A comprehensive deep learning framework for predicting single-cell RNA-seq expression from genomic DNA sequences in zebrafish development.
DanioDecima extends the Borzoi architecture to predict cell-type-specific gene expression patterns across zebrafish developmental stages. The framework implements transfer learning from human and mouse models, enabling cross-species regulatory sequence analysis and synthetic regulatory element design.
- Multi-species Transfer Learning: Systematic comparison of human vs. mouse pretrained models for zebrafish expression prediction
- Cell-Type Specificity: Predictions across 85 cell-type × developmental timepoint combinations
- Attribution Analysis: Model interpretability through genomic feature attribution to assess conservation confounding
- Regulatory Element Design: Directed evolution pipeline for generating cell-type-specific synthetic promoters
- Comprehensive Evaluation: Statistical framework for model performance assessment across developmental stages
- Input: 5-channel sequences (4 DNA bases + gene mask, 524,288bp windows)
- Architecture: 7 CNN blocks + 8 Transformer blocks (1,920 embedding channels)
- Output: Cell-type-specific expression predictions via exponential activation
- Loss Function: TaskWisePoissonMultinomialLoss (Poisson + multinomial components)
- 16 Model Configurations: 4 initialization strategies × 4 replicates each
- Initialization Types: Random, Human-Borzoi, Human-Decima, Mouse-Borzoi pretraining
- Training: Early stopping, gradient accumulation, distributed GPU compute via SLURM
- Data Exploration: Single-cell preprocessing, variance analysis, conservation metrics
- Model Training: Distributed training pipeline with systematic experimental design
- Prediction Generation: Scalable inference with sequence augmentation
- Performance Evaluation: Cell-type-specific metrics and developmental stage analysis
- Attribution Analysis: InputXGradient interpretability and conservation confounding assessment
- Regulatory Design: Evolutionary optimization of synthetic regulatory elements
- Transfer Learning Effectiveness: Quantified performance gains from cross-species pretraining
- Developmental Patterns: Temporal prediction accuracy across zebrafish embryogenesis
- Conservation Analysis: Assessment of model reliance on evolutionary vs. regulatory signals
- Synthetic Design: Generated cell-type-specific promoter elements with motif analysis
daniodecima-applications-main/notebooks/
├── 2_dataset/ # Data exploration and preprocessing
├── 4_evaluation/ # Model evaluation and performance analysis
├── 5_specificity/ # Attribution analysis pipeline
└── 9_design/ # Regulatory element design workflows
daniodecima-main/scripts/
├── decima_finetune.py # Model training
├── decima_predictions.py # Prediction generation
└── submit_*.sh # SLURM job submission scripts
- PyTorch Lightning
- Captum (attribution analysis)
- scanpy (single-cell analysis)
- SLURM (distributed computing)
- GPU compute (H100/H200 recommended)
sbatch daniodecima-main/scripts/submit_decima_finetune.shsbatch daniodecima-applications-main/notebooks/4_evaluation/00_submit_predict_decima.shsbatch daniodecima-applications-main/notebooks/5_specificity/00_submit_combined_attributions.shsbatch daniodecima-applications-main/notebooks/9_design/00_submit_evolve_combined.shPrimary Framework:
@article{lal2024decoding,
title={Decoding sequence determinants of gene expression in diverse cellular and disease states},
author={Lal, Avantika and Karollus, Alexander and Gunsalus, Laura and Garfield, David and Nair, Surag and Tseng, Alex M and Gordon, M Grace and Blischak, John and van de Geijn, Bryce and Bhangale, Tushar and others},
journal={bioRxiv},
pages={2024--10},
year={2024},
publisher={Cold Spring Harbor Laboratory}
}
ZebraHub Dataset:
@article{lange2024multimodal,
title={A multimodal zebrafish developmental atlas reveals the state-transition dynamics of late-vertebrate pluripotent axial progenitors},
author={Lange, Merlin and Granados, Alejandro and VijayKumar, Shruthi and Bragantini, Jord{\~a}o and Ancheta, Sarah and Kim, Yang-Joon and Santhosh, Sreejith and Borja, Michael and Kobayashi, Hirofumi and McGeever, Erin and others},
journal={Cell},
volume={187},
number={23},
pages={6742--6759},
year={2024},
publisher={Elsevier}
}This Work:
@article{daniodecima2025,
title={DanioDecima: A DNA sequence-to-function model of zebrafish embryogenesis},
author={Voges, Mathias, et al.},
journal={In preparation},
year={2025}
}This project accompanies the DanioDecima manuscript (in preparation). It is maintained for reproducibility of the published results; critical bug fixes will be made as required. Issues and pull requests are welcome — see CONTRIBUTING.md.
We warmly welcome contributions from the community! Please see CONTRIBUTING.md for how to get started.
This project adheres to the Contributor Covenant code of conduct. By participating, you are expected to uphold this code. Please report unacceptable behavior to opensource@chanzuckerberg.com.
If you believe you have found a security issue, please responsibly disclose by contacting us at security@biohub.org. See SECURITY.md for details.
See LICENSE.md. This repository is a derivative work of Genentech/decima and Genentech/decima-applications and inherits their Non-Commercial Software License v1.0 (commercial use prohibited). Per-fork details are documented in daniodecima-main/FORK_NOTES.md and daniodecima-applications-main/FORK_NOTES.md. See NOTICE for upstream attribution and citation guidance.