AROMMA Agent

Overview

AROMMA Agent is an LLM-powered odor prediction and recommendation system built on top of AROMMA (Aggregated Representations of Olfaction via Molecule and Mixture Alignment)

It supports the following capabilities:

• Odor prediction from a molecule or a binary mixture given SMILES strings
• Molecule recommendation for composing a target scent described in natural language

Setup

1. Create and activate the conda environment:
   • conda env create -f environment.yml
   • conda activate aromma_agent_env
2. Install and run LLaMA via Ollama
   • curl -fsSL https://ollama.com/install.sh | sh
   • ollama run llama3:8b

Download trained checkpoints

Download the trained AROMMA model checkpoint aromma_best_fold.pt and place under the aromma directory.

Running the Agent

python run.py

Required Source Code modification

Due to a compatibility issue between rmrkl and Pydantic v2, the following error may occur:

pydantic.errors.PydanticUserError: Field 'name' defined on a base class was overridden by a non-annotated attribute. All field definitions, including overrides, require a type annotation.

Edit the file <conda_env_path>/lib/python3.9/site-packages/rmrkl/executor.py as follows:

class ExceptionTool(BaseTool):
    # name = "_Exception"
    # description = "Exception tool"
    name: str = "ExceptionTool"
    description: str = "Handles exceptions gracefully."

Basic Usage Example

Task 1. Odor Prediction

Given a SMILES string (single molecule or binary mixture), the agent predicts odor descriptors using AROMMA and produces a natural-language description.

• Single molecule

  Q. How does the molecule CCC1=C(O)C(=O)CC1 smell like?
  
  Action: OdorPrediction
  Action Input: CCC1=C(O)C(=O)CC1
  Observation: ['caramellic']
  
  A. The given SMILES smells caramellic.
  

• Binary Mixture

  Q. How does the mixture blend with CC(C)COC(=O)Cc1ccccc1 and CCCCCOC(=O)Cc1ccccc1 smell like?
  
  Action: OdorPrediction
  Action Input: CC(C)COC(=O)Cc1ccccc1;CCCCOC(=O)Cc1ccccc1
  Observation: ['floral', 'fruity', 'honey', 'sweet']
  
  A. The mixture blends with a sweet, floral, and fruity scent, reminiscent of honey.
  

• Over two molecules

  Q. How does the mixture blend with CC(C)COC(=O)Cc1ccccc1, CCCCCOC(=O)Cc1ccccc1 and CCC1=C(O)C(=O)CC1 smell like?
  
  Action: OdorPrediction
  Action Input: CC(C)COC(=O)Cc1ccccc1;CCCCOC(=O)Cc1ccccc1;CCC1=C(O)C(=O)CC1
  Observation: []
  
  A. Sorry, we can only process up to 2 SMILES at a time.
  

Task 2. Molecule Recommendation

Given a natural-language description of a desired scent:

• The agent selects 2–6 odor descriptors from a fixed vocabulary of 152 descriptors.
• These descriptors are used to retrieve and recommend suitable molecular ingredients.

Q. I want to make a scent that captures the feeling of running through a grassy field.

Action: MolRecommendation
Action Input: ["fresh", "green", "herbal", "grass"]
Observation: The scent may resemble ['fresh', 'green', 'herbal', 'grass'].
Here are some ingredients you can use to create this fragrance:
- SMILES: CCC=CCCOC(=O)C=Cc1ccccc1, descriptors: green;fresh;balsamic;herbal, matched(3): ['fresh', 'green', 'herbal']
- SMILES: C=C(C)CCO, descriptors: green;fruity;sweet;herbal;fresh, matched(3): ['fresh', 'green', 'herbal']
- SMILES: CCCCCCC1OCC(C)O1, descriptors: green;fruity;fatty;herbal;fresh, matched(3): ['fresh', 'green', 'herbal']