FlowTransOP

FlowTransOP is a deep generative framework for the distributional translation of omics signatures across biological domains, developed by the Lauffenburger Lab at MIT in collaboration with the University of Amsterdam and Washington State University. The model addresses a recurring obstacle in translational biology: experimental findings from one system, such as a mouse model or an in vitro cell line, often fail to generalize to another, such as human tissue, and the molecular signatures measured in each domain rarely align on a shared feature space or come from paired samples. Rather than mapping individual samples one-to-one, FlowTransOP learns to transport an entire distribution of signatures from a source domain to a target domain, preserving the population-level structure that single-sample methods discard.

The central methodological contribution is the use of constrained deep flow matching, a class of continuous-time generative models that learn a velocity field carrying probability mass from one distribution to another. Introduced in a May 2026 bioRxiv preprint titled "Distributional Translation of Omics Signatures via Constrained Deep Flow Matching," the approach removes two restrictive assumptions common to prior cross-domain methods: it does not require paired source-target samples, and it does not require a one-to-one correspondence between features (for example, between mouse and human orthologous genes).

FlowTransOP fits within the broader landscape of cross-species and cross-modality transcriptomic models but distinguishes itself by framing the problem in the language of optimal transport and distributional alignment, making it applicable wherever two omics datasets must be reconciled despite differing biology, platforms, or feature sets.

#Key Features

• Distributional translation: Transports the full distribution of omics signatures between domains rather than mapping individual samples, preserving population-level molecular structure.
• No paired samples required: Learns the source-to-target mapping without requiring matched samples across domains, a common constraint in real-world cross-system datasets.
• No one-to-one feature mapping: Operates without requiring explicit one-to-one correspondence between features, enabling translation across species with imperfect gene orthology.
• Constrained flow matching: Uses a constrained deep flow-matching objective to learn the continuous transport between distributions in a shared latent space.
• Reusable pretrained latent space: A latent space pretrained on large transcriptomic compendia can be reused for new case studies without training from scratch.
• Packaged for inference: Distributed as a flowtransop pip package with an inference command-line interface and API, alongside MIT-licensed source code and released weights.

#Technical Details

FlowTransOP couples an autoencoder that embeds omics signatures into a shared latent representation with a flow-matching model that learns a velocity field transporting source-domain latents toward the target-domain distribution, subject to constraints that regularize the transport. The framework was pretrained on large-scale transcriptomic resources, including ARCHS4 (covering both human and mouse RNA-seq) and the L1000 perturbational expression compendium, providing broad coverage of biological states and perturbations. Because translation is performed in the learned latent space, the same pretrained model can be applied to new domains without retraining the underlying representation. Trained model weights are archived on Zenodo (DOI 10.5281/zenodo.20434738), and the implementation is released under the MIT license.

#Applications

FlowTransOP is aimed at translational researchers who need to reconcile omics data across biological systems, most notably translating signatures between mouse models and human tissue. The authors demonstrate this with a metabolic dysfunction-associated steatohepatitis (MASH) liver-disease case study, in which the pretrained latent space is reused to translate disease signatures without training a new model from scratch. More broadly, the framework can support cross-species inference, drug-response extrapolation from cell lines to patients, and any setting where conclusions drawn in one omics domain must be projected onto another that lacks paired measurements or aligned features.

#Impact

By recasting cross-domain omics translation as a constrained distributional transport problem, FlowTransOP offers a principled alternative to feature-aligned, sample-paired mapping methods that often break down across species or platforms. Its ability to reuse a pretrained latent space for downstream case studies, demonstrated on MASH liver disease, lowers the barrier to applying generative translation in new biological contexts. As of mid-2026 the work is a bioRxiv preprint and has not yet undergone formal peer review, so its benchmarks and generalization claims await independent validation; nonetheless, the open MIT-licensed code, released weights, and packaged inference tooling make it a readily adoptable resource for the translational omics community.