MIPS: Multimodal Infinite Polymer Sequence

• Multimodal Infinite Polymer Sequence (MIPS) is a framework that represents polymers as infinite, periodic graphs combining topological and spatial information for property prediction.
• It introduces an induced star-linking graph to convert infinite polymer sequences into tractable finite graphs using message passing and localized graph attention.
• Empirical results show that MIPS achieves state-of-the-art performance on eight polymer property tasks while maintaining invariance under cyclic shifts and repeats.

Multimodal Infinite Polymer Sequence (MIPS) is a pre-training framework for polymer property prediction that systematically addresses the limitations of conventional monomer-based representations. By modeling polymers as infinite sequences of repeating monomer units and integrating both topological and spatial descriptors, MIPS establishes new methodology for comprehensive polymer modeling and property prediction across a wide range of applications (Wang et al., 27 Jul 2025).

1. Infinite Polymer Sequence Formalism

A polymer is conceptualized as an infinite concatenation of a single monomer unit. Let the monomer be represented as a finite graph $G = (V, E, X)$, where $V = \{v_0, ..., v_{|V|-1}\}$ are atom nodes, $E \subset V \times V$ are chemical bonds, and $X \in \mathbb{R}^{d \times |V|}$ encodes atom-level features. The special boundary atoms $v_0$ and $v_{|V|-1}$ represent the sites by which monomers are covalently linked.

The infinite polymer sequence is constructed by defining $V^p = \{v^p_i : i \in \mathbb{Z}\}$ (atom indices in the polymer), $X^p_i = X_{(i \bmod |V|)}$ (attributing each site its monomer features), and two types of edges:

• $E_{\text{intra}}$: Intramonomeric bonds, maintaining the connectivity of the base unit,
• $E_{\text{inter}}$: Bonds linking the right boundary atom of one monomer to the left boundary atom of the next.

Thus, the polymer is modeled as a periodic, infinite graph, where computation over all $V = \{v_0, ..., v_{|V|-1}\}$0 is intractable unless suitable reductions are devised.

To achieve tractable computation, MIPS introduces the induced star-linking graph $V = \{v_0, ..., v_{|V|-1}\}$1, defined as $V = \{v_0, ..., v_{|V|-1}\}$2, $V = \{v_0, ..., v_{|V|-1}\}$3, $V = \{v_0, ..., v_{|V|-1}\}$4, effectively closing the monomer to a ring via a "star-link" at the boundary atoms. This representation undergirds efficient modeling of infinite polymer sequences in downstream architectures.

2. Topological Modeling: Message Passing and Graph Attention

2.1 Message Passing Mechanism (MPM) on Infinite Sequences

Standard Message Passing Mechanisms (MPMs) iteratively update node features:

$V = \{v_0, ..., v_{|V|-1}\}$5

Applying MPMs directly to the infinite $V = \{v_0, ..., v_{|V|-1}\}$6 is not computationally feasible. However, by leveraging the periodic structure, MIPS shows (Proposition 2.1) that applying MPM to the infinite sequence is algebraically equivalent to running MPM on the induced star-linking graph $V = \{v_0, ..., v_{|V|-1}\}$7. Theoretical guarantees (Theorem 2.2) ensure that any GNN built from MPM layers, nodewise transforms, and mean pooling admits $V = \{v_0, ..., v_{|V|-1}\}$8 (Wang et al., 27 Jul 2025).

2.2 Localized Graph Attention (LGA)

Graph Attention Mechanisms (GAMs) generalize message passing by using attention weights:

$V = \{v_0, ..., v_{|V|-1}\}$9

with $E \subset V \times V$0 and $E \subset V \times V$1 encoding learned spatial and path biases.

To constrain attention to a finite neighborhood in $E \subset V \times V$2, MIPS applies Localized Graph Attention (LGA):

$E \subset V \times V$3

and proves (Theorem 2.3) that a sufficiently deep stack of LGA layers reproduces exact results on $E \subset V \times V$4 if the boundary-atom distance in the monomer exceeds $E \subset V \times V$5.

2.3 Repeat and Shift Invariance Test (RSIT)

RSIT evaluates whether a model's predictions are invariant under arbitrary cyclic rotations ("shift") and concatenations ("repeat") of the polymer's sequence (mimicking P-SMILES robustness). The algorithm perturbs input sequences and assesses the worst-case predictive loss across $E \subset V \times V$6 trials. Among multiple strategies, only the star-linking approach maintains zero or near-zero performance drop under this test, as indicated in Table 3.1 of the source (Wang et al., 27 Jul 2025).

2.4 Weisfeiler-Lehman Limitation and Backbone Embedding

Certain "twin polymer graphs"—distinct polymers whose $E \subset V \times V$7 coincide—cannot be distinguished by the Weisfeiler-Lehman (WL) test, limiting MPM and LGA expressivity (Theorem 3.2). MIPS addresses this by backbone embedding: atoms on the shortest path ("backbone") between boundary nodes are augmented with an embedding $E \subset V \times V$8, $E \subset V \times V$9 if on backbone, 0 otherwise; all features are updated $X \in \mathbb{R}^{d \times |V|}$0. Downstream, a Localized Graph Transformer encodes these augmented features, improving the model's ability to distinguish polymer graphs, especially those with side-chain ring structures.

3. Spatial Descriptor Extraction

MIPS supplements topological encodings with spatial descriptors derived from monomer 3D coordinates. For each repeating monomer, the following geometric descriptors are computed:

• Pairwise distances $X \in \mathbb{R}^{d \times |V|}$1
• Bond angles $X \in \mathbb{R}^{d \times |V|}$2
• Dihedral (torsion) angles $X \in \mathbb{R}^{d \times |V|}$3
• Additional higher-order geometric features

Each type of descriptor $X \in \mathbb{R}^{d \times |V|}$4 is individually linearly projected: $X \in \mathbb{R}^{d \times |V|}$5 ($X \in \mathbb{R}^{d \times |V|}$6) and the stack $X \in \mathbb{R}^{d \times |V|}$7 is formed for fusion with topological embeddings.

4. Cross-Modal Fusion of Topological and Spatial Information

The unification of topological ($X \in \mathbb{R}^{d \times |V|}$8) and spatial ($X \in \mathbb{R}^{d \times |V|}$9) representations is achieved via cross-attention:

$v_0$0

where $v_0$1 indicates layer normalization. This architecture allows polymer models to incorporate both sequential bonding pattern information and three-dimensional geometry, yielding a multimodal representation of polymer systems.

5. Empirical Evaluation and Performance

MIPS is evaluated on eight downstream tasks, each corresponding to a DFT-computed polymer property: chain bandgap (Egc), bulk bandgap (Egb), electron affinity (Eea), ionization energy (Ei), crystallization percentage (Xc), dielectric constant (EPS), refractive index (Nc), and atomization energy per atom (Eat). Five-fold cross-validation is used, and performance is measured by RMSE and $v_0$2.

Reported $v_0$3 results:

Method	Egc	Egb	Eea	Ei	Xc	EPS	Nc	Eat
MMPolymer	0.924	0.934	0.925	0.836	0.488	0.779	0.864	0.961
MIPS (ours)	0.926	0.945	0.940	0.846	0.506	0.814	0.877	0.990

MIPS achieves state-of-the-art performance by clear margins across all property prediction tasks (Wang et al., 27 Jul 2025).

6. Theoretical Contributions and Future Directions

MIPS establishes several foundational results:

• Star-linking provides an equivalence between modeling infinite polymer sequences and computations on a finite, augmented graph.
• Localized Graph Attention extends transformer-style architectures to infinite, periodic graphs, maintaining locality and computational feasibility.
• The Repeat and Shift Invariance Test offers a robust invariance benchmark for polymer models.
• Backbone embedding circumvents the expressivity limitations imposed by the Weisfeiler-Lehman test, especially for monomers containing side-chain rings.

Practical attributes include invariance to P-SMILES shifts/repeats and the ability to aggregate two- and three-dimensional molecular information via cross-modal fusion. Future developments are anticipated in the domains of higher-order subgraph priors (e.g., $v_0$4-WL), end-to-end 3D conformation learning, and extension of the framework to more complex polymer architectures such as copolymers and cross-linked or dynamic systems (Wang et al., 27 Jul 2025).