Periodic Topological Deep Learning for Polymer Design and Discovery

Technical Summary: Periodic Topological Deep Learning for Polymer Design and Discovery

Problem Statement
Polymers are critical to applications in energy, healthcare, and materials science, yet their vast chemical space renders systematic discovery via purely experimental methods impractical. While polymer informatics has emerged to address this using machine learning (ML), existing deep learning frameworks suffer from a fundamental representational bottleneck. Most approaches encode polymers as molecular graphs of a single repeating unit, effectively treating macromolecules as small molecules. This strategy fails to capture two critical physical realities: (1) the periodic nature of polymer chains extending across repeating units, and (2) many-body interactions beyond pairwise bonds that arise from non-covalent forces and long-range spatial proximity. Furthermore, standard graph neural networks (GNNs) are restricted to pairwise interactions and often miss long-range non-covalent interactions essential for bulk properties like the glass transition temperature ($T_g$).

Methodology: Periodic-TDL
The authors introduce Periodic-TDL, a deep learning framework designed to capture both periodicity and higher-order interactions across multiple spatial scales. The methodology consists of three core components:

1. Periodic Vietoris–Rips Complex Representation:
   Instead of a single repeating unit, the framework constructs a periodic distance matrix by incorporating translated copies of the repeating unit while preserving chemical connectivity. This matrix defines a periodic Vietoris–Rips (VR) complex. Unlike standard graphs, VR complexes are simplicial complexes that encode interactions between atoms across adjacent repeating units under periodic boundary conditions. By varying the cutoff distance ($\epsilon$), the filtration captures interactions at different scales:

   • $\epsilon_1 = 2.0$ Å: Captures covalent bonds.
   • $\epsilon_2 = 3.0$ Å and $\epsilon_3 = 4.0$ Å: Capture non-covalent interactions and many-body structures.
     This representation is invariant to the arbitrary choice of the repeating unit's boundaries.

2. Hierarchical Simplicial Message Passing (HSMP) Encoder:
   To learn from the periodic VR complex, the authors developed the HSMP encoder. It operates on a nested sequence of simplicial complexes and propagates information in a chemically motivated direction:

   • Multi-head Simplicial Message Passing: Within each filtration level, simplex features are updated via upper-adjacency (coface-mediated) message passing, allowing information flow between simplices of varying dimensions (e.g., from triangles to edges).
   • Cross-Scale Refinement: Information flows from coarser scales (larger $\epsilon$, capturing long-range interactions) to finer scales (smaller $\epsilon$, capturing covalent bonds). A gated mechanism modulates fine-scale features using coarse-scale representations, ensuring that atom and bond embeddings at the covalent scale incorporate geometry-enhanced signals from higher-order topological domains.
   • Curvature-Based Featurization: To equip higher-order simplices with meaningful features, the framework utilizes Forman's discretization of Ricci curvature. This provides a principled initial signal for simplices of any dimension, generalizing naturally to capture local density and geometric bottlenecks.

3. Pretraining and Fine-tuning Protocol:
   Given the scarcity of labeled polymer data, Periodic-TDL employs a two-stage training strategy:

   • Pretraining: The HSMP encoder is pretrained on the PI1M dataset (1 million unlabelled polymers) using self-supervised tasks adapted from the GROVER framework: atom context prediction, bond context prediction, and functional group (FG) prediction.
   • Fine-tuning: The pretrained encoder is fine-tuned on nine downstream regression tasks spanning electronic, optical, physical, and thermal properties. The polymer-level representation is obtained via global mean pooling of atom embeddings at the finest filtration level.

Key Results
The performance of Periodic-TDL was evaluated across nine polymer property prediction tasks and validated against experimental trends:

• Benchmark Performance: Periodic-TDL achieved state-of-the-art performance, attaining the lowest Root Mean Squared Error (RMSE) on eight of nine tasks and the highest $R^2$ on all nine. Notably, for the $T_g$ dataset, the model achieved an RMSE of $\sim35$ °C against a dataset standard deviation of 110 °C, representing a 69% reduction in error relative to the data spread. It consistently outperformed baselines including fingerprint-based models (Morgan NN), sequence-based Transformers (polyBERT, TransPolymer), and graph-based models (polyGNN, MolCLR).
• Chemical Credibility and Trend Validation: The authors quantitatively validated the model's ability to capture physical effects by analyzing a computationally synthesized dataset of 48,208 systematically substituted acrylate and acrylamide polymers.
  • Ester-to-Amide Substitution: The model predicted a mean $T_g$ increase of $\sim55$ °C for ester-to-amide substitutions.
  • $\alpha$-Methylation: The model predicted a mean $T_g$ increase of $\sim14$ °C for backbone $\alpha$-methylation.
  • These predictions were statistically significant ($p < 0.001$) and consistent across matched polymer pairs.
• Experimental Verification: To verify these trends, the authors synthesized three novel polymers (previously unreported in literature) and compared their experimental $T_g$ values against model predictions. The model's predictions agreed with experimental measurements within 4 °C for two of the three polymers. Furthermore, the model correctly predicted the directional trends (sign of $\Delta T_g$) for all six tested polymer pairs (including three synthesized and three from literature), confirming that the model captures underlying physical effects rather than merely optimizing for benchmark metrics.

Significance and Claims
The paper claims that Periodic-TDL represents a significant advancement in polymer informatics by addressing the fundamental limitations of treating polymers as finite molecular fragments. Its primary contributions are:

1. Novel Representation: It is the first framework to incorporate the periodic nature of polymers into a topological deep learning architecture, explicitly encoding interactions across repeating units via periodic Vietoris–Rips complexes.
2. Many-Body Interaction Capture: By utilizing simplicial complexes and hierarchical message passing, the model explicitly captures many-body interactions and long-range non-covalent forces that are inaccessible to standard pairwise graph representations.
3. Physical Interpretability: The framework demonstrates that capturing higher-order topological features and periodicity allows the model to learn chemically meaningful structure-property relationships. The successful prediction of $T_g$ trends for ester-to-amide substitution and $\alpha$-methylation, validated by new experimental synthesis, suggests the model captures the mechanistic origins of these properties (e.g., hydrogen bonding and steric hindrance) rather than just statistical correlations.
4. Scalability and Compatibility: The architecture is designed to be compatible with established self-supervised pretraining pipelines for graph encoders, facilitating transfer learning in data-scarce domains.

The authors acknowledge current limitations, noting that the framework is restricted to linear homopolymers and does not yet handle copolymers, crosslinking, or interactions between distant repeat units beyond the immediate periodic boundary. However, the work establishes a new paradigm for representing macromolecular structure in machine learning.