Efficient Data Selection for Training Genomic Perturbation Models

TL;DR

The paper tackles the slow, multi-round training of graph-based genomic perturbation predictors by introducing GraphReach, a one-shot subset-selection method that maximizes supervision propagation on a knowledge graph. By greedily maximizing the reach of supervision, GraphReach offers a near-optimal guarantee via submodular optimization and a lazy-forward implementation, enabling parallel PerturbSeq experiments and drastic reductions in real-world lab time. Empirical evaluation across four PerturbSeq datasets shows GraphReach delivers substantial speedups (months to weeks) with competitive accuracy and improved stability, even under graph noise, outperforming random selection and traditional active-learning baselines in several settings. The work highlights a practical strategy for accelerating genomic discovery that decouples selection from model initialization, with potential extensions to hybrid approaches and non-graph-based methods.

Abstract

Genomic studies face a vast hypothesis space, while interventions such as gene perturbations remain costly and time-consuming. To accelerate such experiments, gene perturbation models predict the transcriptional outcome of interventions. Since constructing the training set is challenging, active learning is often employed in a "lab-in-the-loop" process. While this strategy makes training more targeted, it is substantially slower, as it fails to exploit the inherent parallelizability of Perturb-seq experiments. Here, we focus on graph neural network-based gene perturbation models and propose a subset selection method that, unlike active learning, selects the training perturbations in one shot. Our method chooses the interventions that maximize the propagation of the supervision signal to the model. The selection criterion is defined over the input knowledge graph and is optimized with submodular maximization, ensuring a near-optimal guarantee. Experimental results across multiple datasets show that, in addition to providing months of acceleration compared to active learning, the method improves the stability of perturbation choices while maintaining competitive predictive accuracy.

Figures (3)

• Figure 1: The difference between active learning and subset selection for training GEARS.
• Figure 2: An overview of the subset selection methodology. Our sole input is the knowledge graph from GO51 roohani2024predicting and a list of candidate perturbations. The subset selection algorithm, such as GraphReach, selects the set of gene perturbations, and they are given to the wet lab for the experimental part. Finally, the single-cell gene expression data is given to GEARS for training and validation.
• Figure 3: Accuracy in the test set with confidence intervals compared to the total physical time required for the training due to wet-lab iterations. GraphReach and Baseline do not require model-input hence they are trained with only one cycle of wet lab experiments taking 30 days. TypiClust and ACS-FW inherently need numerous cycles, and they are each run for 90 and 150 days to highlight the role that the number of cycles plays in the accuracy.

Theorems & Definitions (2)

• Proposition 1
• Proposition 2