LANGER: Lanthanide Affinity Network via Graph and Evolutionary Representations
4 affiliations
- 1Bredesen Center
- 2University of Tennessee
- 3Department of Chemical and Biomolecular Engineering
- 4Department of Mathematics
Abstract
The transition to sustainable energy relies heavily on rare earth elements (REEs), yet their separation remains a critical industrial bottleneck due to their nearly identical physicochemical properties. The discovery of lanmodulin (LanM), a highly selective REE-binding protein, has inspired bio-based separation strategies; however, predicting binding affinity between different LanM variants and REEs remains a significant computational challenge. Here, we present LANGER, a multimodal deep learning framework for predicting REE protein binding affinity (Kd) by synchronously integrating global evolutionary context, continuous ionic chemistry, and local three-dimensional coordinate geometry. Our architecture combines ESM2 protein sequence embeddings, REE physicochemical descriptions, and geometric graph learning (GGL) features derived from topologies extracted from protein-ion complexes. Using a comprehensive dataset of 616 evaluated LanM homologs against 15 REE ions, we demonstrate the robustness of this framework across two rigorous evaluation settings. In a cluster-based validation analysis designed to test extrapolation to held-out LanM sequence clusters, LANGER achieved a root mean square error (RMSE) of 0.212 and a concordance index (CI) of 0.757. More importantly, in a rigorous evaluation protocol that completely excludes individual ions from the training phase, the fully integrated multimodal framework achieved high predictive accuracy (RMSE = 0.119, Pearson correlation coefficient r = 0.860, CI = 0.800), demonstrating its unique ability to transfer learned physicochemical trends to unseen lanthanide elements. This work establishes a scalable, computationally accurate structure recognition platform for the discovery and design of data-driven REEselective proteins.
Related
- Network-based analysis reveals genetic diversity of human-brain expressed genes with high evolutionary rates
- Comparative analysis of graph-based metabolic network representations using graph Kernels
- Using a lanthanide tagged nanobody to visualise Glial fibrillary acidic protein by secondary ion mass spectrometry
- EF-Hand Binding Order and Conformational Control of Lanthanide Ion Binding in M. extorquens Lanmodulin
- Dye-affinity functionalized magnetic nanoparticles enable direct immobilization of cellulase from cell-free microbial culture medium
BibTeX
@misc{nguyen2026langer,
title = {{LANGER: Lanthanide Affinity Network via Graph and Evolutionary Representations}},
author = {Trung Nguyen and Christopher Cotter and Seunghyun Ryu and Cong T. Trinh and Duc Duy Nguyen},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009150/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009150/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - LANGER: Lanthanide Affinity Network via Graph and Evolutionary Representations
AU - Trung Nguyen
AU - Christopher Cotter
AU - Seunghyun Ryu
AU - Cong T. Trinh
AU - Duc Duy Nguyen
PY - 2026
DA - 2026/09/21
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009150/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009150/v1
AB - The transition to sustainable energy relies heavily on rare earth elements (REEs), yet their separation remains a critical industrial bottleneck due to their nearly identical physicochemical properties. The discovery of lanmodulin (LanM), a highly selective REE-binding protein, has inspired bio-based separation strategies; however, predicting binding affinity between different LanM variants and REEs remains a significant computational challenge. Here, we present LANGER, a multimodal deep learning framework for predicting REE protein binding affinity (Kd) by synchronously integrating global evolutionary context, continuous ionic chemistry, and local three-dimensional coordinate geometry. Our architecture combines ESM2 protein sequence embeddings, REE physicochemical descriptions, and geometric graph learning (GGL) features derived from topologies extracted from protein-ion complexes. Using a comprehensive dataset of 616 evaluated LanM homologs against 15 REE ions, we demonstrate the robustness of this framework across two rigorous evaluation settings. In a cluster-based validation analysis designed to test extrapolation to held-out LanM sequence clusters, LANGER achieved a root mean square error (RMSE) of 0.212 and a concordance index (CI) of 0.757. More importantly, in a rigorous evaluation protocol that completely excludes individual ions from the training phase, the fully integrated multimodal framework achieved high predictive accuracy (RMSE = 0.119, Pearson correlation coefficient r = 0.860, CI = 0.800), demonstrating its unique ability to transfer learned physicochemical trends to unseen lanthanide elements. This work establishes a scalable, computationally accurate structure recognition platform for the discovery and design of data-driven REEselective proteins.
ER -