MolCluster: Unsupervised Molecular Mapping

• MolCluster is an unsupervised framework that learns coarse-grained mappings by integrating a GNN with community detection for chemical consistency.
• It employs a predefined group pair loss and a bisection-based resolution control strategy to customize coarse-grained mappings across diverse molecular systems.
• Empirical evaluations on the MARTINI2 dataset demonstrate that MolCluster outperforms traditional clustering and supervised methods in key metrics like precision and F1-score.

MolCluster is an automated, unsupervised framework for coarse-grained (CG) mapping of molecular systems that integrates a graph neural network (GNN) with a community detection algorithm to extract CG representations from molecular graphs. It was proposed to address two limitations emphasized for prior CG workflows: traditional CG approaches rely on fixed mapping rules and require extensive manual intervention, while supervised learning-based CG methods are constrained by limited labeled datasets and by the inability to control mapping resolution. In MolCluster, chemically meaningful groupings are learned without labeled mapping data; a predefined group pair loss preserves target groups, and a bisection strategy provides precise, customizable resolution across molecular systems. On the MARTINI2 dataset, its label-free pretraining strategy yields stronger downstream performance than both traditional clustering and supervised models, and the formulation is presented as a basis for customizable and chemically consistent CG mapping (Zhong et al., 25 Sep 2025).

1. Problem formulation and conceptual position

Coarse-grained modeling simplifies molecular systems by mapping groups of atoms into representative units. Within that setting, MolCluster is positioned against two established baselines. First, rule-based CG schemes employ fixed mapping rules, which limits adaptability across chemically diverse systems. Second, supervised CG mapping methods automate part of the process, but depend on labeled mappings and do not provide direct control over mapping resolution. MolCluster is designed specifically to overcome those limitations through unsupervised representation learning coupled to graph partitioning (Zhong et al., 25 Sep 2025).

The method converts a molecule into a graph whose nodes are atoms and whose edges are bonds. At the modeling level, the goal is not merely to cluster atoms by geometric proximity, but to partition the atomic graph into CG beads that are chemically meaningful and structurally consistent. The framework therefore combines local chemical representation learning with a global partitioning mechanism based on community structure. This suggests an architectural commitment to learned molecular similarity rather than to predetermined fragmentation rules.

A common misconception is that “unsupervised” here means unconstrained or chemically agnostic. In the MolCluster formulation, unsupervised training is paired with explicit chemical and structural constraints, particularly through the predefined group pair loss that preserves domain-knowledge-based groups such as functional groups defined via SMARTS patterns (Zhong et al., 25 Sep 2025).

2. Architecture: GNN-derived similarity graph and Leiden partitioning

MolCluster’s architecture is organized around two coupled components: representation learning by a GNN and graph partitioning by the Leiden algorithm. The molecular input is represented as a graph in which atoms are nodes, with features such as atom type, aromaticity, degree, and hybridization, and bonds are edges, with features such as bond type, length, and direction (Zhong et al., 25 Sep 2025).

Each atom’s one-hot encoding is embedded into a high-dimensional vector, concatenated with a rich set of atomic property features, and then transformed with a Transformer encoder to capture global context. Edge features are processed through an MLP to permit dynamic weighting. Node features are updated iteratively through message passing so that each atom representation reflects its chemical environment. After multiple GNN/Transformer layers, a Gaussian kernel computes pairwise edge weights between atoms, producing an adjacency matrix that encodes atom-wise similarity (Zhong et al., 25 Sep 2025).

This learned adjacency matrix is then passed to the Leiden community detection algorithm, described as an advanced variant of Louvain optimized for modularity. Leiden partitions the molecular graph into densely connected communities, and these communities define the CG beads. Because the partitioning operates on similarities derived from learned atom features rather than from static heuristics alone, the number and composition of CG groups are adapted in a data-driven manner (Zhong et al., 25 Sep 2025).

The architectural significance lies in the division of labor. The GNN learns a chemically informed similarity geometry over atoms, while community detection enforces a discrete partition of the graph. This suggests that MolCluster treats CG mapping as a structured graph partitioning problem rather than as direct node labeling.

3. Objective functions and chemically informed constraints

MolCluster is trained without labeled CG mappings. Its unsupervised objective combines modularity-driven graph partitioning with losses that encode chemical consistency. The training signal is described as a combination of a triplet loss and a predefined group pair loss, with the overall objective

$$\mathcal{L} = \mathcal{L}_{\text{triplet}} + \mathcal{L}_{\text{group}}.$$

The triplet loss is described as enforcing that atoms in the same group are closer in the learned feature space than atoms in different groups (Zhong et al., 25 Sep 2025).

The predefined group pair loss is given explicitly as

$$\mathcal{L}_{\text{group}} = \lambda_{\text{group}} \cdot \sum_{(i, j) \in \mathcal{P}} \left[ 1 - \exp\left(-\frac{\|\mathbf{f}_i - \mathbf{f}_j\|_2^2}{2\sigma^2}\right) \right],$$

where $\mathbf{f}_i$ and $\mathbf{f}_j$ are GNN-learned feature vectors of atoms $i$ and $j$ within the same predefined chemical group, $\mathcal{P}$ is the set of all such pairs, $\lambda_{\text{group}}$ is a loss weight, and $\sigma$ is a Gaussian bandwidth (Zhong et al., 25 Sep 2025).

Functionally, this term penalizes feature dissimilarity among atoms that should remain cohesive, thereby preserving target groups. In practice, those target groups are described as user-defined chemical groups, for example functional groups identified via SMARTS patterns. The method therefore incorporates chemical prior knowledge without requiring full supervision at the bead-assignment level. A plausible implication is that MolCluster occupies an intermediate regime between purely heuristic mapping and label-dependent end-to-end prediction.

The formulation also clarifies another frequent misunderstanding: MolCluster is not simply applying community detection to a raw molecular graph. The communities are extracted from a learned similarity graph whose topology is shaped by feature learning and by explicit cohesion constraints on predefined groups (Zhong et al., 25 Sep 2025).

4. Resolution control through modularity and bisection

A central design point in MolCluster is explicit control of CG resolution. The resolution is defined as

$r = \frac{N_{\text{groups}}}{N_{\text{atoms}}},$

where $$\mathcal{L}_{\text{group}} = \lambda_{\text{group}} \cdot \sum_{(i, j) \in \mathcal{P}} \left[ 1 - \exp\left(-\frac{\|\mathbf{f}_i - \mathbf{f}_j\|_2^2}{2\sigma^2}\right) \right],$$0 is the number of CG beads and $$\mathcal{L}_{\text{group}} = \lambda_{\text{group}} \cdot \sum_{(i, j) \in \mathcal{P}} \left[ 1 - \exp\left(-\frac{\|\mathbf{f}_i - \mathbf{f}_j\|_2^2}{2\sigma^2}\right) \right],$$1 is the atom count (Zhong et al., 25 Sep 2025).

The partitioning itself is governed through the Leiden modularity parameter $$\mathcal{L}_{\text{group}} = \lambda_{\text{group}} \cdot \sum_{(i, j) \in \mathcal{P}} \left[ 1 - \exp\left(-\frac{\|\mathbf{f}_i - \mathbf{f}_j\|_2^2}{2\sigma^2}\right) \right],$$2 in the modularity objective

$$\mathcal{L}_{\text{group}} = \lambda_{\text{group}} \cdot \sum_{(i, j) \in \mathcal{P}} \left[ 1 - \exp\left(-\frac{\|\mathbf{f}_i - \mathbf{f}_j\|_2^2}{2\sigma^2}\right) \right],$$3

In this formulation, a larger $$\mathcal{L}_{\text{group}} = \lambda_{\text{group}} \cdot \sum_{(i, j) \in \mathcal{P}} \left[ 1 - \exp\left(-\frac{\|\mathbf{f}_i - \mathbf{f}_j\|_2^2}{2\sigma^2}\right) \right],$$4 tends to split communities more finely and thus yields higher resolution, i.e., more CG beads (Zhong et al., 25 Sep 2025).

To achieve a user-specified target number of groups, MolCluster applies a bisection search on $$\mathcal{L}_{\text{group}} = \lambda_{\text{group}} \cdot \sum_{(i, j) \in \mathcal{P}} \left[ 1 - \exp\left(-\frac{\|\mathbf{f}_i - \mathbf{f}_j\|_2^2}{2\sigma^2}\right) \right],$$5. Starting from initial bounds $$\mathcal{L}_{\text{group}} = \lambda_{\text{group}} \cdot \sum_{(i, j) \in \mathcal{P}} \left[ 1 - \exp\left(-\frac{\|\mathbf{f}_i - \mathbf{f}_j\|_2^2}{2\sigma^2}\right) \right],$$6 and $$\mathcal{L}_{\text{group}} = \lambda_{\text{group}} \cdot \sum_{(i, j) \in \mathcal{P}} \left[ 1 - \exp\left(-\frac{\|\mathbf{f}_i - \mathbf{f}_j\|_2^2}{2\sigma^2}\right) \right],$$7, it iteratively halves the interval, reruns the partitioning, and continues until the resulting $$\mathcal{L}_{\text{group}} = \lambda_{\text{group}} \cdot \sum_{(i, j) \in \mathcal{P}} \left[ 1 - \exp\left(-\frac{\|\mathbf{f}_i - \mathbf{f}_j\|_2^2}{2\sigma^2}\right) \right],$$8 is within the desired tolerance of $$\mathcal{L}_{\text{group}} = \lambda_{\text{group}} \cdot \sum_{(i, j) \in \mathcal{P}} \left[ 1 - \exp\left(-\frac{\|\mathbf{f}_i - \mathbf{f}_j\|_2^2}{2\sigma^2}\right) \right],$$9 (Zhong et al., 25 Sep 2025).

This mechanism is important because resolution control is identified as essential for multiscale modeling. The bisection strategy is described as robust and system-agnostic, such that the same $\mathbf{f}_i$0 yields comparable mapping granularity for molecules of different topology or size. The practical implication presented in the source material is that the same framework can be deployed for small molecules, polymers, and even proteins (Zhong et al., 25 Sep 2025).

5. Empirical evaluation on MARTINI2

MolCluster is evaluated on the MARTINI2 dataset against classical clustering methods and supervised deep learning models, including DSGPM-TP. The reported clustering metrics include Adjusted Mutual Information (AMI), Cut Precision, Cut Recall, and Cut F1-Score. MolCluster is reported to outperform both classical clustering methods—HDBSCAN, FINCH, spectral clustering, Graclus, and METIS—and learned approaches such as DSGPM and DSGPM-TP (Zhong et al., 25 Sep 2025).

The key quantitative results reported for selected methods are as follows:

Method	AMI / Cut F1	Type Prec. (F1)
HDBSCAN	0.35±0.04 / 0.27	—
Spectral Clustering	0.75±0.06 / 0.54	—
DSGPM-TP	0.85±0.03 / 0.71	0.85 (0.82)
MolCluster	0.88±0.01 / 0.79	0.89 (0.87)
MolCluster (finetuned)	0.91±0.00 / 0.83	—

The same evaluation reports Cut Precision, Recall, and F1 of 0.79 for MolCluster and 0.83 for MolCluster (finetuned). For bead type prediction, MolCluster reports Type Precision 0.89 and F1 0.87, compared with 0.85 and 0.82 for DSGPM-TP (Zhong et al., 25 Sep 2025).

Several points follow directly from these results. First, label-free pretraining is not merely a pretext task; it improves downstream CG mapping quality relative to both traditional clustering and supervised baselines. Second, pretraining plus finetuning yields further gains, which objectively indicates that the strongest configuration in the reported experiments is not the frozen unsupervised model alone. Third, visualization via t-SNE is described as confirming that learned representations generalize well, grouping molecules by physical and chemical similarity even without explicit supervision (Zhong et al., 25 Sep 2025).

6. Relation to adjacent clustering methods in molecular machine learning

MolCluster belongs to a broader set of molecular ML methods that use clustering, but its target object and objective are distinct. In molecular-orbital-based machine learning, clustering has been used to partition pair-energy data into chemically intuitive local regression domains. Regression-clustering in MOB-ML partitions feature–label pairs into clusters fitted by local linear regression models, with subsequent cluster-wise regression and classification; later work replaced supervised clustering with an unsupervised Gaussian mixture model and scalable Gaussian process regression for improved efficiency and transferability (Cheng et al., 2019, Cheng et al., 2022). Those methods cluster orbital-feature datapoints for energy learning, whereas MolCluster clusters atoms into CG beads for mapping.

Another nearby line is hierarchical clustering inside geometric GNNs for long-range interaction modeling. Multi-stage Clustered Global Modeling (MCGM) introduces adaptive multi-resolution atomic clusters to distill and disseminate global context in geometric GNN backbones, improving long-range molecular property prediction (Pan et al., 26 Sep 2025). That use of clustering operates as a message-passing augmentation for energy and force models, not as a discrete coarse-graining map. This suggests that MolCluster’s distinctive contribution is the coupling of learned atomic representations to community detection specifically for CG partition generation.

A more general misconception is to treat all molecular clustering methods as interchangeable. The data provided across related work indicate otherwise: local density clustering has been used for molecular clump extraction in spectroscopic data cubes (Luo et al., 2021), and determinant-space clustering has been used to characterize multi-determinant wave functions in quantum chemistry (Sun et al., 2022). MolCluster differs from both in the object being clustered, the graph representation employed, and the downstream purpose.

Within CG mapping specifically, the combination of unsupervised learning, predefined group preservation, and explicit resolution control is the defining profile of MolCluster. The source material therefore frames it not simply as another clustering algorithm, but as a customizable and chemically consistent CG mapping model with a label-free pretraining pathway and competitive downstream accuracy on MARTINI2 (Zhong et al., 25 Sep 2025).