Polymer property prediction drops chemistry-rule fragmentation: geometry-based fragments at a chosen radius, with roughly 5–10% gains on several properties
Synopsis
Polymer property prediction has long relied on molecular fragments cut by chemical rules, which may suit unfamiliar structures poorly and offer limited control over fragment size. A geometry-based fragmentation method casts fragmentation as a radius-constrained center-selection problem, where the number of fragments follows from a given radius rather than being fixed in advance, and changing the radius yields fragments at different scales. Feeding these multiscale fragments into a multimodal model that combines p-SMILES with 3D conformations lowers prediction error by roughly 5–10% relative to the baseline on several properties of a polymer benchmark, though the 3D modality's contribution varies by property.
Interpretation
Molecular fragments can be partitioned using only the geometric positions of atoms in three dimensions, without invoking predefined chemical bond-breaking rules, and the spatial scale of the fragments is set directly by a user-chosen radius. Earlier fragmentation methods such as BRICS identify chemically meaningful bonds through predefined decomposition rules, giving chemically interpretable fragments, but their applicability is limited for previously unseen, structurally diverse or chemically unfamiliar molecular and material systems, and they offer limited control over fragment size. For the same monomer, BRICS produces relatively small fragments as a consequence of rule-based decomposition, whereas geometry-based fragmentation generates fragments whose spatial extent is governed by the prescribed radius, so varying the radius systematically yields fragments at different characteristic scales.
Fragmentation is formulated as a radius-constrained optimization: representative atoms are chosen as centers so that every atom lies within a given radius of some center, with the number of centers not fixed in advance. The classical center-selection problem requires the number of centers to be specified beforehand; here a pooling radius is fixed instead, and the number of centers emerges from that radius. The solution is approximated with a greedy farthest-point strategy that repeatedly selects the atom farthest from its nearest chosen center until the covering radius falls below the pooling distance; atoms are then assigned to the nearest representative within a cutoff distance, atoms beyond every representative's cutoff form singleton fragments, and hydrogen atoms are excluded from center selection.
Multiscale fragment information enters a polymer property prediction model through a fragment attention module, in a model that uses both p-SMILES and 3D conformations generated from p-SMILES with RDKit. Sequence models based only on p-SMILES encode molecules as one-dimensional strings and may not explicitly capture the hierarchical and multiscale organization of molecular structures. On the Polymer Genome polymer benchmark, following the TransPolymer evaluation protocol with five-fold cross-validation and average validation RMSE, three fragment radii resolve local atomic composition, localized functional groups and extended topological frameworks, and the 3D encoder MACE-MP is frozen during training.
Adding 3D conformational information improves prediction accuracy for several polymer properties relative to models based exclusively on p-SMILES. Against the TransPolymer baseline, the multimodal model achieves roughly 5–10% improvement in predictive performance for several properties. The gain is property-dependent: several TransPolymer properties improve when 3D structural information is added, the remaining properties show averaged improvements within the observed fluctuations, and one property shows no clear improvement.
Perspective
The demonstrated scope is seven DFT-computed properties on the Polymer Genome polymer benchmark, with 3D conformations generated by RDKit through classical force-field optimization and evaluated by five-fold cross-validation average validation RMSE.
The 3D modality's contribution is property-dependent, with averaged improvements for some properties falling within observed fluctuations and one property showing no clear improvement; ablation differences for the fragment branch mostly fall within error bars; the attribution analysis gives importance distributions without resolving whether a functional group raises or lowers a property; and only a single conformer is supplied to the model at a time rather than an ensemble.