LaCo

L2aCo is the official code repository for the paper "Expanding Knowledge Boundaries via LLM-Grounded Alignment for Drug Combination Recommendation", accepted at KDD 2026 AI4Science track.

The project explores how to use large language model (LLM) grounded embeddings to enhance drug and cell line representations, improving drug combination prediction for long-tail and cold-start cell lines.

Key Contributions

• Introduces L2aCo, a model-agnostic knowledge alignment framework.
• Uses LLM-inferred semantic profiles to augment drug and cell line representations with knowledge beyond experimental measurements.
• Aligns semantic representations with conventional molecular and cellular features at the representation level.
• Demonstrates improved generalization on long-tail and novel cell lines.

Repository Structure

• graph_based/ — graph-based drug combination prediction implementation with GNN models and optional LLM embeddings.
• feature_based/ — feature-based methods, currently organized into three folders:
  • DeepDDS/
  • DFFNDDS/
  • SynergyX/ These methods are placeholders for future code, data, and READMEs.
• Combinational_Drug_Recommendation.pdf — the full paper PDF.

Graph-Based Module

The graph_based/ module is the most complete implementation in this repository and includes:

• dcb_main.py — main training and evaluation entry point.
• dataloader.py — data loading and batching logic.
• model.py — GNN model definitions such as GCN, HeteroGAT, and KGNN.
• layers.py — regularization and layer utilities.
• loss_util.py — loss functions and evaluation metrics.
• utils.py — model saving, logging, and evaluation helpers.
• generate_embed_bge.py — helper script for generating BGE / LLM embeddings.
• datasets/ — dataset input files and processed dataset artifacts.
• ckpts/ — checkpoints output root.
• logs/ — runtime logs.
• wandb/ — optional W&B output.

See graph_based/README.md for more details.

Feature-Based Module

The feature_based/ module currently contains three feature-based method folders:

• DeepDDS
• DFFNDDS
• SynergyX

These folders are intended to support comparison and validation of LLM-augmented feature-based drug combination prediction. Code, data, and per-method documentation will be added later.

Environment and Dependencies

Recommended Python version: 3.8+.

Install dependencies for the graph-based module:

pip install -r graph_based/requirements.txt

The current graph_based/requirements.txt includes:

• torch==2.5.1+cu121
• dgl==2.4.0+cu121
• numpy==1.26.3
• scikit-learn==1.7.2
• tqdm==4.67.3
• wandb==0.23.0

Choose a compatible torch and dgl installation for your CUDA environment.

Data Preparation

The graph-based experiments currently require the following dataset files:

• datasets/kg/entities.dict
• datasets/kg/relations.dict
• datasets/kg/train_new.tsv

The dataset is available on Hugging Face:

• https://huggingface.co/datasets/Matthewmtf/LaCo

Download the dataset and place it under the repository root so the directory structure remains consistent.

Quick Start

Run a training example in graph_based/:

cd graph_based
python dcb_main.py --gpu 0 --model SAGE --dataset drugcombdb --debug --aug --llm gpt-4o-mini --setting S1

Common models:

• GCN
• HGAT
• KGNN
• SAGE

Common LLM options:

• gpt-4o-mini
• gpt-3.5-turbo
• gpt-5
• llama3-8b-chat
• Baichuan2-chat
• llama
• qwen

Design Philosophy

L$^2$aCo is designed to:

1. Build base representations for drugs and cell lines from experimental features.
2. Expand representation boundaries with LLM-derived semantic profiles.
3. Fuse experimental and semantic knowledge through representation-level alignment to improve performance on long-tail and rare cell lines.

This plugin-style approach enables L2aCo to enhance existing drug combination predictors without redesigning their core architectures.

Citation

If you use this repository or the associated ideas, please cite:

@inproceedings{ma2026expanding,
  title={{Expanding Knowledge Boundaries via LLM-Grounded Alignment for Drug Combination Recommendation}},
  author={Ma, Tengfei and He, Yuqin and Ren, Zhonghao and Song, Bosheng and Li, Qian and Zeng, Xiangxiang},
  booktitle={Proceedings of the 2026 ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)},
  year={2026},
  series={AI4Science Track},
  location={Jeju Island, Republic of Korea}
}