Bence Szalai

In October 2024, I twote that "something is deeply wrong" with what we now call virtual cell models. A lot has happened since then. How am I updating? New blog post: ekernf01.github.io/virtual-cell...

Beeswarm plot of the prediction error across different methods of double perturbations showing that all methods (scGPT, scFoundation, UCE, scBERT, Geneformer, GEARS, and CPA) perform worse than the additive baseline.

Line plot of the true positive rate against the false discovery proportion showing that none of the methods is better at finding non additive interactions than simply predicting no change.

Our paper benchmarking foundation models for perturbation effect prediction is finally published 🎉🥳🎉

www.nature.com/articles/s41...

We show that none of the available* models outperform simple linear baselines. Since the original preprint, we added more methods, metrics, and prettier figures!

🧵

Thanks to all the coauthors, Eszter Varga, Daniel Dimitrov,
@juliosaezrod.bsky.social, László Hunyady and especially to first author Szilvia Barsi who led this project from start to finish! 7/7

GitHub - basvaat/RIDDEN_tool: RIDDEN command-line interface: Data-driven inference of receptor activity for cell-cell communication studies RIDDEN command-line interface: Data-driven inference of receptor activity for cell-cell communication studies - GitHub - basvaat/RIDDEN_tool: RIDDEN command-line interface: Data-driven inference o...

RIDDEN is available as an easy to use command line tool at github.com/basvaat/RIDD.... 6/n

Using an IO dataset of #nivolumab-treated renal cancer patients, we found that:
- #PD1 or #PDL1 expression didn't predict survival
- PD1 activity (from RIDDEN) did associate with survival.
Showing that receptor activity could be a more effective biomarker than expression. 5/n

RIDDEN’s receptor-specific signatures align with regulons of transcription factors downstream of receptors. 4/n

We benchmarked RIDDEN on various receptor perturbation datasets, including the recent in vivo Immune Dictionary data www.nature.com/articles/s41..., and found good predictive performance. 3/n

We used prior knowledge from omnipathdb.org & perturbation data from clue.io to build models for receptor activity inference. Instead of focusing on receptor expression, we predicted activity from the expression of receptor-regulated genes. 2/n

Our tool, RIDDEN (Receptor actIvity Data Driven inferENce) for predicting #receptor activity from #transcriptomics data is published in
PLOS Computational Biology journals.plos.org/ploscompbiol... 1/n

Thanks all the co-authors, Gema Sanz, @kristofszalay.com and especially Gerold Csendes, who led this project! 5/5

Several popular Perturb-seq based benchmark datasets lack heterogeneity, making it difficult to distinguish between strong and weak models. 4/5

Gene embeddings from foundation models align more closely with gene regulatory networks than with signaling networks, which may underlie their weaker performance in perturbation tasks. 3/5

Even the most trivial baseline (mean of train samples) outperformed recent foundation models, while basic ML models using biologically meaningful features won by a large margin. 2/5

Single-cell foundation models, trained on large-scale scRNA-seq datasets, are increasingly used for post-perturbation RNA-seq prediction.

But how do they actually perform?

Our new paper from @turbine-ai.bsky.social is now out in BMC Genomics. bmcgenomics.biomedcentral.com/articles/10....

1/5

Hello #AACR25 !

PhD/Postdoc in Spatial Omics for Cardiology Post a job in 3min, or find thousands of job offers like this one at jobRxiv!

We have an opening for a PhD/Postdoc in Spatial Omics for Cardiology in Heidelberg 🫀 #job #academicjob #phd #postdoc #cadiology 👇
jobrxiv.org/job/heidelbe...

Pretraining virtual cells is useless the way we do it now

Turns out the way we usually pre-train foundational cell models adds very little information to the system - definitely not enough to make drug effect predictions work. Not what I've expected.
#virtualcells #foundationalmodels #compbio

blog.turbine.ai/p/pretrainin...

easter holiday is coming: 2 revised versions submitted today

You’ve probably heard about how AI/LLMs can solve Math Olympiad problems ( deepmind.google/discover/blo... ).

So naturally, some people put it to the test — hours after the 2025 US Math Olympiad problems were released.

The result: They all sucked!

Does @iscb.bsky.social have a statement on this? We need to be there to support bioinformatics researchers.

based on some recent papers / presentations, the hill(s) I'll die on:
1) Raw IC50 values != good metrics for comparing cancer drug sensitivity
2) Raw TPMs / counts are meaningless for comparing gene expression similarity
3) Diverging color palettes need a meaningful midpoint.

Postdoctoral fellow - Saez-Rodriguez Group Your group Saez-Rodriguez Research Group Your supervisor Julio Saez-Rodriguez Your role As a postdoctoral fellow in the Saez Rodriguez group, you will develop and apply computational methods and tools...

Post-doc opening to join our lab @ebi.embl.org to develop&apply #bioinformatics and #machine-learning methods to study intra-/extra cellular networks to extract disease mechanisms from #single-cell and #spatial multiomic data - please share 🙏 : embl.wd103.myworkdayjobs.com/EMBL/job/Hin...

Flier for the ICSB2025 conference in Dublin 5-9 October 2025

We will be hosting the International Conference on Systems Biology (ICSB) in Dublin this year – 5-9th October icsb2025.com/registration/ #ICSB2025

It's been over 24 hours since #AdventOfCode 2024 ended!

This AoC season:
- 273,313 users collected at least one star
- 3,654,949 stars were collected

Since 2015:
- 779 users have all 500 stars
- 1,077,226 users collected at least one star
- 23,170,305 were collected

holiday season is coming: 2 papers submitted today

Thanks to all the coauthors, Eszter Varga, Daniel Dimitrov, @juliosaezrod.bsky.social , László Hunyady and especially to first author Szilvia Barsi who led this project from start to finish! 6/6

Using an IO dataset of nivolumab-treated renal cancer patients, we found that:
- PD1 or PDL1 expression didn't predict survival
- PD1 activity (from RIDDEN) did associate with survival.
This shows that receptor activity could be a more effective biomarker than expression. 5/n

RIDDEN’s receptor-specific signatures align with regulons of transcription factors downstream of receptors. 4/n

We benchmarked RIDDEN on various receptor perturbation datasets, including the recent in vivo Immune Dictionary data nature.com/articles/s41..., and found good predictive performance. 3/n

We combined prior knowledge from OmniPath omnipathdb.org with LINCS L1000 perturbation transcriptomics clue.io to build models that infer receptor activity. Instead of focusing on receptor expression, we predict activity using the expression of receptor-regulated genes. 2/n