Population-aware Generative Modeling of Lignin Ensemble
4 affiliations
- 1Laboratory of Organic Electronics, Department of Science and Technology
- 2Linköping University
- 3AI4S AB
- 4Wallenberg Wood Science Center
Abstract
Lignin is a structurally heterogeneous biopolymer whose monomer composition, interunit linkage types, branching, and molecular-weight distribution vary across plant sources and among individual polymer chains. This heterogeneity is essential for understanding lignin self-assembly and morphology, but it is difficult to represent in molecular dynamics simulations because existing rule-based and stochastic generators are generally limited to relatively small structures or populations. Here, we present a machine learning workflow for generating statistically controlled, polydisperse lignin ensembles and converting them into simulation-ready molecular systems. We trained autoregressive molecular language models on 1.16M lignin structures generated with LigninGraphs and represented as canonical SMILES to serve as a generative prior. This generative prior provides a chemically informed distribution for efficient sampling of lignin structures during reinforcement learning. We then introduce a population-aware reinforcement learning scheme that conditions the generative model toward feedstock-specific ensemble properties. Rather than optimizing each generated molecule independently toward the target properties, the generated molecules are evaluated collectively as a population and rewarded according to their contribution to reproducing the target ensemble statistics, including monomer fractions, linkage distributions, branching, number- and weight-average molecular weights, and polydispersity. Molecular-fingerprint and dimensionality-reduction analyses showed that the generated structures remained within the chemical manifold learned from the training data while producing structures not duplicated in the original library. To connect molecular generation directly with simulation, we further extended the structure-construction workflow to automate three-dimensional conversion, residue and linkage-patch assignment, force-field topology generation, and solvent packing. The resulting framework provides an end-to-end route from experimentally informed population statistics to diverse, polydisperse lignin systems suitable for large-scale molecular dynamics simulations. By treating the molecular population, rather than an individual molecular structure, as the target of generative modeling, this approach enables molecular heterogeneity to be incorporated explicitly into atomistic simulations and enables systematic investigation of how molecular heterogeneity governs lignin aggregation, morphology, and self-assembly.
Related
- ChemEidos: Synthesis-Aware Multimodal Foundation Models for Domain-Specific Generative Design
- Surface-Aware Generative Design of Selective Inhibitors for Protein–Membrane Adsorption
- Generative design of high-fidelity microstructures using physics-aware machine learning.
- From Top to Bottom: Manufacturing Process‐Context Aware Resolution of Energy Device Electrodes Through a 3D Diffusion Generative Model
- PCAM: proteochemometric assay-aware modeling for robust bioactivity estimation and uncertainty quantification
BibTeX
@misc{camposano2026populationaware,
title = {{Population-aware Generative Modeling of Lignin Ensemble}},
author = {Anthony Val Camposano and Hampus Gummesson Svensson and Karl Löwenmark and Aleksandar Y. Mehandzhiyski and Igor Zozoulenko},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008225/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008225/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Population-aware Generative Modeling of Lignin Ensemble
AU - Anthony Val Camposano
AU - Hampus Gummesson Svensson
AU - Karl Löwenmark
AU - Aleksandar Y. Mehandzhiyski
AU - Igor Zozoulenko
PY - 2026
DA - 2026/09/02
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008225/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008225/v1
AB - Lignin is a structurally heterogeneous biopolymer whose monomer composition, interunit linkage types, branching, and molecular-weight distribution vary across plant sources and among individual polymer chains. This heterogeneity is essential for understanding lignin self-assembly and morphology, but it is difficult to represent in molecular dynamics simulations because existing rule-based and stochastic generators are generally limited to relatively small structures or populations. Here, we present a machine learning workflow for generating statistically controlled, polydisperse lignin ensembles and converting them into simulation-ready molecular systems. We trained autoregressive molecular language models on 1.16M lignin structures generated with LigninGraphs and represented as canonical SMILES to serve as a generative prior. This generative prior provides a chemically informed distribution for efficient sampling of lignin structures during reinforcement learning. We then introduce a population-aware reinforcement learning scheme that conditions the generative model toward feedstock-specific ensemble properties. Rather than optimizing each generated molecule independently toward the target properties, the generated molecules are evaluated collectively as a population and rewarded according to their contribution to reproducing the target ensemble statistics, including monomer fractions, linkage distributions, branching, number- and weight-average molecular weights, and polydispersity. Molecular-fingerprint and dimensionality-reduction analyses showed that the generated structures remained within the chemical manifold learned from the training data while producing structures not duplicated in the original library. To connect molecular generation directly with simulation, we further extended the structure-construction workflow to automate three-dimensional conversion, residue and linkage-patch assignment, force-field topology generation, and solvent packing. The resulting framework provides an end-to-end route from experimentally informed population statistics to diverse, polydisperse lignin systems suitable for large-scale molecular dynamics simulations. By treating the molecular population, rather than an individual molecular structure, as the target of generative modeling, this approach enables molecular heterogeneity to be incorporated explicitly into atomistic simulations and enables systematic investigation of how molecular heterogeneity governs lignin aggregation, morphology, and self-assembly.
ER -