Authors:

Eric Xu

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Predicting how humans perceive smells has mainly focused on monomolecular odors, but inferring the perception of natural smells, which are complex combinations of molecules, remains a challenge. Gaining a stronger understanding of the extent to which different representations of olfactory information agree, including chemical compositions, human ratings for perceptual qualities, and textual descriptions, could improve performance of models on tasks that involve perception of natural smells. This project aims to understand the mapping from complex mixtures to perception through the analysis of different odorant representations, obtained from a natural smell dataset and the outputs of text embedding models and Large Language Models (LLMs). We used the Volatile Compounds in Food Dataset, which contains binary encodings for the presence of over 8000 chemical compounds within over 1300 foods. Due to the scarcity of human-rated perception datasets for natural smells, we used an LLM to obtain proxies for human odor perceptions. Using representational similarity analysis, we quantified the strength of the relationship between different odorant representations through correlations between pairwise distances of samples. As a control, we used the Gemini text-embedding-004 model to obtain embeddings for raw-text names of natural smells. Using the same model, we also obtained embeddings of LLM- generated perceptual descriptions for each natural smell. Through representational similarity analysis, we computed pairwise distance matrices for each representation using well-suited distance metrics. Then, we calculated the Pearson correlation coefficient of pairwise distances between matching elements in these matrices. Preliminary results using the Jaccard distance metric for chemical composition representations and Euclidean distance metric for raw-text embeddings yield correlations from 0.21 to 0.43 with p- values from 1.2×10−19 to 0.023. We are currently investigating the relationship between odorant chemical composition and LLM- generated perceptual representations using other distance metrics and improving prompting techniques to obtain more informative odor perceptual descriptions. Automated Chemical Reasoning Agent for Drug Design Gavin Ye, Nada Amin Harvard College | Cabot House | Computer Science | 2028 Recent developments in frontier AI language models have observed the success of language model agents in various domains, such as code generation. While a few recent studies did demonstrate the possibility of automated research with agentic LLMs, their research scope is mainly limited to computer science and math. Here, we introduce Ethereal: an agentic reasoning model for automated molecular and drug designing tasks. By treating molecules as an additional modality, important chemical information that was previously inaccessible to traditional LLMs, such as their 3D structure, can be provided through external graph and stereochemistry encoders. With the addition of downstream molecular validation tools such as docking simulations and external evaluation models, Ethereal can be applied to downstream drug designing tasks against a specific drug target, after several rounds of supervised and reinforcement learning training. As the first chemical reasoning model for drug design, Ethereal marks a crucial step towards fully automated drug development and downstream natural science discovery. Harvard Summer Undergraduate Research Village Kempner Research in Artificial and Natural Intelligence for Undergraduates with Mentorship

Source:

Harvard / Computer Science / 2028

Topics:

No topics listed

Co-authors:

Eric Xu