Technical Summary: FlowGRN

Problem Statement

Inferring Gene Regulatory Networks (GRNs) from single-cell RNA sequencing (scRNA-seq) data is critical for understanding cellular dynamics and designing therapeutic strategies. However, this task faces two primary challenges:

1. Lack of Temporal Information: scRNA-seq provides static snapshots of cell states rather than continuous time-series data. Key dynamic changes, such as cell transitions, are often unobserved due to sparse sampling, making it difficult to restore cellular dynamics and determine the directionality of regulations.
2. Dropout Noise and High Dimensionality: scRNA-seq data suffers from "dropout" events, where expressed genes are not detected (recorded as zeros). In high-dimensional gene spaces (tens of thousands of genes), standard distance metrics like Euclidean distance become less informative due to the curse of dimensionality and are severely biased by dropout noise. Existing methods often either oversimplify system dynamics, rely on computationally intractable neural ODEs for high dimensions, or discard temporal information entirely, rendering regulatory directions unidentifiable.

Methodology: FlowGRN

FlowGRN is a scalable, dropout-robust framework that integrates three core components to reconstruct cellular trajectories and infer directed GRNs:

1. Dropout-Robust Cell Similarity Measure

To address the bias introduced by dropout noise in distance calculations, FlowGRN proposes a novel similarity measure ($d_{raw}$).

• Mechanism: Instead of using all genes, the measure calculates the average L1 distance only over the set of genes that are non-zero in both cells being compared ($S_{x,y}$). This ensures that technical zeros (dropouts) do not inflate the distance.
• Geometric Awareness: To capture the underlying manifold geometry of gene expression data, a k-nearest neighbor (kNN) graph is constructed using $d_{raw}$ as edge weights. The final similarity metric is the shortest path distance on this graph, computed via Dijkstra's algorithm. This approach is resilient to dropout and high dimensionality.

2. Trajectory Reconstruction via Conditional Flow Matching (CFM)

FlowGRN utilizes Conditional Flow Matching (CFM) and Score Matching to learn the vector fields driving cellular transitions, avoiding the computational cost and instability of traditional Neural ODE integration.

• Optimal Transport (OT): The model uses the OT plan, guided by the dropout-robust similarity measure, to define target vector fields between consecutive time snapshots.
• Log-Domain Transformation: To handle the non-negativity constraint of gene expression values during trajectory integration, data is transformed into a log domain. This transformation preserves differentiability and captures dynamics across different expression scales.
• Reconstruction: The learned vector field is used to solve an initial value problem (IVP), reconstructing continuous trajectories for each cell lineage from the static snapshots.

3. GRN Inference via dynGENIE3

Once trajectories are reconstructed, FlowGRN employs dynGENIE3, a dynamic extension of the GENIE3 algorithm, to infer the GRN.

• Dynamic Modeling: For each target gene, a random forest model predicts its rate of change ($dx/dt$) based on the reconstructed trajectory, rather than static expression levels.
• Edge Type Assignment: While dynGENIE3 provides regulatory strength (importance scores), FlowGRN determines the sign of regulation (activation vs. inhibition) by analyzing the signs of the Jacobian matrix of the learned flow field ($v_\theta$) across the trajectory. This allows for the inference of signed, directed networks.

Key Contributions

1. Novel Similarity Measure: A cell similarity metric robust to dropout noise and the curse of dimensionality, utilizing non-zero gene intersections and kNN geodesic distances.
2. Scalable Trajectory Reconstruction: Application of CFM and score matching to reconstruct non-linear cellular trajectories in high-dimensional spaces (up to 1,783 genes), overcoming the limitations of traditional Neural ODEs.
3. Integration of Dynamics and Inference: A pipeline that uses reconstructed trajectories to feed dynGENIE3, enabling the inference of identifiable (directed) GRNs with edge types (activation/inhibition), a capability often missing in static methods.
4. State-of-the-Art Performance: Demonstration of superior performance on the BEELINE benchmark compared to existing methods.

Experimental Results

The authors evaluated FlowGRN on the BEELINE benchmark, comprising synthetic, curated, and experimental datasets.

• Synthetic and Curated Data: FlowGRN achieved the highest average rank for both Area Under Precision-Recall Curve (AUPRC) and Early Precision Ratio (EPR) among all tested models (including LEAP, SCODE, GRISLI, and GENIE3). It notably outperformed static models (like GENIE3) in datasets with long cascades (LL) and cyclic dynamics (CY), where directionality is crucial.
• Edge Type Inference: In tasks requiring the identification of activation vs. inhibition edges, FlowGRN achieved the best performance in 7 out of 10 datasets, significantly surpassing the second-best model (GRNVBEM).
• Experimental Data: On 7 real-world experimental datasets (e.g., hESC, mHSC), FlowGRN ranked first or second in 10 out of 14 tasks (across AUPRC and EPR metrics). It demonstrated scalability up to 1,783 genes.
• Ablation Studies:
  • Trajectory Necessity: Removing the trajectory reconstruction step (using static data with dynGENIE3 or direct Jacobian inference) resulted in significantly lower performance, confirming that trajectory reconstruction is indispensable for capturing indirect regulations.
  • Dropout Robustness: Ablating the dropout-robust similarity measure led to performance drops, particularly in high-dimensional settings, validating its role in mitigating noise.

Significance and Limitations

Significance:
The paper claims that FlowGRN addresses the "dilemma" in GRN inference between the need for temporal information and the challenges of high dimensionality and dropout noise. By combining CFM for dynamics modeling with a robust similarity measure and dynGENIE3 for inference, FlowGRN offers a precise, scalable, and identifiable solution that outperforms current state-of-the-art methods in both accuracy and the ability to model complex dynamics (cycles and long cascades).

Limitations:
The authors acknowledge specific limitations:

• Data Scarcity: Performance drops on datasets with very few cells (e.g., mDC, mESC), a common issue for deep learning models requiring sufficient sample sizes.
• Unified Framework: The current method separates dynamic modeling (CFM) and network inference (dynGENIE3) rather than unifying them into a single interpretable framework.
• Future Work: The authors suggest future directions involving physics-informed neural networks (PINNs) to restrict the CFM model with biological constraints and aligning trajectories more closely with cell distribution shifts.