README.md

metadata

license: mit
language:
  - en
tags:
  - biology
  - single-cell
  - scRNA-seq
  - macrophage
  - IBD
  - inflammatory-bowel-disease
  - perturbation
  - gene-knockout
  - scGPT
  - transformers
metrics:
  - pearsonr
  - roc_auc

MacroPert — IBD Macrophage Perturbation Model

A fine-tuned scGPT model for predicting macrophage polarization state and IBD disease status from single-cell RNA-seq, combined with in silico gene knockout predictions for 13 IBD target genes.

Model Description

Macrophages in inflammatory bowel disease (IBD) exist along a continuous spectrum between pro-inflammatory (M1) and anti-inflammatory (M2) states, rather than in discrete classes. MacroPert captures this continuous polarization spectrum and predicts how individual gene knockouts shift macrophage state.

The repository contains:

File	Description
scgpt_ibd_v2/best_model.pt	scGPT v2 fine-tuned weights — continuous polarization model
scgpt_ibd_v2/test_metrics.json	Test set performance metrics
scgpt_ibd_v1/best_model.pt	scGPT v1 weights — discrete 4-class baseline (archived)
scgpt_ibd_v1/test_metrics.json	v1 test metrics
results/ibd_ko_predictions_combined.json	In silico KO predictions for 8 IBD target genes

Performance

Metric	scGPT v1 (4-class)	scGPT v2 (continuous)
IBD AUROC	0.915	0.989
Polarization Pearson r	—	0.909
Classification Accuracy	0.754	—

Training Data

Fine-tuned on four IBD macrophage scRNA-seq datasets:

GEO Accession	Cells	Condition
GSE134809	13,794	IBD macrophages (primary)
GSE116222	~8,000	Ulcerative colitis macrophages
GSE182270	~6,000	IBD macrophage substates
GSE148810	~5,000	Crohn's disease macrophages

Model Architecture

• Base: scGPT pretrained transformer
• Fine-tuning objective:

  L = 0.5 × L_soft_contrastive + 0.3 × L_MSE(pol) + 0.2 × L_BCE(ibd)
  

• Polarization score: pol_score = z-score(M1_score − M2_score) using literature-based gene signatures
  • M1 genes: TNF, IL1B, IL6, CXCL10, NOS2, CD80, CD86, CCL5, CXCL9, PTGS2, IRF5, HIF1A, CXCL8, IL12A
  • M2 genes: CD163, MRC1, ARG1, IL10, TGFB1, CCL18, CD209, FOLR2, SOCS3, HMOX1, CLEC7A
• Soft contrastive loss: Gaussian kernel exp(−d²/2σ²) over polarization score distances — shapes embedding space continuously without discrete cluster boundaries
• Output heads: polarization regression + IBD binary classification

In Silico KO Predictions

KO effects for 13 IBD target genes were predicted using two methods:

Gene	Method	Key Predicted Effect
HIF1A	CellOracle (GRN)	PTGS2↓, CXCL8↓, CD74↓
IRF5	CellOracle (GRN)	ISG15↓, IFI6↓, IFIT3↓, IRF7↓
IL6	OmniPath propagation	JAK2↓, IL6ST↓, JAK1↓, IL6R↓, TYK2↓
SOCS3	OmniPath propagation	JAK2↓, STAT5A↓, STAT1↓, STAT3↓; AKT1↑
TNF	OmniPath propagation	TNFRSF1A↓, TNFRSF1B↓, PIK3CG↓, AKT1↓
TGFB1	OmniPath propagation	PIK3R1↓, RAC1↓, TGFBR1↓, GRB2↓
IL1B	OmniPath propagation	IL1R2↓, MYD88↓, STAT3↓, NR3C1↓
PTGS2	OmniPath propagation	IL4R↓, IL2RG↓, NFE2L2↓

Full predictions (top 15 up/down per gene) are in results/ibd_ko_predictions_combined.json.

Usage

import torch
import json

# Load model weights
checkpoint = torch.load("scgpt_ibd_v2/best_model.pt", map_location="cpu")

# Load KO predictions
with open("results/ibd_ko_predictions_combined.json") as f:
    ko_predictions = json.load(f)

# Example: top downregulated genes after IL6 KO
il6_ko = ko_predictions["IL6"]
print(il6_ko["top_downregulated"][:5])
# [['JAK2', -0.207], ['IL6ST', ...], ['JAK1', ...], ...]

Citation

If you use this model, please cite the underlying datasets and tools:

• scGPT: Cui et al., Nature Methods 2024
• CellOracle: Kamimoto et al., Nature 2023
• OmniPath: Türei et al., Nature Methods 2021
• GSE134809: Smillie et al., Cell 2019