Grounding olfactory perception in language: Benchmarks and models for generating natural language odor descriptions

Area of Science:

• Computational chemistry
• Neuroscience
• Artificial intelligence

Background:

• Deep learning models predict odor perception from molecular structure but are limited by fixed vocabularies and scarce data.
• Evaluating free-form natural language odor descriptions lacks standardized metrics.

Purpose of the Study:

• To develop a benchmark and metrics for evaluating free-form odor descriptions.
• To create a model for generating odor descriptions from molecular structure.
• To enable odor perception prediction from neural data.

Main Methods:

• Introduced Odor Description and Inference Evaluation Understudy (ODIEU) benchmark with 10,000+ molecular descriptions.
• Developed a model-based metric using fine-tuned Sentence-BERT (SBERT) for evaluating text descriptions.
• Created CIRANO (Chemical Information Recognition and Annotation Network for Odors), a transformer-based structure-to-text model.
• Utilized an invertible SBERT model for neural-to-text predictions from mouse olfactory bulb data.

Main Results:

• Fine-tuned SBERT models reveal syntactic information in free-form odor descriptions.
• CIRANO achieves human-comparable performance in generating odor descriptions from molecular structures.
• Neural-to-text predictions align well with human descriptions.

Conclusions:

• ODIEU and CIRANO establish a standardized framework for olfactory language generation and evaluation.
• The study enhances odor classification and analysis by bridging molecular structure, neural data, and human perception.
• This work opens new avenues for understanding and predicting olfactory experiences.