Dual-Path Knowledge-Augmented Contrastive Alignment Network for Spatially Resolved Transcriptomics - AAAI 2026 Oral

Thank you for your attention. This is the official codebase for DKAN.

Environment

python=3.11.11
pytorch-lightning==2.4.0
torch==2.4.1+cu121
torchvision==0.19.1+cu121

More details of our environments are provided in dkan.yaml.

Data Preparation

All data used in this study is available at Huggingface. Please feel free to reach me if you have any trouble in downloading the data. After downloading all data, you will obtain below data directory:

	./data/
	├── her2st
    │ 		├── gt_features_224
    │ 		├── n_features_5_224
    │ 		├── ST-cnts
    │ 		├── ST-imgs
    │ 		├── ST-spotfiles
    │ 		├── ST-text
    				└── her2st_ncbi_biobert.pt
    │ 		└── ...
	├── skin
    │ 		└── ...
	└── stnet
    		└── ...

Image Feature Extraction

If you wish to use other foundation model alternatives, such as Conch, take HER2+ dataset as an example, the command for feature extraction can be found below:

python extract_features.py --config her2st/her2st --test_mode internal --extract_mode target --encoder conch
python extract_features.py --config her2st/her2st --test_mode internal --extract_mode neighbor --encoder conch

Gene Feature Extraction

Our gene feature extraction have multiple steps:

1. Gene summary retrieval from external public gene databases, such as NCBI and GO-Term
2. Gene summary refinement by LLM, such as GPT-4o, DeepSeek-V3, DeepSeek-R1, and Llama-2
3. Gene embedding, such as Conch, Plip, BioBERT and BioGPT

To provide more gene representation alternatives and ensure reproducibility, more details about using other free accessed LLMs can be found at gene_embedding_notes. All codes, gene summaries and gene features are also released in this repository.

Training and Evaluation

Before starting training, please make sure the directories in ./config/*/*.yaml are updated accordingly to your work environments.

# Training
python main.py --config_name skin/skin --mode cv --gpu 0
python main.py --config_name her2st/her2st --mode cv --gpu 0
python main.py --config_name stnet/stnet --mode cv --gpu 0

# Evaluation
python test.py --dataset skin --model_name your_name --gpu 0

Reproducing Reported Results

Please see ./evaluation/metric.ipynb to reproduce the results reported in our paper.

TODO

• Release our ST data on Hugging Face

• Release our gene summary and embeddings

• Release our prompts we used in this study

• Release raw data from NCBI database

• Release code for training and evaluation

• Release code for reproducing our results

• Release code for image features with foundation models

• Release code for gene summary from external gene databases

• Release code for gene summary refinement via LLMs

• Release code for gene feature extraction with text encoders

• Release paper on Arxiv

Acknowledgement

Our study builds upon previous studies: TRIPLEX, HisToGene. Many thanks for their opensource and the contributions to this community.

Citation

Please kindly cite our paper if our work could be helpful for your study. Thank you.

@inproceedings{zhang2025dkan,
  title={Dual-Path Knowledge-Augmented Contrastive Alignment Network for Spatially Resolved Transcriptomics},
  author={Zhang, Wei and Chu, Jiajun and Liu, Xinci and Tong, Chen and Li, Xinyue},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={40},
  number={15},
  pages={12807--12815},
  year={2026}
}