A Foundation Chemical Language Model for Comprehensive Fragment-Based Drug Discovery 1. FragAtlas-62M is a groundbreaking chemical language model specifically designed for fragment-based drug discovery. It is trained on the largest fragment dataset to date, comprising over 62 million molecules from the ZINC-22 database. This model achieves an impressive 99.90% chemical validity in generated fragments, making it a powerful tool for medicinal chemistry. 2. The model not only maintains a high coverage of known ZINC fragments (53.55%) but also generates 22.04% novel structures with practical relevance. This balance between rediscovery and novelty is crucial for fragment-based drug discovery, as it ensures both the reliability of known fragments and the potential for new discoveries. 3. FragAtlas-62M is built on a GPT-2 architecture with 42.7M parameters. It uses a 128-token context window and is trained using HuggingFace Transformers. The model's architecture and training methodology are optimized for fragment-level SMILES modeling, ensuring efficient and high-throughput generation capabilities. 4. The model's performance is validated across 12 molecular descriptors and three fingerprint methods. The generated fragments closely match the training distribution, with all effect sizes being less than 0.4. This indicates that the model generalizes well and maintains the key properties of the training set without systematic bias. 5. FragAtlas-62M demonstrates substantial overlap in chemotype space between novel and rediscovered molecules, as shown by t-SNE visualizations and distance analysis. The distance ratios (NR/NN and NR/RR) are consistently near 1.0 across all three fingerprint types, indicating minimal distributional shifts. 6. The model is released with training code, preprocessed data, documentation, and model weights, making it accessible for further research and practical applications. This open release lowers the barrier to entry for groups with modest computational resources and encourages rapid follow-up work and experimental validation. 7. While FragAtlas-62M is a significant advancement, it has limitations. It does not explicitly model stereochemistry, geometric relationships, or fragment-to-fragment connectivity rules. Future work should focus on integrating conditional controls, structural information, and methods for automated molecule construction to broaden its practical applications. 📜Paper: arxiv.org/abs/2509.19586 #FragAtlas62M #ChemicalLanguageModel #FragmentBasedDrugDiscovery #MedicinalChemistry #AIinDrugDiscovery #OpenSource #Research