Technical Summary: FlowGRN+

Problem Statement

Inferring Gene Regulatory Networks (GRNs) from single-cell RNA sequencing (scRNA-seq) data is a critical challenge in computational biology. While scRNA-seq provides high-resolution snapshots of gene expression, the destructive nature of the assay prevents tracking individual cells over time, resulting in static data that obscures underlying dynamic processes.

Existing methods face significant limitations:

• Static Approaches: Classical methods (e.g., GENIE3, TIGRESS) treat data as static snapshots, failing to capture temporal dynamics.
• ODE-Based Models: Methods like SCODE and GRISLI attempt to model dynamics via Ordinary Differential Equations but often suffer from oversimplification and scalability issues in high-dimensional, non-linear systems.
• Deep Learning Limitations: While Conditional Flow Matching (CFM) offers a promising framework for learning vector fields from distributional shifts, the previous state-of-the-art model, FlowGRN, exhibits specific shortcomings:
  1. Temporal Incoherence: It relies on linear interpolation between discrete time points, creating discontinuities in the learned vector field at interval boundaries.
  2. Dropout Sensitivity: Linearly interpolated reference trajectories are susceptible to noise and "dropouts" (zero values), leading to unstable reference velocity vectors.
  3. Lack of Automation: It requires well-separated snapshot distributions, often necessitating manual human inspection and clustering to partition cells, which hinders reproducibility.
  4. Manifold Misalignment: Reconstructed trajectories may drift away from the intrinsic data manifold.

Methodology: FlowGRN+

The authors propose FlowGRN+, an enhanced framework that integrates spline fitting and manifold projection into the CFM pipeline to address the limitations of FlowGRN. The method operates in a log-transformed gene expression space and follows a five-step training pipeline:

1. Pseudotime Binning and Sampling

Instead of relying on clustering algorithms to define time intervals, FlowGRN+ bins cells evenly by count along a pre-computed pseudotime axis ($\tau$). This stratified sampling strategy automates the process and removes the need for manual parameter tuning.

2. Spline Fitting for Reference Trajectories

To generate stable reference trajectories, the method samples cells from consecutive pseudotime bins and fits smoothing splines (B-splines with curvature regularization) rather than linear or cubic interpolations.

• Mechanism: The spline $\gamma(\tau)$ minimizes a loss function balancing data fitting and curvature ($\lambda_{reg}$).
• Benefit: This approach links non-local observations, mitigates the impact of dropouts through smoothing, and scatters potential discontinuities across the pseudotime range, providing a more robust training signal for the CFM model.

3. Manifold Projection Scheme

To address the "overshooting" problem where splines may produce reference velocities pointing off the data manifold, FlowGRN+ introduces a projection scheme:

• Local Geometry Estimation: For a given cell $x$, the method identifies $k$-nearest neighbors (kNN) to estimate the local tangent space of the data manifold.
• Spectral Filtering: A weighted local covariance matrix is computed. To avoid instability in selecting the intrinsic dimension, the authors apply spectral filtering (shrinkage) to the eigenvalues, generating an approximate projection matrix $P(x)$.
• Refinement: Both the sampled cell positions and the reference velocity vectors are projected onto the tangent space and the manifold, respectively. An affine retraction operator ensures points remain close to the manifold during integration.

4. Conditional Flow Matching Training

The CFM model $v_\theta(x, \tau)$ is trained to match the projected reference velocities using a metric-aware loss function. The loss incorporates the projection matrix $P(x)$ as a metric, penalizing errors in the tangent space while adding a small regularization term to prevent trajectories from drifting off-manifold during integration.

5. Trajectory Reconstruction and GRN Inference

Once trained, the vector field is integrated (using tangent-projected velocities and retraction steps) to reconstruct continuous cell trajectories. These trajectories are mapped back to the gene expression domain and fed into dynGENIE3 to infer the directed GRN.

Key Contributions

The paper outlines three primary contributions:

1. FlowGRN+ Framework: An improved GRN inference model that utilizes spline fitting to generate dropout-robust reference trajectories, significantly enhancing the temporal coherence of learned dynamics compared to linear interpolation.
2. Manifold Projection Scheme: A novel method to refine reference velocity vectors by projecting them onto the local tangent space of the data manifold, effectively mitigating spline overshooting and ensuring trajectories adhere to the data support.
3. Empirical Evaluation: A comprehensive evaluation on the BEELINE benchmark demonstrating that FlowGRN+ achieves top-tier performance in trajectory smoothness and competitive GRN inference metrics, while reducing reliance on manual intervention.

Experimental Results

The authors evaluated FlowGRN+ on seven experimental datasets from the BEELINE benchmark (human and mouse cells) using two primary metrics:

• AUPRC (Area Under Precision-Recall Curve): Measures overall precision-recall trade-off.
• EPR (Early Precision Ratio): Measures precision among top-ranked edges, emphasizing the recovery of high-confidence regulatory relationships.

Key Findings:

• Performance: FlowGRN+ achieved top-tier performance, ranking in the top two models for AUPRC in 6 out of 14 settings and for EPR in 10 out of 14 settings.
• Comparison with FlowGRN: FlowGRN+ showed a higher EPR than the original FlowGRN in 10 of 14 settings, indicating superior recovery of top-ranked edges. However, its AUPRC was higher in only 4 of 14 settings, suggesting the improvement is concentrated in the most confident predictions rather than the full ranked list.
• Smoothness: Ablation studies confirmed that the combination of smoothing splines and manifold projection yields the smoothest reconstructed trajectories (lowest normalized Total Variation) compared to linear interpolation, cubic splines, or models without projection.
• Ablation Insights:
  • Spline fitting alone (without CFM learning) resulted in poor GRN inference (near-zero EPR), confirming the necessity of learning a coupled, state-dependent vector field.
  • The projection scheme improved EPR in 66.67% of settings and reduced trajectory roughness in 71.43% of settings.
  • Smoothing splines outperformed cubic and PCHIP splines in both EPR rank and trajectory smoothness.

Significance and Limitations

Significance:
FlowGRN+ provides a practical, automated framework for reconstructing cell trajectories and inferring GRNs from scRNA-seq data. By improving the temporal coherence of dynamics and aligning trajectories with the data manifold, it addresses critical reproducibility and stability issues in previous CFM-based approaches. The insights regarding spline fitting and manifold projection may be applicable to other CFM-based models of cellular dynamics.

Limitations:
The authors modestly acknowledge several constraints:

• Branching Dynamics: The smooth manifold assumption may be violated in datasets with complex branching trajectories, and the uniqueness of ODE solutions may limit the modeling of such bifurcations.
• Computational Cost: The projection scheme, particularly the repeated kNN queries and matrix inversions, increases training time compared to FlowGRN.
• Downstream Model Sensitivity: The performance of the final GRN inference depends heavily on the choice of the downstream model (e.g., dynGENIE3), and the authors note that different model classes can materially affect network recovery.
• Future Work: The paper suggests future directions include better extraction of networks from learned dynamics and constraining CFM models to align with known biological pathways.