Work

Privileged structure-based molecular fingerprints for organic electronic materials: towards intuitive machine learning interpretation

Year: 2026

Type: article

Abstract: Molecular descriptors are central to the performance and interpretability of QSPR models, yet most existing fingerprints for organic electronics lack chemical relevance or interpretability. Here, we present the Organic Electronic Fingerprint (OEFP), a structure-based representation tailored for OLED and OPV materials. OEFP was constructed from a manually curated OLED dataset and publicly available OPV and chromophore datasets to ensure structural diversity. Synthetically accessible substructures were identified using fragmentation and ring decomposition methods, which capture the conjugated π-bonds crucial for organic electronic materials, and subsequently encoded as individual bits. In case studies of OPV HOMO energy prediction, OEFP achieved up to 13.7% lower MAE and 1.6% higher R2 compared to domain-mismatched structure-based fingerprint under random splits. Under scaffold-based splits designed to assess generalization, OEFP demonstrated substantially improved generalization, reducing MAE by 50–70% and increasing R2 by over 150% relative to the same baseline and achieved performance comparable to ECFP baselines evaluated at a similar fingerprint length when using count-based representations. SHAP analysis further enabled intuitive substructure-level interpretation, while OEFP’s chemically meaningful and synthetically relevant design supports rational molecular generation. Together, these results establish OEFP as an effective and interpretable molecular representation for machine learning applications in organic electronics. Benchmarks across diverse datasets demonstrate that the Organic Electronic Fingerprint (OEFP) provides a domain-specific and interpretable molecular representation, capturing chemically meaningful motifs relevant to organic electronic materials. Compared with conventional fingerprints, OEFP consistently improves predictive accuracy and interpretability, while enabling direct substructure–property insights. These results establish OEFP as a valuable resource for advancing machine learning–based molecular property prediction and AI-assisted molecular design in organic electronics.

Source: Journal of Cheminformatics

DOI: 10.1186/s13321-026-01169-7

Authors: Tae Wook Yang, Seung Hyun Jo, Min Chul Suh

Institution: Kyung Hee University

Language: English

Topic: Machine Learning in Materials Science

Subfield: Materials Chemistry

Field: Materials Science

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $2,390

Funders: Ministry of Trade, Industry and Energy, Ministry of Science and ICT, South Korea, Korea Evaluation Institute of Industrial Technology