[Paper Review] Cross-modality Matching and Prediction of Perturbation Responses with Labeled Gromov-Wasserstein Optimal Transport

Jayoung Ryu, Charlotte Bunne+3 more|arXiv (Cornell University)|May 1, 2024

Plant Surface Properties and Treatments6 citations

TL;DR

Extends GWOT/COOT with perturbation labels to improve cross-modality alignment and predict perturbation responses across modalities in multi-modal single-cell perturbation data.

ABSTRACT

It is now possible to conduct large scale perturbation screens with complex readout modalities, such as different molecular profiles or high content cell images. While these open the way for systematic dissection of causal cell circuits, integrated such data across screens to maximize our ability to predict circuits poses substantial computational challenges, which have not been addressed. Here, we extend two Gromov-Wasserstein Optimal Transport methods to incorporate the perturbation label for cross-modality alignment. The obtained alignment is then employed to train a predictive model that estimates cellular responses to perturbations observed with only one measurement modality. We validate our method for the tasks of cross-modality alignment and cross-modality prediction in a recent multi-modal single-cell perturbation dataset. Our approach opens the way to unified causal models of cell biology.

Motivation & Objective

• Motivated by multi-modal perturbation screens where cells are profiled with different readouts (RNA, proteins, images) and labeled by perturbations.
• Develops labeled extensions of entropic GWOT (EGWOT) and COOT to leverage perturbation labels in cross-modality alignment.
• Demonstrates that label-informed GWOT/COOT improve cross-modality matching and enable out-of-sample prediction of perturbation responses.
• Provides an open-source implementation and benchmarking on a multi-modal perturbation dataset to enable unified causal modeling of cell biology.

Proposed method

• Define label-identity constraint B^l enforcing l-compatible couplings where T_ij>0 only if l_x_i = l_y_j.
• Propose Labeled Entropic-regularized GWOT (Labeled EGWOT) with Sinkhorn-based updates and prove structural form of the l-compatible OT plan.
• Show that labeled GWOT cost can be accelerated when cost is a sum-of-functions form, enabling faster computations for large perturbation screens.
• Adapt Labeled COOT (and ECOOT) to jointly optimize sample transport per label with a shared global feature transport across labels.
• Use the learned coupling T to train a cross-modality predictor (MLP) that predicts RNA from protein measurements via sampling j ~ Multinomial(T_i· / sum T_i·).
• Benchmark against unlabeled OT/ GWOT, DAVAE, and per-label variants on a multi-modal perturbation dataset to assess matching, prediction, and feature matching.

Experimental results

Research questions

• RQ1 Can perturbation labels improve cross-modality alignment between RNA and protein (or other modalities) in perturbation screens?
• RQ2 Do labeled GWOT/COOT extensions yield better sample matching and predictive accuracy for perturbation responses compared to unlabeled methods?
• RQ3 Does incorporating label information enable accurate out-of-sample prediction of perturbation effects in one modality from observations in another?

Key findings

• Labeled GWOT-based methods (EGWOT and ECOOT) outperform unlabeled OT/GWOT baselines for matching and prediction.
• Incorporating perturbation labels enables sharing information across perturbations, improving global topology learning and sample couplings.
• Per-label GWOT variants improve over no-label approaches but are surpassed by fully labeled approaches due to cross-label information sharing.
• Compared with DAVAE baselines, labeled GWOT methods achieve better matching and prediction, while DAVAE shows weaker predictive performance.
• The framework yields superior sample matching, prediction, and feature matching, suggesting improved interpretability of cross-modality features.
• Code and data are made available (Genentech/Perturb-OT) for reproducible benchmarking.