Unifying Polymer Modeling and Design via a Conformation-Centric Generative Foundation Model

TL;DR

Polymers require global conformational information for accurate modeling, yet existing approaches rely largely on monomer-level descriptors. PolyConFM introduces a conformation-centric generative foundation model that represents a polymer conformation as a sequence of repeating-unit conformations with orientation frames, and is pretrained in two phases: a MAR-based reconstruction of ${\mathcal C}^{u}$ and an SO(3) diffusion for rotations, conditioned on the polymer graph ${\mathcal G}$. A large MD-derived dataset of over $50{,}000$ polymers supports this pretraining, enabling generation of conformations and global embeddings that improve downstream tasks. Finetuning shows state-of-the-art performance in downstream polymer property prediction (across eight datasets) and in polymer design under various conditioning signals, demonstrating a universal backbone that links polymer structure, properties, and design. This conformation-focused framework advances polymer informatics by enabling structurally informed generation and design with physics-informed priors.

Abstract

Polymers, macromolecules formed from covalently bonded monomers, underpin countless technologies and are indispensable to modern life. While deep learning is advancing polymer science, existing methods typically represent the whole polymer solely through monomer-level descriptors, overlooking the global structural information inherent in polymer conformations, which ultimately limits their practical performance. Moreover, this field still lacks a universal foundation model that can effectively support diverse downstream tasks, thereby severely constraining progress. To address these challenges, we introduce PolyConFM, the first polymer foundation model that unifies polymer modeling and design through conformation-centric generative pretraining. Recognizing that each polymer conformation can be decomposed into a sequence of local conformations (i.e., those of its repeating units), we pretrain PolyConFM under the conditional generation paradigm, reconstructing these local conformations via masked autoregressive (MAR) modeling and further generating their orientation transformations to recover the corresponding polymer conformation. Besides, we construct the first high-quality polymer conformation dataset via molecular dynamics simulations to mitigate data sparsity, thereby enabling conformation-centric pretraining. Experiments demonstrate that PolyConFM consistently outperforms representative task-specific methods on diverse downstream tasks, equipping polymer science with a universal and powerful tool.

Figures (7)

• Figure 1: Overview of the proposed PolyConFM.a.The overall scheme of PolyConFM: PolyConFM employs polymer conformation generated by itself as input to provide global structural information for downstream tasks, while the modeling module can also assist the design module via virtual screening to prioritize candidates, thereby positioning PolyConFM as a unified backbone that seamlessly bridges polymer structure, property, and design. b.The frame-based polymer representation: The complete polymer conformation is decomposed into a sequence of repeating-unit conformations with identical SMILES strings and distinct 3D structures, overlapping at those key atoms (e.g., atom-1 of the current repeating unit aligns with atom-3 of the preceding repeating unit). Here, the orientation transformation contained in the corresponding frame is denoted as $\mathcal{O} = (\bm R, \bm t)$, where rotation transformation $\bm R \in \mathbb{R}^{3 \times 3}$ is calculated via the Gram-Schmidt procedure on vectors $\bm v_1$ and $\bm v_2$ and translation transformation $\bm t \in \mathbb{R}^{3}$ corresponds to the 3D coordinate of atom-3. c.Conformation-centric generative pretraining: PolyConFM is first trained to reconstruct repeating-unit conformations via masked autoregressive modeling and then trained to generate the required orientation transformations for assembling them to recover the corresponding polymer conformation, thereby enabling it to capture dependencies among repeating units for global structure modeling while simultaneously unlocking conformation generation capability for downstream tasks. d.The inference framework of PolyConFM: We directly run inference with the pretrained PolyConFM to generate the corresponding polymer conformation, which is then added to the input to derive polymer embedding for downstream tasks.
• Figure Supplementary Figure 1: Supplementary Figure 0: The distribution of repeating-unit counts per conformation within the conformation dataset, which is further partitioned into training, validation, and test sets.
• Figure Supplementary Figure 2: Supplementary Figure 1: Several visualization examples of TorsionalDiff (i.e., the best baseline) and PolyConFM on the polymer conformation generation task.
• Figure Supplementary Figure 3: Supplementary Figure 2: The t-SNE visualization of PolyConFM on the downstream polymer property prediction task, where the ground‑truth property values determine point colors.
• Figure Supplementary Figure 4: Supplementary Figure 3: The scatter plots of PolyConFM on the downstream polymer property prediction task, covering eight typical polymer property datasets.