MixRx

Authors: Risha Surana, Cameron Saidock, Hugo Chacon Affiliation: University of Southern California

Abstract

MixRx uses Large Language Models (LLMs) to classify drug combination interactions as Additive, Synergistic, or Antagonistic, given a multi-drug patient history. We evaluate the performance of GPT-2 and Mistral-7B-Instruct models, along with fine-tuned variations of each. Our experiments show that the fine-tuned Mistral-7B-Instruct model achieves an average accuracy of ~81.5%, including robustness to “messy” clinical input formats. These results demonstrate the feasibility of applying LLMs to biological interaction prediction and suggest a pathway toward real-time clinical decision support, particularly in emergency medicine settings where interaction complexity and time constraints are significant factors.

Overview

Most clinical drug interaction checkers evaluate only pairwise drug relationships and depend on exact drug name spelling. Real patients, however, often take 3–10 or more drugs simultaneously, and clinical notes frequently contain abbreviations, misspellings, and non-standard naming.

MixRx addresses this by:

1. Constructing multi-drug combinations where all pairwise synergy metrics are known.
2. Converting those combinations into structured LLM prompts with relevant synergy context.
3. Using an LLM to infer a combination-level classification.
4. Validating predictions through a rule-based synergy scoring algorithm.

Repository Structure

mixrx/
│
├── Models/                         # Model weights, checkpoints, and fine-tuned outputs
│   └── mistral/                    # Fine-tuned Mistral-7B-Instruct model artifacts
│
├── evaluate/                       # Evaluation scripts and metric calculations
│
├── preprocessing/                  # Data preprocessing and prompt generation
│   └── messy/                      # Input perturbation logic (spelling swaps, truncations)
│
├── .gitignore
├── README.md
│
├── embeddings.ipynb                # Embedding and representation analysis
│
├── final.csv                       # Full processed synergy dataset
├── final_reduced.csv               # Cleaned dataset for prompting
├── final_reduced_messy.csv         # Perturbed dataset for robustness testing
│
├── generate_messy.py               # Script to create spelling/formatting perturbation

Installation

Requires Python 3.9+.

pip install -r requirements.txt

If using a hosted LLM API (OpenAI, Anthropic, etc.):

export OPENAI_API_KEY=your_api_key

Workflow

1. Preprocess the Synergy Data

python preprocessing/generate_validation_data.py

2. Create LLM Prompt Data

python preprocessing/generate_model_data.py

3. Generate Messy Input Variant (optional)

python generate_messy.py

4. Run Evaluation / Model Inference

Open:

synergy.ipynb

Citation

If using MixRx in academic work, please cite:

Surana, R., Saidock, C., & Chacon, H. (2024).
MixRx
University of Southern California.