←
Constantin Ahlmann-Eltze (@const-ae.bsky.social)238 posts
Constantin Ahlmann-Eltze (@const-ae.bsky.social) avatar

Constantin Ahlmann-Eltze (@const-ae.bsky.social)

@const_ae

Postdoc at UCL with @JL_Reading. Previously at EMBL working with @wolfgangkhuber. Biostats, R, cancer immunology Find me on 🦋 at const-ae.bsky.social

◷ Joined December 2010▧ 62 Photos and videos
Clint Miller avatar
⟳ Constantin Ahlmann-Eltze (@const-ae.bsky.social) retweeted
Clint Miller@clintomics·3 Jan 2025

New single-cell analysis framework, LEMUR, disentangles covariates and latent cell states from multi-condition data to predict counterfactual gene expression and identify cell neighborhoods with similar DEGs without clustering. doi.org/10.1038/s41588-024-0… @NatureGenet @wolfgangkhuber

Constantin Ahlmann-Eltze (@const-ae.bsky.social) avatar
Constantin Ahlmann-Eltze (@const-ae.bsky.social)@const_ae·6 Jan 2025

For the Monday morning crowds :) Also, take a look at my blog post, in which I compress the full lemur package into 100 lines of code and explain the core algorithm. const-ae.name/post/2025-01-0…

After 4y in the making, I am super excited that my main PhD project is published 🎉🥳🎉🎉🥳 nature.com/articles/s41588-0… LEMUR is a tool to analyze multi-condition single-cell data and model differential expression as a continuous function of the cell-state space. Some highlights⬇️

Post attachment
Constantin Ahlmann-Eltze (@const-ae.bsky.social) avatar
Constantin Ahlmann-Eltze (@const-ae.bsky.social)@const_ae·3 Jan 2025

In addition, we find for each gene the neighborhood of cells that are - close together in cell-type space, and - have maximal differential expression. This aggregation improves DE detection as we can optimally adjust our "cluster" resolution for each gene.

Post attachment
Constantin Ahlmann-Eltze (@const-ae.bsky.social) avatar
Constantin Ahlmann-Eltze (@const-ae.bsky.social)@const_ae·3 Jan 2025

In the end, we have a parametric model of multi-condition scRNA data, which allows us to predict the expression of any cell under any treatment. This is also called cluster-free differential expression. But the output can be disorientingly large: one LFC per gene and cell.

Constantin Ahlmann-Eltze (@const-ae.bsky.social) avatar
Constantin Ahlmann-Eltze (@const-ae.bsky.social)@const_ae·3 Jan 2025

But more intuitively, think of this approach as optimal transport where you match the __space__ occupied by cells instead of individual cells between conditions. On top, this means you don't need to worry about abundance changes within cell types.

Post attachment
Constantin Ahlmann-Eltze (@const-ae.bsky.social) avatar
Constantin Ahlmann-Eltze (@const-ae.bsky.social)@const_ae·3 Jan 2025

At its core, LEMUR is a matrix factorization with a _twist_! Instead of finding a single subspace R to approximate the data, I find one subspace per condition R(x). Inspired by generalized linear models (GLMs), I implement this as solving a regression on subspaces.

Post attachment
Aino-Maija Leppä avatar
⟳ Constantin Ahlmann-Eltze (@const-ae.bsky.social) retweeted
Aino-Maija Leppä@amleppa·25 Nov 2024

Extremely proud to see my PhD project exploring intra-tumor heterogeneity in complex karyotype AML finally out in @NatureGenet ! 🎉👩🏻‍🔬 They say that 10,000 hours make you an expert, so I guess after more than five years of work I’m half-way there 😅 ⬇️1/5 nature.com/articles/s41588-0…

Charlotte Bunne avatar
⟳ Constantin Ahlmann-Eltze (@const-ae.bsky.social) retweeted
Charlotte Bunne@_bunnech·25 Sep 2024

🚀Interested in doing a PostDoc in my lab? Consider applying for a🇨🇭SNSF Swiss Postdoctoral Fellowship 2024. If you have any questions, join the information event tomorrow via Zoom (register via shorturl.at/zP93m) or reach out to us via bunnelab.com! 🧬Who should apply? Check that you fulfill the requirements. And: We are in particular interested in candidates with background in computational biology or machine learning applied to the biomedical domain!

📢 #Call for SNSF Swiss Postdoctoral Fellowships 2024 This highly competitive call is aimed at researchers who wish to obtain a position in a non-profit research institution in Switzerland for a period of 12 to 24 months. 🗓️Deadline: 2 December 2024 sohub.io/6afa

Constantin Ahlmann-Eltze (@const-ae.bsky.social) avatar
Constantin Ahlmann-Eltze (@const-ae.bsky.social)@const_ae·21 Sep 2024

Great thread by Anshul, one of the pioneers of deep learning for biology, summarizing our preprint!

biorxiv.org/content/10.1101/… Nice benchmark of single cell "foundation models" (scGPT, scFoundation) and GEARS (a GNN model) further hyped as "virtual cell models" against linear baselines on perturbation prediction. Long-story short: they can't beat the linear baselines. 1/