Manifold Interpolating Optimal-Transport Flows for Trajectory Inference

Authors: Guillaume Huguet, Daniel Sumner Magruder, Alexander Tong, Oluwadamilola Fasina, Manik Kuchroo, Guy Wolf, Smita Krishnaswamy

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our method on simulated data with bifurcations and merges, as well as sc RNA-seq data from embryoid body differentiation, and acute myeloid leukemia treatment. We compare our results to those other methods that preform population flows including Trajectory Net [44] which is based on a CNF that is regularized to achieve efficient paths, and Diffusion Schrödinger s Bridge (DSB) [10] which is an optimal transport framework for generative modeling. The baseline measure in the quantitative results corresponds to the average distance between the previous and next timepoints for a given time t.
Researcher Affiliation Academia Guillaume Huguet1 D.S. Magruder2 Alexander Tong1 Oluwadamilola Fasina2 Manik Kuchroo2 Guy Wolf1 Smita Krishnaswamy2 1Université de Montréal; Mila Quebec AI Institute 2 Yale University
Pseudocode Yes We describe the training procedure of the GAE in algorithm 2 in the supplementary material. We describe the overall training procedure of MIOFlow in algorithm 1.
Open Source Code Yes Code is available here: https://github.com/KrishnaswamyLab/MIOFlow
Open Datasets Yes Code and links to publicly available data are provided in the supplemental material. We use Dyngen [6] to simulate a sc RNA-seq dataset from a dynamical cellular process.
Dataset Splits Yes Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] All training details are provided in the supplement.
Hardware Specification No Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] Compute resources are discussed in the supplement.
Software Dependencies No In practice, to compute the Wasserstein between discrete distributions, we use the implementation from the library Python Optimal Transport [14].
Experiment Setup No Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] All training details are provided in the supplement.