northenlab/soilfm2-graph-tower-joint-v0.1

SoilFM2 Graph Tower (Joint Model)

Graph neural network component of SoilFM2 for soil microbiome analysis.

Developed by: Trent Northen, Lawrence Berkeley National Laboratory

Performance

Task	Metric	Score
Metabolite Consumption	Accuracy	78.72%
Pathway Prediction	Accuracy	94.66%
Pathway Prediction	AUROC	0.98

Architecture

• Hidden dimension: 512
• GNN layers: 4 (Multi-head attention)
• Dual task heads: Consumption (12-class) + Pathway (569-dim binary)

Quick Start

from soilfm2 import SoilFM2

model = SoilFM2.from_pretrained(
    checkpoint_dir="YOUR_USERNAME/soilfm2-graph-tower-joint-v0.1",
    literature_model="northenlab/soilfm-qwen2.5-14b-literature-cpt"
)

response = model.predict(
    query="What metabolites boost nitrogen fixation?",
    files={"16S": "community.tsv"}
)

## Files

- `pytorch_model.pt`: Model checkpoint (4.6 GB)
- `soilfm_integrated_kg.pt`: Knowledge graph (71 MB)
- `integrated_node_mappings.pkl`: Node mappings (54 MB)
- `config.json`: Model configuration

## License

MIT License