TL;DR
This paper introduces DeltaNMF, a novel neural network-based approach to Non-negative Matrix Factorization (NMF) for single-cell RNA sequencing data. DeltaNMF effectively disentangles changes in gene program usage between conditions (e.g., disease vs. control) from the emergence of entirely new programs, a critical challenge in current methods. It achieves this through a two-stage fitting process and by incorporating biological priors from foundation models, leading to more interpretable and accurate biological discoveries.
What Problem Does This Paper Address?
Single-cell RNA sequencing (scRNA-seq) generates massive datasets that researchers analyze using dimensionality reduction techniques like NMF to identify "gene programs" – sets of genes that are co-expressed and often represent biological processes. A common use case is comparing cells from different conditions, such as healthy versus diseased tissue, or treated versus untreated cells. However, standard NMF struggles to distinguish between two key types of differences:
- Quantitative shifts: A gene program exists in both conditions but is used more or less in one.
- Qualitative shifts: An entirely new gene program emerges in one condition, or a program disappears.
Standard NMF, by fitting a single set of programs to all data, conflates these scenarios. This ambiguity can obscure genuine biological discoveries, making it hard to tell if observed differences are due to altered activity of known processes or the appearance of novel cellular states. This is particularly problematic in case-control studies where precise identification of disease-specific mechanisms is crucial.
Methods
DeltaNMF is a neural network reformulation of NMF. This allows for greater flexibility in defining the objective function and incorporating biological knowledge. Key innovations include:
- Neural Network Formulation: NMF is framed as a differentiable optimization problem solvable by neural networks. This enables the use of standard deep learning tools like GPU acceleration and flexible loss functions. Non-negativity is enforced using activation functions like Softplus.
- Foundation Model Integration: Gene similarity information, learned by large language models for scRNA-seq (like scGPT and TranscriptFormer), is incorporated as a regularization term. This "graph-Laplacian regularization" encourages genes within a program that are similar in the foundation model's learned space to have similar loadings in the NMF program, improving biological coherence.
- Two-Stage Case-Control Architecture: This is the core innovation for differential analysis.
- Stage 1: Baseline gene programs are learned solely from control cells.
- Stage 2: Additional "case-specific" programs are learned for case cells, while the baseline programs from Stage 1 are kept active and fixed. An orthogonality penalty is applied to ensure case-specific programs are distinct from baseline ones. This explicitly separates changes in existing programs from the emergence of new ones.
Datasets
- Synthetic Data: Used to validate the method's ability to recover known ground truth programs and to test the separation of distinct vs. shifted programs.
- Coronary Artery Disease (CAD) Perturb-seq Data: Used to identify disease-specific programs in endothelial cells.
- Alzheimer's Disease (AD) Perturb-seq Data: Used to identify disease-specific programs in macrophages.
The paper also leverages pre-trained gene embeddings from scGPT and TranscriptFormer, which are large-scale foundation models trained on millions of single-cell transcriptomes.
Key Insights and Ideas
- Neural NMF is more flexible: Reframing NMF as a neural network opens the door to easily incorporating complex biological priors and structured fitting procedures that were previously difficult to implement.
- Foundation models enhance biological coherence: Using gene similarities from foundation models significantly improves the biological interpretability of NMF programs, making them more tightly linked to known biological processes and pathways.
- Two-stage fitting is crucial for case-control studies: Explicitly separating baseline programs from case-specific ones is essential for accurately identifying novel disease mechanisms, rather than just observing altered activity of existing pathways.
- DeltaNMF is computationally efficient: The GPU-accelerated neural network approach offers a substantial speedup over existing methods like consensus NMF.
Key Results
- Improved Program Coherence: Foundation model regularization improved program coherence by 2.4-fold on protein interaction networks.
- Accurate Program Recovery: DeltaNMF accurately recovered ground-truth programs in synthetic data.
- Computational Speedup: DeltaNMF achieved a ~20x speedup over consensus NMF while maintaining comparable accuracy.
- Differential Program Discovery: In CAD Perturb-seq data, DeltaNMF successfully identified disease-specific programs, including the cerebral cavernous malformation (CCM) pathway, and clearly distinguished them from baseline endothelial programs with altered activity.
- AD Macrophage States: DeltaNMF identified distinct macrophage states in AD, including an APOE-linked lipid-lysosome program.
- Robustness: The method's performance was robust across different regularization strengths and foundation models.
Limitations and Future Work
- Discrete Case-Control Assumption: The two-stage architecture assumes discrete case-control separation, making it less suitable for continuous processes like differentiation trajectories.
- Control Purity Assumption: The method assumes control cells represent "pure" baseline states, which might not always hold if controls have subtle perturbations.
- Orthogonality Penalty: The orthogonality penalty, while effective for separation, might overly restrict programs that legitimately share components.
- GPU Requirements: While standard, GPU requirements might limit accessibility for some users.
Future work could involve:
- Semi-supervised Formulations: Incorporating perturbation metadata directly into the loss function to link programs to specific genetic drivers.
- Hierarchical Architectures: Modeling shared and condition-specific programs across multiple conditions.
- Extending to Continuous Processes: Adapting the framework for analyzing continuous biological processes.
Summary updated