KANN-MFI

Predict Mitochondrial Fitness from single-cell RNA-seq using Kolmogorov-Arnold Neural Networks

What it does

You give it: A Seurat object with normalized counts
It gives you: Mitochondrial Function Index (MFI) for every cell

The pipeline:

1. Auto-calculates metabolic scores (Mitophagy, OXPHOS, TCA, Glycolysis, UPRmt, mt%)
2. Computes MFI from those scores
3. Trains a neural network to predict MFI from ~200 mitochondrial genes
4. Finds important genes that drive mitochondrial quality

Quick Start (Complete Noob)

Install (one time)

install.packages(c("Seurat", "qs", "dplyr", "torch", "ggplot2"))
torch::install_torch()  # Downloads PyTorch (~500MB)

Run (every time)

library(Seurat)
source("kann_mfi.R")

# Load your data
seu <- readRDS("my_seurat_object.rds")

# Step 1: Calculate MFI (auto-calculates everything!)
seu <- calculate_MFI(seu)

# Step 2: Train neural network
fit <- train_kann(seu, epochs = 300)

# Step 3: Predict for all cells
seu <- predict_kann(seu, fit)

# Step 4: Plot
plot_predictions(seu)

# Step 5: Find important genes
imp <- perm_importance(seu, fit)
plot_importance(imp)

# Step 6: Save everything
save_kann(fit, dir = "my_model")
saveRDS(seu, "seurat_with_MFI.rds")

That's it! Check out example.R for a complete annotated walkthrough.

---

Requirements

What you need:

• A Seurat object with normalized counts (run NormalizeData() if you haven't)
• That's it!

What gets calculated automatically:

• Mitophagy score (27 genes)
• OXPHOS score (Complex I-V genes)
• TCA cycle score
• Glycolysis score
• UPRmt score (15 genes)
• Mitochondrial % (from MT- genes)
• MFI (combines all of the above)

All you need is raw RNA counts! 🎉

---

What is MFI?

Mitochondrial Function Index = a single number that summarizes mitochondrial health

MFI = (Mitophagy + OXPHOS + UPRmt) - penalties

Higher MFI = healthier mitochondria

Penalties for:

• High mitochondrial DNA %
• OXPHOS/Glycolysis imbalance
• Excess TCA activity

---

🔬 Output

New columns in your Seurat object:

• Mitophagy - mitophagy activity
• OXPHOS1 - oxidative phosphorylation
• TCA1 - TCA cycle activity
• Glyco1 - glycolysis
• UPRmtScore - mitochondrial unfolded protein response
• percent.mt - mitochondrial fraction
• MFI_v1 - calculated mitochondrial quality index ⭐
• MFI_pred - KANN prediction of MFI ⭐

Files created:

• pred_vs_obs.pdf - How well the model works
• importance.pdf - Top genes that determine MFI
• importance_all_genes.csv - Complete gene rankings
• my_model/ - Saved model (reusable!)

---

Examples

Basic usage

source("kann_mfi.R")
seu <- readRDS("data.rds")

# Everything in 4 lines:
seu <- calculate_MFI(seu)
fit <- train_kann(seu, epochs = 300)
seu <- predict_kann(seu, fit)
plot_predictions(seu)

Reuse saved model

source("kann_mfi.R")

# Load model
fit <- load_kann("my_model")

# Apply to new data
new_seu <- readRDS("new_data.rds")
new_seu <- calculate_MFI(new_seu)
new_seu <- predict_kann(new_seu, fit)

Compare cell types

# After running the pipeline
library(ggplot2)

# MFI by cell type
ggplot(seu@meta.data, aes(x = cell_type, y = MFI_v1)) +
  geom_boxplot() +
  theme_classic()

# Top genes for each cell type
for (ct in unique(seu$cell_type)) {
  seu_sub <- subset(seu, cell_type == ct)
  imp <- perm_importance(seu_sub, fit, repeats = 3)
  print(head(imp, 10))
}

---

Understanding the Results

Good performance:

• R² > 0.7 - Model captures most MFI variation
• RMSE < 0.5 - Predictions are close to observed

Importance scores (ΔMSE):

• > 0.01 - Gene strongly affects MFI
• 0.001-0.01 - Moderate effect
• < 0.001 - Weak effect

Top genes usually include:

• Mitophagy receptors (BNIP3, FUNDC1, PINK1, PRKN)
• OXPHOS genes (Complex I, Complex V)
• Quality control (LONP1, CLPP, HSPD1)

---

Files in this repo

kann-mfi/
├── kann_mfi.R          # 🔴 Main file - source this!
├── example.R           # Complete working example
├── README.md           # This file
├── LICENSE             # MIT License
└── logo.png            # Hex sticker

Just download kann_mfi.R and you're good to go!

---

FAQ

**Q: I'm so lost what do I do?
A: Create an issue on the issue tab :)

Q: I only have a count matrix, not a Seurat object
A:

seu <- CreateSeuratObject(counts = your_matrix)
seu <- NormalizeData(seu)
# Now use the pipeline!

Q: Some genes are missing from my dataset
A: No problem! The pipeline uses whatever genes are available and tells you the coverage %.

Q: Can I use my own gene sets?
A: Yes! Pass your own genes to train_kann():

my_genes <- c("PINK1", "PRKN", "MFN1", "MFN2", ...)
fit <- train_kann(seu, features = my_genes)

Q: How long does training take?
A: ~2-5 minutes for 10,000 cells on a laptop. Permutation importance takes longer (10-30 min).

Q: Can I use this on mouse data?
A: Yes! Gene names should be in the same format as your Seurat object (e.g., Ppargc1a for mouse).

Q: What if I already have metabolic scores?
A: The pipeline checks for existing scores and skips them. Use force_recalc=TRUE to recalculate:

seu <- calculate_MFI(seu, force_recalc = TRUE)

---

Citation

@software{kann_mfi_2025,
  author = {<Fahd Qadir aka Dragonmasterx87>},
  title = {KANN-MFI: Kolmogorov-Arnold Networks for Mitochondrial Fitness Prediction},
  year = {2025},
  url = {https://github.com/FMJLabTulane/kann-mfi}
}

---

📄 License

MIT License - Free to use, modify, and share!

---

Help

• Issues? Open an issue on GitHub
• Questions? Check example.R for detailed walkthrough
• Email: mqadir@tulane.edu

---

Acknowledgments

Based on:

• Kolmogorov-Arnold representation theorem
• Liu et al. (2024) "KAN: Kolmogorov-Arnold Networks" arXiv:2404.19756
• Seurat: Hao et al. (2021) Cell