Molecular Graph Foundation Models

• MGFMs are specialized graph foundation models that pre-train on extensive molecular data to enable transfer across diverse tasks.
• They employ architectures including GNNs, LLMs, and hybrid systems with contrastive, generative, and reconstructive objectives.
• Adaptation strategies such as fine-tuning, prompt tuning, and parameter-efficient methods support robust property prediction and molecule generation.

Molecular Graph Foundation Models (MGFMs) are domain-specific graph foundation models specialized to molecular graphs, where atoms are nodes, bonds are edges, and the representation may additionally incorporate 2D/3D structure and textual or other molecular annotations. In the graph foundation model formulation, a model is expected to benefit from pre-training on broad graph data and then adapt to a wide range of downstream graph tasks; in the molecular case, this becomes pre-training on large corpora of molecules, followed by adaptation to property prediction, reaction prediction, molecule generation, docking or drug–target interaction, and related tasks (Liu et al., 2023). Contemporary surveys place MGFMs within a broader modular framework of backbone architectures, pretraining strategies, and adaptation mechanisms, and treat them as a central domain-specific branch of graph foundation models (Wang et al., 21 May 2025).

1. Conceptual foundations

The core formalization of an MGFM is a pretrained representation function over molecular graphs,

$f_\theta:\mathcal{G}\to\mathbb{R}^d,$

with downstream predictors built on top of the learned representation for atom-, bond-, or molecule-level tasks (Liu et al., 2023). In the molecular specialization described in the literature, downstream mappings include node tasks, edge tasks, and graph tasks, with the molecular graph $G\in\mathcal{G}$ providing the shared input object. This formulation distinguishes MGFMs from ordinary task-specific GNNs, which are typically trained end-to-end on a single benchmark and have limited ability to leverage massive unlabeled corpora or transfer broadly across tasks (Liu et al., 2023).

Four characteristics recur in the graph-foundation literature and transfer directly to molecules: pre-training on broad graph data, adaptation to diverse downstream tasks, emergence, and homogenization. In the molecular setting, emergence is associated with abilities such as in-context molecular reasoning or zero-shot generation once scale is sufficiently large, while homogenization denotes one backbone serving atom-, bond-, and molecule-level formats rather than a separate model per task (Liu et al., 2023). This framing is broad enough to include graph-only molecular encoders, sequence-centric models that treat molecules as SMILES or SELFIES, and graph–language hybrids.

A recurrent misconception is that a “foundation model” in chemistry must be a gigantic LLM. The survey literature and position papers do not support that restriction. One position paper argues that “primitive GFMs” are already present, including domain-specific molecular or atomistic models such as DiG, MoleBERT, Graphium, Uni-Mol+, GPS++, JMP-1, DPA-2, MACE-MP-0, and MACEmp (Mao et al., 2024). Another concrete counterexample is MiniMol, which is explicitly presented as a foundation model for molecular learning with only 10 million parameters (Kläser et al., 2024). This suggests that, in the molecular domain, “foundation” is tied more to breadth of pretraining data, task coverage, and transfer behavior than to parameter count alone.

A second important distinction concerns backbone limitations. The graph-foundation survey notes familiar constraints of classical GNNs—small parameter count, over-smoothing, over-squashing, and limited expressivity relative to transformers—and these issues are particularly relevant in chemistry because many molecular tasks depend on long-range interactions, non-local functional-group effects, and 3D geometry (Liu et al., 2023).

2. Architectural families

The standard taxonomy separates MGFMs into GNN-based, LLM-based, and hybrid GNN+LLM systems. The same taxonomy appears in both the early graph-foundation survey and later comprehensive surveys, and it captures most of the current molecular literature (Liu et al., 2023).

Family	Typical molecular representation	Representative exemplars
GNN-based	2D/3D molecular graphs	GROVER, Mole-BERT, MiniMol, Graphium, GTFM
LLM-based	SMILES, SELFIES, or graph-to-text serialization	LLM4Mol, MolecularGPT, GP-GPT, CaR, GraphsGPT
Hybrid GNN+LLM	Joint graph and text/image/knowledge representations	MolFM, MoMu, GIT-Mol, InstructMol, MolCA, UniMoT, GIMLET

In GNN-based MGFMs, the backbone is usually a message-passing network or graph transformer. A standard message-passing update is

$$h_v^{k+1} = U_k\Big(h_v^k,\; \sum_{u \in N(v)} M_k(h_v^k, h_u^k, \mathbf{X}^e_{(u,v)})\Big),$$

where $h_v^k$ is the embedding of atom $v$ at layer $k$, $\mathbf{X}^e_{(u,v)}$ denotes bond attributes, and $M_k,U_k$ are message and update functions (Liu et al., 2023). Graph transformers replace local aggregation with self-attention over graph elements and structural encodings. In chemistry this family includes graph-centric pretraining systems such as GROVER, Mole-BERT, Graphium, and MiniMol.

LLM-based MGFMs treat molecular inputs as language-like objects. Two input routes are described in the GFM survey: graph-to-token, where graph elements become special tokens, and graph-to-text, where structure is serialized into text such as edge lists, adjacency descriptions, or SMILES (Liu et al., 2023). In the molecular case this directly covers SMILES- and SELFIES-based models, as well as prompt-based systems such as LLM4Mol.

Hybrid GNN+LLM architectures combine structured graph encoders with language encoders. The survey distinguishes GNN-centric, symmetric, and LLM-centric hybrids, and molecular examples include Text2Mol, MoMu, MoleculeSTM, CLAMP, GIMLET, and the multimodal foundation model of Su et al. (Liu et al., 2023). A common alignment mechanism is graph–text contrastive learning: $$\mathcal{L}_{\text{GTCL}} = \mathbb{E}_{x, y^+, y^-} \left[ -\log \frac{\exp(k(x, y^+))} {\exp(k(x, y^+)) + \exp(k(x, y^-))} \right],$$ where $x$ is a graph embedding, $G\in\mathcal{G}$0 a matched text, $G\in\mathcal{G}$1 a mismatched text, and $G\in\mathcal{G}$2 a similarity score (Liu et al., 2023).

Several representative systems enlarge the modality set beyond graph and text. MolFM learns jointly from molecular structures, biomedical texts, and knowledge graphs through cross-modal attention between atoms, molecule-entity neighbors, and semantically related texts (Luo et al., 2023). MMELON aggregates graph, image, and text views, with single-view models pretrained on datasets of up to 200M molecules and then fused into combined representations (Suryanarayanan et al., 2024). MolGraph-LarDo adds a different hybridization strategy by calibrating LLM-produced molecular descriptions with a domain-specific small model, instantiated as RDKit, and then aligning graph and text via a CLIP-style loss (Zhang et al., 2024).

3. Pretraining objectives and representation learning

MGFM pretraining inherits the main graph-foundation objectives: contrastive, generative or reconstructive, and language-model-style objectives over textual molecular representations. The same-scale contrastive pattern uses two augmented views $G\in\mathcal{G}$3 of the same molecule and an InfoNCE loss,

$G\in\mathcal{G}$4

with augmentations such as atom masking, bond dropping, substructure perturbation, or conformer variation (Liu et al., 2023). Cross-scale contrastive learning aligns atom-level and graph-level embeddings, while graph–text contrastive learning aligns molecular structures with textual descriptions, assay metadata, or literature-derived semantics.

Generative and reconstructive objectives include adjacency or bond reconstruction, masked feature reconstruction, and autoregressive graph generation. For example, a variational graph autoencoder style objective reconstructs adjacency by

$G\in\mathcal{G}$5

while masked atom-feature reconstruction can be written as

$G\in\mathcal{G}$6

over masked atoms $G\in\mathcal{G}$7 (Liu et al., 2023). Molecular graph transformers such as GROVER also use property-prediction pretext tasks, predicting motifs or graph-level chemical attributes from learned embeddings.

Sequence-based MGFMs use language modeling and masked language modeling over SMILES or related serializations: $G\in\mathcal{G}$8 with masked-token prediction as the MLM analogue (Liu et al., 2023). This route is especially natural for LLM-based systems and for hybrid systems where graph and sequence views are both available.

The scaling question is central. The graph-foundation survey emphasizes that emergent abilities tend to appear only when parameter counts and training data scale up, and that data quality and diversity are crucial (Liu et al., 2023). In the molecular setting, one supervised infrastructure paper introduces ToyMix, LargeMix, and UltraLarge datasets covering nearly 100 million molecules and over 3000 sparsely defined tasks, totaling more than 13 billion individual labels of both quantum and biological nature (Beaini et al., 2023). MiniMol gives a compact instantiation of the same principle: it is pretrained on approximately 6 million molecules, 526 million labeled data points, and roughly 3300 graph- and node-level tasks of both quantum and biological nature (Kläser et al., 2024). MMELON shows an alternative large-scale route, pretraining its graph and text models on the same 200 million molecules (Suryanarayanan et al., 2024).

These examples indicate that MGFM pretraining has split into at least two viable regimes. One regime emphasizes large unlabeled corpora with self-supervised objectives; the other emphasizes very large, sparse, supervised multi-task corpora. The available evidence does not imply that one regime universally dominates the other, but it does show that both are active and technically mature.

4. Adaptation, downstream tasks, and task transfer

The standard adaptation toolbox comprises vanilla fine-tuning, parameter-efficient fine-tuning, prompt tuning, and in-context learning (Liu et al., 2023). In the generic molecular formulation, property prediction is written as

$G\in\mathcal{G}$9

with $$h_v^{k+1} = U_k\Big(h_v^k,\; \sum_{u \in N(v)} M_k(h_v^k, h_u^k, \mathbf{X}^e_{(u,v)})\Big),$$0 typically much smaller than the pretrained backbone (Liu et al., 2023). Reaction prediction, drug–target interaction, and synthesis-related tasks follow the same template: encode structured molecular inputs with the pretrained backbone, then add a lightweight task head.

Prompt-style adaptation is attractive in low-data regimes. The GFM survey discusses GraphPrompt, GPPT, GPF, and related methods that recast downstream classification as prompt-conditioned similarity or link prediction (Liu et al., 2023). In molecular applications, this can mean using learned task vectors, prototype molecules, or class tokens to adapt a pretrained encoder to toxicity, solubility, or activity labels without full backbone updates.

Fine-tuning behavior is not uniform across pretraining regimes. RoFt-Mol benchmarks eight fine-tuning methods on six pretrained molecular encoders and classifies them into weight-based, representation-based, and partial fine-tuning mechanisms (Liu et al., 30 Aug 2025). Its central empirical conclusions are that, under few-shot fine-tuning, supervised pretraining yields better performance than self-supervised pretraining, that self-supervised MGFMs favor weight-based methods such as WiSE-FT and $$h_v^{k+1} = U_k\Big(h_v^k,\; \sum_{u \in N(v)} M_k(h_v^k, h_u^k, \mathbf{X}^e_{(u,v)})\Big),$$1-SP, and that supervised MGFMs favor representation-based methods such as Feature-map or Surgical FT (Liu et al., 30 Aug 2025). The same study also reports that linear probing is consistently the worst method for self-supervised pretrained models across splits and few-shot settings (Liu et al., 30 Aug 2025). This is a direct warning against assuming that parameter-efficient adaptation behaves the same way for all MGFM backbones.

MiniMol provides a concrete, graph-based example of economical adaptation. The backbone is kept fixed as a feature extractor, and downstream ADMET models are small MLP heads trained on top of the pooled fingerprint $$h_v^{k+1} = U_k\Big(h_v^k,\; \sum_{u \in N(v)} M_k(h_v^k, h_u^k, \mathbf{X}^e_{(u,v)})\Big),$$2 (Kläser et al., 2024). On the TDC ADMET group, the best MiniMol variant has mean rank 3.6 compared with MolE at 5.4, outperforms MolE on 17 tasks, and reaches top-1 performance on 8 tasks and top-3 performance on 10 tasks (Kläser et al., 2024). This result is notable because it demonstrates strong transfer from a shared molecular graph encoder without full end-to-end retuning.

5. Representative systems and model lineages

Several concrete systems illustrate how the MGFM design space has diversified.

MolFM is a multimodal molecular foundation model that jointly learns from molecular structures, biomedical texts, and knowledge graphs. Its graph encoder is a 5-layer GIN initialized from GraphMVP; its text pathway uses a 6-layer transformer, and its knowledge pathway uses TransE-style embeddings over a molecule-centric knowledge graph (Luo et al., 2023). The pretraining objective combines structure–text contrastive loss, cross-modal matching, masked language modeling with multimodal context, and knowledge-graph embedding. On cross-modal retrieval, MolFM reports 12.13% and 5.04% absolute gains under the zero-shot and fine-tuning settings, respectively (Luo et al., 2023). This places MolFM among the clearest examples of a hybrid MGFM that treats structured chemistry, literature, and explicit biomedical relations as coequal data sources.

MoMu is an earlier molecular multimodal foundation model that aligns a GIN-based graph encoder with scientific text retrieved from S2ORC. Its pretraining corpus contains 15,613 molecule graph–document pairs and approximately 37 million paragraphs of text (Su et al., 2022). MoMu uses contrastive graph–text learning with graph augmentations and multiple text views, and it is evaluated on cross-modal retrieval, molecule captioning, molecular property prediction, and zero-shot text-to-graph generation (Su et al., 2022). Historically, it marks an early direct bridge between molecular graphs and natural language.

MolGraph-LarDo shows a different multimodal route: rather than merely aligning graph and text, it first improves the textual modality itself by combining an LLM with a domain-specific small model. In the reported implementation, the LLM is Mistral-7B-Instruct-v0.2 and the DSM is RDKit 2023.3.2, which provides calibrated molecular metrics such as molecular weight and LogP (Zhang et al., 2024). The resulting MD-Text descriptions are aligned with GIN graph embeddings by a symmetric graph–text contrastive objective. On MoleculeNet benchmarks, MolGraph-LarDo consistently improves over GCN, GIN, Hu et al., GROVER, and MolCLR-GIN; for example, on QM7 it reports MAE 74.91 compared with GROVER at 80.23 (Zhang et al., 2024).

MMELON broadens the notion of a molecular foundation model from graph–text to graph–image–text. Its graph component is a TokenGT-based transformer operating on molecular graphs tokenized as atom and bond tokens with spectral positional encodings, and it is pretrained with masked feature prediction, corrupted edge prediction, and Betti number prediction (Suryanarayanan et al., 2024). Single-view graph, image, and text models are each pretrained on up to 200M molecules and then fused by an attention-based aggregator (Suryanarayanan et al., 2024). The resulting multi-view model is validated on 18 tasks and applied to screen compounds against a large set of GPCR targets, including 33 Alzheimer’s disease-related GPCRs (Suryanarayanan et al., 2024).

CoMole exemplifies the generative branch of MGFM research. It is framed as a family of Controllable Molecular Generative Foundation Models built with a unified motif-aware graph diffusion pipeline, followed by supervised fine-tuning and reinforcement learning (Zhu et al., 14 May 2026). Across three heterogeneous benchmarks spanning materials and drug discovery, CoMole ranks first in controllability on all nine targets, reduces MAE by up to 48.2% relative to the strongest baselines, and maintains validity above 0.94 without rule-based correction or post-hoc filtering (Zhu et al., 14 May 2026). Its central architectural move is to operate in a motif-aware graph space rather than an atom-wise action space, thereby making controllable generation more stable.

Taken together, these systems show that “MGFM” is not a single architecture class. It already includes graph-only multi-task encoders, graph–text aligners, graph–text–knowledge systems, graph–image–text aggregators, and controllable generative models. This suggests that the unifying concept is the pretrain–adapt paradigm over molecular graph structure, rather than any one modality choice.

6. Reliability, interpretability, and open problems

The graph-foundation surveys identify several persistent challenges that intensify in molecular applications: data quantity and quality, heterogeneity and domain shift, scalability, over-smoothing or over-squashing, evaluation of open-ended generative tasks, and the combined problem of safety, interpretability, and scientific validity (Liu et al., 2023). In molecules, these issues appear as biased chemical corpora, assay noise, scaffold and size shifts, difficulty evaluating generated structures, and the risk that a model produces chemically implausible or hazardous outputs with unwarranted confidence.

A major recent reliability theme is out-of-distribution detection. Mole-PAIR does not introduce a new MGFM backbone; instead, it adds a post-training scoring head to existing molecular foundation encoders and trains it by preference optimization over ID–OOD pairs (He et al., 29 Sep 2025). The paper frames the key failure mode as chemical hallucination and reports up to 45.8%, 43.9%, and 24.3% improvements in AUROC under distribution shifts of size, scaffold, and assay, respectively (He et al., 29 Sep 2025). This is important because it shifts part of the MGFM safety problem from backbone design to post hoc reliability calibration.

Interpretability is an independent line of development. Foundation Molecular Grammar induces an interpretable molecular language by using a multimodal foundation model to render molecules as images, describe them as text, and align information across modalities during graph-grammar induction (2505.22948). The resulting framework can be used as a drop-in replacement for prior grammar learning approaches in molecular generation and property prediction and is explicitly designed to provide built-in chemical interpretability (2505.22948). In a different but related direction, CoMole’s motif-aware action space can be read as an interpretability device as well as a control mechanism, because the generator’s decisions are made over chemically meaningful motifs rather than raw atom-level edits (Zhu et al., 14 May 2026).

Future directions proposed in the graph-foundation surveys translate cleanly to molecules: larger and more diverse molecular graph corpora; unified benchmarks spanning property prediction, reaction prediction, synthesis planning, generative tasks, and cross-modal retrieval; backbone exploration beyond transformers; unified pretext tasks that combine masked atom or bond features, motif prediction, 3D geometry prediction, and graph–text alignment; prompt-based tuning for new molecular tasks; reinforcement learning from scientist feedback; and knowledge distillation from large MGFMs into lightweight task models (Liu et al., 2023). A plausible implication is that the next phase of MGFMs will be defined less by whether the model is graph-based or language-based, and more by how effectively it combines structural pretraining, multimodal alignment, robust adaptation, and reliability control within a single molecular modeling stack.