---
license: cc-by-4.0
tags:
  - biology
  - gene-regulatory-network
  - foundation-model
  - cross-species-transfer
  - plant
  - Nicotiana-benthamiana
---

# nbgrn-transfer — cross-species gene-regulatory-network transfer to *Nicotiana benthamiana*

Models, data, and inferred networks for the paper
**"Conditioning a gene-regulatory foundation model on transcription-factor sequence enables ortholog-free cross-species transfer: the *Nicotiana benthamiana* network from soybean and Arabidopsis."**

A masked-expression gene-regulatory foundation model (**GRN-FM**) with a single explicit
TF×target weight tensor **W** is trained in a data-rich source species (soybean or Arabidopsis)
and transferred **zero-shot** to data-poor *N. benthamiana* — either through reciprocal-best-hit
orthologs, or, with **no ortholog lookup for the regulators**, by a **sequence-conditioned** model
whose TF rows are generated from ESM-2 protein embeddings (`W = φ(ESM(TF))·Vᵀ`). Transfer is scored
by relatedness-aware **leave-study-out** prediction of tobacco's own held-out-study expression
(no regulatory gold standard is used).

## Contents

### `models/`
| file | description |
|---|---|
| `arabidopsis_grn_fm_W.npz` | Arabidopsis GRN-FM explicit weight tensor W (source) |
| `arabidopsis_grn_fm_seq_d256.pt` | Arabidopsis **sequence-conditioned** GRN-FM (φ + V, rank d=256; ESM-2 `esm2_t12_35M_UR50D`) |
| `soybean_grn_fm_W.npz` | Soybean GRN-FM weight tensor W (source; same model as the companion `dirmy/soygrn`) |
| `tobacco_within_grn_fm_W.npz` | Within-species tobacco GRN-FM W (upper-bound reference) |

### `data/`
| file | description |
|---|---|
| `tobacco_gene_tpm.npz` | *N. benthamiana* compendium: 56,738 genes × 1,540 samples (NbT2T reference) |
| `tobacco_sample_meta.csv` | per-sample SRA Study/BioProject metadata (153 studies) |
| `tobacco_protein_esm2.npz` | ESM-2 mean-pooled embeddings of tobacco proteins |
| `tobacco_tf_genes.txt` | 1,203 ortholog-defined candidate tobacco TFs |
| `tobacco_to_arabidopsis_rbh.tsv` | tobacco↔Arabidopsis reciprocal-best-hit orthologs (12,773) |
| `tobacco_to_soybean_rbh.tsv` | tobacco↔soybean reciprocal-best-hit orthologs (13,981) |

### `networks/`
Inferred TF→target edge lists (columns: `TF`, `target`, `score`):
`tobacco_within_clr.tsv`, `tobacco_within_grnboost2.tsv`, `tobacco_within_fm.tsv`
(within-species), and `tobacco_transfer_from_arabidopsis.tsv` (zero-shot Arabidopsis→tobacco transfer).

## Reproduce

Code: **https://github.com/k821209/nbgrn-transfer**
```
python src/grn_transfer_eval.py     # ortholog transfer (soybean/Arabidopsis → tobacco)
python src/grn_transfer_seq.py      # sequence-conditioned, ortholog-free transfer
python src/grn_fewshot.py           # within-species data-efficiency curve
```

## Companion

Within-species soybean GRN benchmark: **`dirmy/soygrn`** (DOI 10.57967/hf/9365).

## Citation

Kang, Y. *Conditioning a gene-regulatory foundation model on transcription-factor sequence
enables ortholog-free cross-species transfer.* (2026). Data & models: this repository.

License: CC-BY-4.0.