Excited to share our new work on Meta Flow Matching (MFM)! 🎉 MFM helps improve the generalization of personalized treatment response predictions by conditioning on learned features of the initial distribution. Work led by @lazar_atan and @NZhang211.

Excited to share our latest work on Metric Flow Matching (MFM)! Loved working with this team. Amazing to see how far we've come computationally since I started working in this field.

Happy to introduce our work on DEM which builds on incredible work on FAB by @SilkyDogfish, Vincent Stamper, @GregorSimms @bschoelkopf and @jmhernandez233 and Equivariant FM by @leonklein26 Andreas Krämer and @FrankNoeBerlin.

Excited to share our latest work on discrete diffusion models and the DDPP objective. A new approach for simulation-free fine-tuning. Check it out if you’re interested!.

Check out this extension of DynGFN (with @lazar_atan and @jasonhartford) which improves scalability with a more efficient parameterization. Reducing trajectory lengths from n^2 --> n.

Happening in ~2 hours! Code available here:

Had a great time working with this team. See our first work on flow matching for protein generation. More to come!.

Very cool application of Flow matching! Multi-Ligand binder design by learning flows from a prior with separated ligands and target.

Looking forward to presenting this work with @Guillaume_hu tomorrow!.

Checkout our new protein model @DreamFoldAI thread by first author @Guillaume_hu. Great work by the team and proud of our results!.

@KrishnaswamyLab @CellCellPress @christophertape @Maria_Ramos_Z Optimal transport is incredibly useful in single-cell analysis, but how do you handle controls? @christophertape already has a great thread on the biological impact, so in this thread I’ll focus on the ML and computational aspects. (1/n)

Thanks for having me! Had a great time at @HelmholtzMunich and many inspiring discussions.

I wholeheartedly agree😀! Great working with @christophertape and @Maria_Ramos_Z on this!.

Great tutorial on Trellis created by @Maria_Ramos_Z which shows how to combine hierarchical clustering and optimal transport for compare a large number of single-cell distributions.

Happy to be defending my thesis in a few hours! If you want to hear about our work in graphs, optimal transport, and deep learning, hope to see you there!

Awesome blog posts on the state of the field by @michael_galkin and @mmbronstein!.

Great working with this incredible team!.The most recent version (with updated results) is available on the open review site here:

Hope to see you there! TL;DR GflowNets for Bayesian (Dynamic) Gene Regulatory Network discovery.

Can't recommend working with @christophertape and lab enough! An awesome opportunity.

@jonkhler @SilkyDogfish @GregorSimms @bschoelkopf @jmhernandez233 @leonklein26 @FrankNoeBerlin @smnlssn @__kraemer__ Thanks Jonas for your thoughts! I believe your understanding is correct. When we test on the 55 particle Lennard-Jones system we get forces on the order of 10^30. In practice we clip and it seems to work fairly well. Definitely interested in how well it extends to proteins :).

@zdhnarsil

@KrishnaswamyLab @CellCellPress @christophertape @Maria_Ramos_Z Thank you for reading!.Code: Tutorial: Raw data: Preprocessed data: (10/n).

@alexechu_ @bose_joey @json_yim Hi @alexechu_, I was also curious about this. (1) It turns out that FM and diffusion have slightly different conditional paths as shown in Figure 3 of (2) the conditionals for the VE ODE are an optimal transport map for a linear noise schedule.

Great feature by @kevanchu and @dow_lab on our work on Trellis studying the Tumor microenvironment in PDOs.

@KrishnaswamyLab @CellCellPress @christophertape @Maria_Ramos_Z I’m excited to see what types of methods can be developed using this data. High-level summary, 25m cells, 44 features in a tensor of 10 patients, 11 drug combinations, 3 co-cultures states, with 3 replicates, which presents many opportunities for splitting and validation. (9/n).

@KrishnaswamyLab @CellCellPress @christophertape @Maria_Ramos_Z Trellis embeds distributions to vectors where the Wasserstein distance between distributions is equivalent to the L^1 distance between vectors. On a tree with hierarchical clustering, we compute the cell abundances in each node of the tree weighted by the edge weight. (7/n).

Grace choreographs a Tod's Shoes commercial via @richtong.

@KrishnaswamyLab @CellCellPress @christophertape @Maria_Ramos_Z The L^1 distance between these vectors is now the Wasserstein distance. This means that when used to create manifold-based embeddings between distributions which only require nearest neighbor distributions, Trellis scales log-linearly in the number of cells and samples. (8/n)

@yuanzhi_zhu @iScienceLuvr Yes! Cool to see what we call the minibatch-OT trick also works in these settings!. Some works that may interest the authors (@Yiheng_Li_Cal @Chenfeng_X) that also use this trick:.OT-CFM Multisample-FM A-OT

@KrishnaswamyLab @CellCellPress @christophertape @Maria_Ramos_Z This means that we can greatly reduce technical variation by comparing how treatments affect the cell distribution within the same plate. (5/n).

@KrishnaswamyLab @CellCellPress @christophertape @Maria_Ramos_Z Since this dataset is so large, it was collected using many individual experiments. This leads to technical batch effects. What saves us here is that each experiment has its own controls. (3/n)

@Pseudomanifold @unifr Congrats Bastian.

@KrishnaswamyLab @CellCellPress @christophertape @Maria_Ramos_Z Single cell datasets are getting larger and larger. Heren @maria_ramos_z profiled >25 million cells in 3,360 separate samples spanning patient, drug, and microenvironment specific effects. At this scale it’s difficult to understand the data let alone draw insights. (2/n).

@KrishnaswamyLab @CellCellPress @christophertape @Maria_Ramos_Z Since this dataset is so large, it was collected using many individual experiments. This leads to technical batch effects which can drown out subtle biological signals. What saves us here is that each individual experiment has its own controls. (4/n)

@k_neklyudov @UMontreal @Mila_Quebec Congrats Kirill!.

@KrishnaswamyLab @CellCellPress @christophertape @Maria_Ramos_Z We developed Trellis as an extremely fast and interpretable method to understand how a treatment changes the observed cell states. Trellis is able to measure the changes in changes of cell state distributions by how much work it takes to transport over a cell state tree. (6/n)

@Pseudomanifold @unifr Would love to :) Hopefully I'll make it at some point.

@KrishnaswamyLab @YaleMed @Yale @YaleCompsci @YaleGenetics Congratulations!.