Saranyaks06/BioGPT-KG-: BioGPT
The exponential proliferation of biomedical literature presents both a challenge and an opportunity for drug discovery, particularly in identifying latent drug-disease associations that remain obscured within unstructured text. This study proposes a novel NLP pipeline integrating BioGPT and Galactica with dynamic Knowledge Graph (KG) reasoning. By employing domain-optimized NER, context-aware relation extraction, and graph-based inference, we demonstrate a robust methodology for automated hypothesis generation in drug repurposing. Proposed Pipeline Our framework operates in three distinct phases: Domain-Optimized NER: Utilizing BioGPT for pre
Resource information
- Format
- zip
- Original repository
- 20007183
- DOI
- 10.5281/zenodo.20007183
- Credit basis
- Creators in Zenodo record metadata
- Access
- Open files available from the original Zenodo deposit
- Last checked
- Nov 5, 2026
- License
- cc-by-4.0
- Last record update
- Nov 5, 2026
About this resource
The exponential proliferation of biomedical literature presents both a challenge and an opportunity for drug discovery, particularly in identifying latent drug-disease associations that remain obscured within unstructured text. This study proposes a novel NLP pipeline integrating BioGPT and Galactica with dynamic Knowledge Graph (KG) reasoning. By employing domain-optimized NER, context-aware relation extraction, and graph-based inference, we demonstrate a robust methodology for automated hypothesis generation in drug repurposing. Proposed Pipeline Our framework operates in three distinct phases: Domain-Optimized NER: Utilizing BioGPT for precise entity extraction (Drugs, Diseases, Proteins). Context-Aware Relation Extraction: Transformers (Galactica 1.3B) fine-tuned on curated drug-disease pairs. Graph-Based Inference: Predicting novel, latent associations using Graph Neural Networks (GNNs) and uncertainty quantification. Mathematical Formulation Relation Extraction Objective Knowledge Graph Embedding Phase 1: Data Loading & Preprocessing -Importing the PubMed/CORD-19 dataset and visualizing samples. Dataset Shape: (50000, 3) Sample Data: PMID | Year | Abstract 34021834 | 2023 | Metformin significantly reduces tumor growth in... 34055122 | 2022 | We investigate the role of AMPK activation in... 33901234 | 2023 | Alzheimers disease pathology is linked to ins... 33887711 | 2024 | A novel deep learning framework for drug disc... 33765432 | 2021 | Type 2 Diabetes mellitus is a chronic metabol... Phase 2: Dynamic Knowledge Graph Updates -Demonstrating the continuous update mechanism of the KG as new literature is processed. Phase 3: Knowledge Graph Embedding (RotatE) -Training the RotatE model using PyKEEN to learn vector representations for link prediction. Phase 4: Benchmarking & Evaluation -Comparative analysis of our Proposed Framework against PubTator and traditional ML baselines. Method | Precision | Recall | F1 PubTator (Rule-based) | 0.72 | 0.65 | 0.68 BERT-Baseline | 0.79 | 0.76 | 0.77 Proposed (BioGPT+KG) | 0.88 | 0.84 | 0.86 Phase 5: Latent Therapeutic Prediction -Top candidates for Metformin repurposing. Conclusion The proposed framework outperforms traditional PubTator baselines (Precision 0.88 vs 0.72), confirming the value of integrating Large Language Models with Knowledge Graph inference for drug repurposing. Comparsion methods Baseline 1: PubMedBERT (Small Medical LLM) We utilize the Microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext model, fine-tuned using the HuggingFace Trainer API. Before Training: Loading Microsoft/BiomedNLP-PubMedBERT... Training... Step | Training Loss | Validation Loss | Accuracy | F1-Score -------|---------------|-----------------|----------|---------- 100 | 0.6823 | 0.6501 | 0.5900 | 0.5200 500 | 0.5500 | 0.5800 | 0.7200 | 0.6900 1000 | 0.4820 | 0.4400 | 0.7680 | 0.7400 1500 | 0.4050 | 0.3600 | 0.7950 | 0.7800 2000 | 0.3210 | 0.3500 | 0.8120 | 0.8000 2500 | 0.2450 | 0.3400 | 0.8250 | 0.8300 --- Final Test Evaluation --- precision recall f1-score support No Association 0.81 0.80 0.81 1250 Association 0.83 0.84 0.83 1250 accuracy 0.82 2500 macro avg 0.82 0.82 0.82 2500 weighted avg 0.82 0.82 0.82 2500 Baseline 2: BioBERT (Manual PyTorch Loop) We fine-tune dmis-lab/biobert-v1.1 using a custom training loop to maximize control over the optimization process. ======== Epoch 1 / 5 ======== Training... Average training loss: 0.50 ======== Epoch 2 / 5 ======== Training... Average training loss: