC-22 fragment subset (Huang et al., 2020, Ali et al., 6 Mar 2026, Ho et al., 23 Sep 2025). Across these uses, curation is not a cosmetic preprocessing step but a methodological choice that fixes the admissible chemical space, the representation, and the downstream claims about transferability, modularity, and generative validity.

1. Concept and scope

In this literature, ZINC-curated resources are not monolithic. They range from fragment dictionaries with explicit geometries and quantum properties to task-specific repurposing libraries and foundation-model corpora. What unifies them is that each begins from a much larger ZINC parent space and then applies formal inclusion criteria before analysis or model training (Huang et al., 2020, Ali et al., 6 Mar 2026, Ho et al., 23 Sep 2025).

A central example is AGZ7, a combined GDB17/ZINC AMON dictionary in which the ZINC query universe is given as 816,777,192 molecules, but the released resource contains about 139,643 AMONs after graph-based fragmentation and a 7-heavy-atom cutoff, $N_I \le 7$ (Huang et al., 2020). A second example is a MYH9-oriented ZINC library that starts from 6,004 molecules and retains 5,000 structurally valid, descriptor-complete compounds as a common node set for network analysis (Ali et al., 6 Mar 2026). A third is the ZINC-22 fragment subset used for FragAtlas-62M, where the curated corpus contains 62,015,589 molecules after preprocessing and is defined by heavy-atom bins H08 to H16 and LogP bins M500 to P240 (Ho et al., 23 Sep 2025).

These examples show that ZINC-curated collections are best understood as chemically constrained subspaces rather than generic subsets. A plausible implication is that conclusions drawn from such resources depend strongly on the curation rule itself: fragmentation, descriptor completeness, or fragment-region selection each induces a different notion of “coverage.”

2. Curation pipelines and admissibility rules

The curation logic differs substantially across resources.

For AGZ7, only the SMILES strings of the parent molecules are used for fragmentation; the resulting fragments are hydrogen-saturated so that dangling valences are capped, and only fragments with no more than 7 non-hydrogen atoms are retained (Huang et al., 2020). The dataset records compositional and constitutional isomers separately. Charged queries were discarded unless the formal charges occurred as paired and opposite charges in the same molecule, and neutral molecules with separated formal charges were also excluded (Huang et al., 2020). Conformer generation is graph-based, and for each graph only one conformer is kept: the approximate global minimum conformer, identified by force-field relaxation with RDKit ETKDG embedding followed by MMFF94 minimization, or UFF when MMFF94 parameters are missing; OpenBabel is used for difficult cases (Huang et al., 2020).

For the MYH9-oriented ZINC library, preprocessing is descriptor-centric. The cleaning step includes normalizing descriptor headers by synonym matching, recomputing missing descriptors from valid SMILES when possible using RDKit, and removing rows with invalid or unparsable SMILES or missing essential descriptors (Ali et al., 6 Mar 2026). The purpose is explicit: every compound must be represented consistently across all similarity layers, avoiding missing-node artifacts when comparing networks (Ali et al., 6 Mar 2026).

For FragAtlas-62M, preprocessing is optimized for generative modeling rather than descriptor harmonization. Exact string duplicates are removed; SMILES are not canonicalized for training, in order to preserve representational diversity; canonicalization is applied only for evaluation; and generated outputs are post-processed with RDKit sanitization, with invalid molecules discarded at evaluation time (Ho et al., 23 Sep 2025).

The methodological contrast is significant. AGZ7 curates for local chemical transferability and quantum consistency, the MYH9 library for cross-descriptor comparability, and FragAtlas-62M for large-scale fragment language modeling. This suggests that “curation” in the ZINC context is inseparable from the intended computational task.

3. Representations, descriptors, and property layers

ZINC-curated resources rely on multiple representational schemes rather than a single canonical molecular encoding.

In the MYH9 study, structural similarity is defined on Morgan fingerprints, specifically circular fingerprints with radius 2 and 2,048 bits, using the Tanimoto coefficient

$$S_{ij}^{\mathrm{Tan}}=\frac{|A_i\cap A_j|}{|A_i\cup A_j|}= \frac{|A_i\cap A_j|}{|A_i|+|A_j|-|A_i\cap A_j|},$$

while descriptor-specific similarity is built from z-score-normalized xLogP, HBD, HBA, molecular weight, and rotatable bond count through

$S_{ij}=\frac{1}{1+|z_i-z_j|}.$

The authors explicitly frame these descriptors as complementary: xLogP and MW produce gradient-like organization, HBD and HBA reflect discrete functional-group counts, and ROTB separates rigid from flexible molecules (Ali et al., 6 Mar 2026).

AGZ7, by contrast, is a geometry-rich quantum dataset. It covers the elements $$\{ \mathrm{H, B, C, N, O, F, Si, P, S, Cl, Br, Sn, I} \}$$ and stores optimized geometries with total energy and its decomposition, Mulliken atomic charges, dipole moment vectors, quadrupole tensors, electronic spatial extent, eigenvalues of all occupied orbitals, LUMO, gap, isotropic polarizability, harmonic frequencies, reduced masses, force constants, IR intensity, normal coordinates, rotational constants, zero-point energy, internal energy, enthalpy, entropy, free energy, and heat capacity at $298.15\ \mathrm{K}$ and 1 atm (Huang et al., 2020). The level of theory is B3LYP/cc-pVTZ, with cc-pVTZ-PP used for Sn and I, and optimized molecular geometries are stored in extended XYZ format (Huang et al., 2020).

FragAtlas-62M uses character-level SMILES as its primary representation. The model is GPT-2 based, with 6 transformer layers, 12 attention heads, 768-dimensional embeddings, a 128-token context window, 42.7 million parameters, and a character-level vocabulary of size 42 (Ho et al., 23 Sep 2025). The training objective is the standard autoregressive factorization

$p(x)=\prod_{t=1}^{T} p(x_t\mid x_{<t}),$

optimized by sequence negative log-likelihood (Ho et al., 23 Sep 2025).

Taken together, these resources illustrate three complementary views of curated ZINC chemistry: graph fragments with quantum observables, descriptor-complete network nodes, and tokenized fragments for generative modeling.

4. Major curated resources derived from ZINC

The main ZINC-curated resources described in the cited papers can be summarized as follows.

Resource	Curation basis	Reported scale
AGZ7	ZINC/GDB17 fragmentation into hydrogen-saturated AMONs with $N_I \le 7$	about 139,643 AMONs
MYH9-oriented ZINC library	preprocessing to structurally valid, descriptor-complete compounds	5,000 compounds from 6,004
ZINC-22 fragment subset / FragAtlas-62M corpus	ZINC fragment-region filtering plus generative-model preprocessing	62,015,589 molecules

AGZ7 is the most explicitly chemistry-centric of the three. Of its about 139,643 AMONs, 98,218 are unique to ZINC, 13,048 are unique to GDB17, and 28,377 are shared between the two sources (Huang et al., 2020). The paper emphasizes that ZINC contributes important chemistries beyond GDB17, including environments such as $>\mathrm{B}^{-}$, $>\mathrm{Si}<$, $>\mathrm{P}^{-}$, and $$S_{ij}^{\mathrm{Tan}}=\frac{|A_i\cap A_j|}{|A_i\cup A_j|}= \frac{|A_i\cap A_j|}{|A_i|+|A_j|-|A_i\cap A_j|},$$0, which makes the ZINC-derived part particularly valuable for bioorganic and medicinal-chemistry-relevant space (Huang et al., 2020).

The MYH9-oriented library is smaller but analytically deeper at the network level. After preprocessing to 5,000 compounds, the study constructs six $$S_{ij}^{\mathrm{Tan}}=\frac{|A_i\cap A_j|}{|A_i\cup A_j|}= \frac{|A_i\cap A_j|}{|A_i|+|A_j|-|A_i\cap A_j|},$$1-nearest-neighbor graphs with $$S_{ij}^{\mathrm{Tan}}=\frac{|A_i\cap A_j|}{|A_i\cup A_j|}= \frac{|A_i\cap A_j|}{|A_i|+|A_j|-|A_i\cap A_j|},$$2: one structural network and five descriptor networks. The structural SMILES/Morgan network has 5,000 nodes and 17,112 weighted edges; the descriptor networks have 20,466 edges for xLogP, 26,961 for HBD, 22,896 for HBA, 15,865 for MW, and 24,760 for ROTB (Ali et al., 6 Mar 2026).

FragAtlas-62M scales curated ZINC fragments to foundation-model size. The source is the April 2025 ZINC-22 fragment subset; after preprocessing, the corpus contains 62,015,589 molecules, roughly 2 billion SMILES tokens, with average SMILES length $$S_{ij}^{\mathrm{Tan}}=\frac{|A_i\cap A_j|}{|A_i\cup A_j|}= \frac{|A_i\cap A_j|}{|A_i|+|A_j|-|A_i\cap A_j|},$$3 characters, range 8–86, and interquartile range 27–35 (Ho et al., 23 Sep 2025).

5. Analytical frameworks built on curated ZINC spaces

A defining feature of ZINC-curated work is that the curated library is usually only the starting point; the principal scientific claims arise from the analysis performed on top of it.

For AGZ7, the validation target is transferability of local chemistry. The paper compares 2-body distance distributions and 3-body angle distributions between AGZ7 and a random test set of about 100k larger GDB17 molecules with $$S_{ij}^{\mathrm{Tan}}=\frac{|A_i\cap A_j|}{|A_i\cup A_j|}= \frac{|A_i\cap A_j|}{|A_i|+|A_j|-|A_i\cap A_j|},$$4, and reports that the domains overlap essentially perfectly (Huang et al., 2020). This is the basis for the claim that local chemical environments of larger targets are covered by the fragment dictionary. The same paper then uses AMON-based kernel ridge regression on seven rigid GDB17 molecules and reports that very low absolute prediction errors are reached rapidly as the size and number of AMONs increase (Huang et al., 2020).

For the MYH9 library, the emphasis is mesoscopic organization in chemical space. Community detection is performed with the Louvain–Leiden approach. The number of communities varies by descriptor—98 for the SMILES network, 149 for xLogP, 5 for HBD, 11 for HBA, 154 for MW, and 9 for ROTB—and the reported modularity values range from 0.638 to 0.985, while the abstract summarizes $$S_{ij}^{\mathrm{Tan}}=\frac{|A_i\cap A_j|}{|A_i\cup A_j|}= \frac{|A_i\cap A_j|}{|A_i|+|A_j|-|A_i\cap A_j|},$$5–0.99 for the strongest cases (Ali et al., 6 Mar 2026). Significance is assessed against 50 degree-preserving randomized null models, and the empirical $$S_{ij}^{\mathrm{Tan}}=\frac{|A_i\cap A_j|}{|A_i\cup A_j|}= \frac{|A_i\cap A_j|}{|A_i|+|A_j|-|A_i\cap A_j|},$$6-value is effectively 0, reported as $$S_{ij}^{\mathrm{Tan}}=\frac{|A_i\cap A_j|}{|A_i\cup A_j|}= \frac{|A_i\cap A_j|}{|A_i|+|A_j|-|A_i\cap A_j|},$$7 (Ali et al., 6 Mar 2026).

The same study introduces a co-clustering matrix over 12,497,500 molecular pairs,

$$S_{ij}^{\mathrm{Tan}}=\frac{|A_i\cap A_j|}{|A_i\cup A_j|}= \frac{|A_i\cap A_j|}{|A_i|+|A_j|-|A_i\cap A_j|},$$8

to quantify cross-descriptor robustness (Ali et al., 6 Mar 2026). The distribution is heavily skewed toward low agreement: 9,332,928 pairs co-cluster in 0 descriptors, 6,608,897 in 1 descriptor, and 1,803,216 in 2 descriptors, while only 4,323 pairs agree across all 6 descriptors (Ali et al., 6 Mar 2026). The paper interprets this as evidence that descriptors are highly complementary, yet still contain a sparse high-consensus core.

FragAtlas-62M shifts the analytical focus from coverage and modularity to generative quality. The model generates 62,015,589 SMILES, of which 61,951,924 are valid, corresponding to 99.90% validity (Ho et al., 23 Sep 2025). After deduplication and canonicalization, there are 42,597,827 unique canonical valid molecules; 33,210,363 are rediscovered from ZINC and 9,387,465 are novel (Ho et al., 23 Sep 2025). The reported ZINC coverage is 53.55%, and novel molecules constitute 22.04% of unique canonical generated molecules (Ho et al., 23 Sep 2025). Distributional matching is assessed across 12 descriptors and three fingerprint methods, with all reported effect sizes below 0.4 and many descriptor-wise Cohen’s $$S_{ij}^{\mathrm{Tan}}=\frac{|A_i\cap A_j|}{|A_i\cup A_j|}= \frac{|A_i\cap A_j|}{|A_i|+|A_j|-|A_i\cap A_j|},$$9 values near zero, such as 0.001 for QED and 0.009 for LogP (Ho et al., 23 Sep 2025).

6. Uses, interpretations, and limitations

The practical value of ZINC-curated resources depends on the level at which they intervene.

AGZ7 is intended as a benchmark and training resource for next-generation quantum machine learning, especially where local chemical representations and transferability across chemical subspaces are central (Huang et al., 2020). Because the data are consistently optimized and computed at a common DFT level, the dataset is positioned as attractive for reproducible benchmarking and for fragment-based QML models that can be queried “on the fly” for new targets (Huang et al., 2020).

The MYH9 library is not presented as a direct drug predictor but as an ordered search space for downstream screening. Its consensus hubs and high-betweenness compounds are proposed as starting points for docking, virtual screening, scaffold hopping, pharmacophore analysis, and pathway-focused repurposing workflows (Ali et al., 6 Mar 2026). The paper explicitly states that its main conclusion is not that active drugs are directly predicted, but that a large candidate library is reduced into robust similarity modules, consensus-stable pairs, and central backbone compounds (Ali et al., 6 Mar 2026).

FragAtlas-62M is framed as a fragment-specific foundation model for fragment-based drug discovery rather than a general molecular generator (Ho et al., 23 Sep 2025). The released resources include training code, preprocessed data, documentation, and model weights (Ho et al., 23 Sep 2025). At the same time, the paper states several limitations: no explicit stereochemistry modeling, no geometric or 3D structural reasoning, no fragment-to-fragment connectivity rules for automated fragment linking, no built-in synthetic-route awareness, and the need for fine-tuning in task-specific downstream use (Ho et al., 23 Sep 2025).

A recurring misconception is that ZINC-curated automatically implies either exhaustive chemical realism or direct translational prediction. The cited papers do not support that interpretation. Instead, they support a narrower and more technical view: curation defines a chemically and computationally tractable subspace, and the success of the resource depends on how well that subspace matches the target task.

7. Position within contemporary chemical informatics

The cited work places ZINC-curated chemistry at the intersection of three mature but still rapidly interacting paradigms: fragment-based representation, network-based chemical-space analysis, and foundation-scale generative modeling.

Fragment-based representation is exemplified by AGZ7, where a highly compressed AMON dictionary is proposed as a transferable basis for learning local chemistry from larger parent spaces (Huang et al., 2020). Network-based analysis is exemplified by the MYH9 study, which treats a curated ZINC library as a multi-layer graph and then quantifies modularity, co-clustering, minimum spanning tree topology, and betweenness centrality (Ali et al., 6 Mar 2026). Foundation-scale modeling is exemplified by FragAtlas-62M, which treats the complete ZINC-22 fragment subset as the substrate for a 42.7M-parameter chemical LLM with high validity and controlled novelty (Ho et al., 23 Sep 2025).

This trajectory suggests a broader synthesis. ZINC-curated resources increasingly function not merely as screening libraries but as formal computational objects: transferable fragment dictionaries, statistically validated network manifolds, and training corpora for generative priors. Within that synthesis, the central technical question is no longer whether ZINC is large, but how a given curation protocol converts that large and commercially grounded database into a representation whose inductive bias is appropriate for quantum prediction, chemical-space organization, or fragment discovery.