Scholarlink

Drug-induced liver injury (DILI) is a leading cause of late-stage clinical attrition due to its unpredictable onset and potential for severe hepatotoxicity. Therefore, it is necessary to predict hepatotoxic liabilities early and cheaply. The standard approach for this problem is adopting machine learning algorithms traditionally or novel deep learning methods like recurrent neural networks and large language model embeddings. However, there is still a lack of comprehensive assessments of how well these approaches combine. In this study, we benchmark four machine learning models on the 1278-compound DILIRank dataset: (i) extreme gradient boosting (XGBoost); (ii) an ExtraTrees classifier trained on molecular fingerprints reduced via principal component analysis (PCA); (iii) a graph convolutional network (GCN) that encodes atom “bond topology; and (iv) an attention-based recurrent neural network (ARNN) applied to SMILES strings. Large language model (LLM) embeddings enrich both the tabular and graph inputs. The best single learner” ExtraTrees + PCA—achieves an AUC of 0.917. A weight-optimized soft-voting ensemble that fuses fingerprint, graph, LLM, and ARNN outputs further improves performance to an AUC of 0.921 while balancing sensitivity (0.72) and specificity (0.92). These results demonstrate that integrating orthogonal molecular representations yields more reliable hepatotoxicity predictions and offers a practical route for early DILI screening in drug-development pipelines.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DeepGPT-DILI: Integrating Graph Convolutional Networks and Large Language Model Embeddings for Accurate Drug-Induced Liver Injury Prediction

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Minh Huu Nhat Le,
  • Uyen Khoi Minh Huynh,
  • Hong Xuan Ong,
  • Phat K. Huynh,
  • Minh-Toan Dinh,
  • Han Hong Huynh,
  • Hien Quang Kha,
  • Phat Ky Nguyen,
  • Xuan-Loc Huynh,
  • An Thuy Vo,
  • Thanh-Minh Nguyen,
  • Thanh-Huy Nguyen,
  • Quan Nguyen,
  • Nguyen Quoc Khanh Le

摘要

Drug-induced liver injury (DILI) is a leading cause of late-stage clinical attrition due to its unpredictable onset and potential for severe hepatotoxicity. Therefore, it is necessary to predict hepatotoxic liabilities early and cheaply. The standard approach for this problem is adopting machine learning algorithms traditionally or novel deep learning methods like recurrent neural networks and large language model embeddings. However, there is still a lack of comprehensive assessments of how well these approaches combine. In this study, we benchmark four machine learning models on the 1278-compound DILIRank dataset: (i) extreme gradient boosting (XGBoost); (ii) an ExtraTrees classifier trained on molecular fingerprints reduced via principal component analysis (PCA); (iii) a graph convolutional network (GCN) that encodes atom “bond topology; and (iv) an attention-based recurrent neural network (ARNN) applied to SMILES strings. Large language model (LLM) embeddings enrich both the tabular and graph inputs. The best single learner” ExtraTrees + PCA—achieves an AUC of 0.917. A weight-optimized soft-voting ensemble that fuses fingerprint, graph, LLM, and ARNN outputs further improves performance to an AUC of 0.921 while balancing sensitivity (0.72) and specificity (0.92). These results demonstrate that integrating orthogonal molecular representations yields more reliable hepatotoxicity predictions and offers a practical route for early DILI screening in drug-development pipelines.