Fine-tuning DeepSeek-OCR-2 for Molecular Structure Recognition.

Area of Science:

• Computational Chemistry
• Artificial Intelligence
• Cheminformatics

Background:

• Optical Chemical Structure Recognition (OCSR) is vital for digitizing chemical information from literature.
• Existing Vision-Language Models struggle with direct OCSR fine-tuning due to training instabilities.

Purpose of the Study:

• To adapt DeepSeek-OCR-2 for robust OCSR via image-conditioned SMILES generation.
• To develop a stable fine-tuning strategy for improved OCSR performance.
• To evaluate the model's effectiveness on diverse chemical structure datasets.

Main Methods:

• Formulated OCSR as image-conditioned SMILES generation.
• Implemented a two-stage progressive supervised fine-tuning strategy (LoRA to selective full-parameter tuning).
• Trained on a combined corpus of synthetic (PubChem) and real-world (USPTO-MOL) chemical diagrams.

Main Results:

• The fine-tuned model, MolSeek-OCR, achieved accuracy comparable to top image-to-sequence models for OCSR.
• Performance remained inferior to state-of-the-art image-to-graph models.
• Post-training methods (reinforcement learning, data curation) did not enhance sequence-level fidelity for exact SMILES matching.

Conclusions:

• The proposed fine-tuning strategy enhances OCSR capabilities for image-to-sequence tasks.
• Further research is needed to bridge the performance gap with image-to-graph approaches.
• Strict sequence-level fidelity is a critical challenge for exact SMILES generation in OCSR.