Broad Obesity-1 Perturbation Prediction

CrunchDAO Hackathon Submission — Broad Obesity-1 Challenge

Predicts single-cell gene expression responses to unseen genetic perturbations using a cross-attention transformer (DeltaInterpolationNetV9) with scGPT biological embeddings and a Retrieve-and-Shift cell generation procedure.

Phase 2: See Crunch_2 for the extension to double-gene perturbations.

---

Method Overview

The model takes a per-gene feature vector (cross-perturbation expression profile + control statistics) and a scGPT biological embedding, then attends over all 122 training perturbation deltas to predict the mean expression shift for the query gene.

Key components:

• Cross-attention over training deltas — interpolates training perturbation responses rather than predicting from scratch, suited for the low-sample regime (122 perturbations)
• Enriched query and key paths — both fuse expression features with scGPT embeddings via learned gating, enabling the model to find biologically similar training perturbations even when expression profiles differ
• Auxiliary std head — predicts per-HVG gene standard deviations for distribution-accurate cell generation
• Confidence calibration — conditioned on attention entropy and embedding outlier distance, shrinks predictions toward the population mean for out-of-distribution genes
• Retrieve-and-Shift — samples cells from KNN-selected training perturbations and affine-transforms them to match predicted mean and HVG standard deviations
• Ensemble of 16 models — 8 trained on full data, 8 with leave-one-out validation, weighted by held-out Pearson correlation
• Blended proportion prediction — neural proportion head + KNN lookup on predicted mean deltas

See Method description.md for the full rationale.

---

Repository Structure

.
├── main.py                        # All model code: train(), infer(), evaluate()
├── Method description.md          # Full method and rationale write-up
├── notebook.ipynb                 # Development notebook
├── resources/
│   ├── embeddings/
│   │   ├── embedding_info.json    # Metadata for the combined embedding file
│   │   ├── gene_embedding_order.txt  # Gene names aligned to embedding matrix
│   │   └── gene_embeddings_combined.npy  # [NOT TRACKED] 512-dim scGPT gene embeddings
│   └── scGPT_human/
│       ├── args.json              # scGPT whole-human checkpoint training config
│       └── vocab.json             # scGPT gene vocabulary
└── data/                          # [NOT TRACKED] Competition data (see below)

---

Data

The model trains on the Broad Obesity-1 Challenge dataset (not included — available from CrunchDAO):

File	Description
data/obesity_challenge_1.h5ad	scRNA-seq matrix: 88,202 cells × 21,592 genes, 122 gene knockouts + NC controls
data/program_proportion.csv	Cell program annotations (pre_adipo, adipo, lipo, other) per perturbation
data/predict_perturbations.txt	List of genes to predict at inference time
data/genes_to_predict.txt	List of output genes (defaults to all if absent)
data/obesity_challenge_1_local_gtruth.h5ad	Local ground truth for evaluate() (optional)
data/program_proportion_local_gtruth.csv	Ground truth proportions for evaluate() (optional)

---

Resources

Large binary files are not tracked in git. You need:

Gene embeddings — extracted from the scGPT whole-human checkpoint:

python extract_foundation_embeddings.py \
    --gene_list data/obesity_challenge_1.h5ad \
    --out_dir resources/embeddings/

This produces resources/embeddings/gene_embeddings_combined.npy (512-dim, ~42 MB) and gene_embedding_order.txt.

Model checkpoint — produced by train():

resources/ensemble_v9.pt

---

Setup

pip install torch numpy pandas anndata scanpy scikit-learn scipy tqdm matplotlib

The scGPT vocabulary (resources/scGPT_human/vocab.json) is included. The full scGPT model weights are only needed to re-extract embeddings.

---

Usage

All entry points are in main.py:

Train

from main import train
train(
    data_directory_path  = "data/",
    model_directory_path = "resources/",
)
# Saves: resources/ensemble_v9.pt

Infer

from main import infer
import pandas as pd, anndata as ad

perts = pd.read_csv("data/predict_perturbations.txt", header=None)[0].tolist()
genes = ad.read_h5ad("data/obesity_challenge_1.h5ad").var_names.tolist()

infer(
    data_directory_path              = "data/",
    prediction_directory_path        = "outputs/",
    prediction_h5ad_file_path        = "outputs/prediction.h5ad",
    program_proportion_csv_file_path = "outputs/predict_program_proportion.csv",
    model_directory_path             = "resources/",
    predict_perturbations            = perts,
    genes_to_predict                 = genes,
)

Evaluate (requires local ground truth)

from main import evaluate
results = evaluate(
    data_directory_path   = "data/",
    model_directory_path  = "resources/",
    output_directory_path = "outputs/",
)

---

Key Hyperparameters

All hyperparameters are defined in the Config dataclass at the top of main.py:

Parameter	Value	Description
d_model	256	Transformer hidden dim
n_heads	8	Attention heads
n_layers	4	Transformer encoder layers
epochs	1200	Max training epochs
n_ensemble	16	Ensemble size (8 full + 8 LOO)
k_retrieve	10	KNN neighbours for Retrieve-and-Shift
cells_per_pert	100	Synthetic cells generated per perturbation
hvg_extra_w	4.0	Extra loss weight on highly variable genes

---

Official Leaderboard Scores

Scores from the CrunchDAO Broad Obesity-1 leaderboard (held-out test set — not reproducible locally):

Metric	Leaderboard Score
Pearson Delta	0.081
MMD	0.083
L1 Distance	0.150