Drug Molecular and Route Similarity-Aware Retrieval-Augmented LLM Generation with Knowledge Graph for Semantic Reasoning
Abstract
To address the inherent limitations of large language models (LLMs) in chemical synthesis—hallucinations, and low knowledge dense—this study presents a Knowledge Graph–enhanced Retrieval-Augmented Generation framework (KG-RAG-LLM) with a multi-scale calculation of molecular and route similarity. Empirical evaluations demonstrate that KG-RAG-LLM significantly mitigates hallucinations, increasing the proportion of concrete chemical information in generated responses to 39.52%–40.87%, drastically outperforming baseline LLMs and standard literature-based RAG pipelines in response completeness and accuracy. To overcome the inability of traditional 2D linear fingerprints to capture 3D spatial properties and reaction path variations, we further developed a multidimensional molecular and route similarity calculation combining SimHash, core similarity, and FCFP, as an alternative to Route CGR representations coupled with UMAP dimensionality reduction. This formulation accurately distinguishes between subtle and distinct synthetic strategies, demonstrating superior precision over traditional clustering methods for fuzzy retrieval of synthetic routes. Overall, this work provides a reliable, knowledge-dense solution for intelligent retrosynthetic Semantic Reasoning.
Related
- SimSJSAlert: A Similarity-Augmented Multi-View Learning Framework with Scaffold Alerts for Drug-Induced Stevens-Johnson Syndrome Risk Assessment
- A retrieval-augmented agent bridges virtual screening and automated synthesis
- Beyond Chemistry Knowledge: Embedding Pedagogical Content Knowledge through an AI-Augmented Retrosynthesis Program for Chemistry Education
- OptoMatKG: A Traceable Knowledge Graph Framework with Semantic Retrieval for Optoelectronic Materials Science
- Molecular function-guided unsupervised similarity-based enrichment for low-data virtual screening
BibTeX
@misc{lao2026drug,
title = {{Drug Molecular and Route Similarity-Aware Retrieval-Augmented LLM Generation with Knowledge Graph for Semantic Reasoning}},
author = {Jiaxin Lao and Liansuo Zu},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010270/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010270/v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Drug Molecular and Route Similarity-Aware Retrieval-Augmented LLM Generation with Knowledge Graph for Semantic Reasoning
AU - Jiaxin Lao
AU - Liansuo Zu
PY - 2026
DA - 2026/10/12
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010270/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010270/v1
AB - To address the inherent limitations of large language models (LLMs) in chemical synthesis—hallucinations, and low knowledge dense—this study presents a Knowledge Graph–enhanced Retrieval-Augmented Generation framework (KG-RAG-LLM) with a multi-scale calculation of molecular and route similarity. Empirical evaluations demonstrate that KG-RAG-LLM significantly mitigates hallucinations, increasing the proportion of concrete chemical information in generated responses to 39.52%–40.87%, drastically outperforming baseline LLMs and standard literature-based RAG pipelines in response completeness and accuracy. To overcome the inability of traditional 2D linear fingerprints to capture 3D spatial properties and reaction path variations, we further developed a multidimensional molecular and route similarity calculation combining SimHash, core similarity, and FCFP, as an alternative to Route CGR representations coupled with UMAP dimensionality reduction. This formulation accurately distinguishes between subtle and distinct synthetic strategies, demonstrating superior precision over traditional clustering methods for fuzzy retrieval of synthetic routes. Overall, this work provides a reliable, knowledge-dense solution for intelligent retrosynthetic Semantic Reasoning.
ER -