Artur Szałata
@arturszalata
Followers
776
Following
1K
Media
11
Statuses
165
Machine learning for molecular biology. @ELLISforEurope PhD student @fabian_theis lab. @EPFL_en alumnus.
Joined August 2020
Interested in predicting transcriptomic effects of perturbations? Check out our @NeurIPSConf D&B spotlight living perturbation prediction benchmark & new drug perturbation dataset: - paper: https://t.co/GBZZGmbtes ! - benchmarking platform: https://t.co/HsXMMVAMrc🧵1/8
3
5
38
An interesting article in @QuantaMagazine about our recent work on why external filters will never work for AI Safety/Alignment
quantamagazine.org
Large language models such as ChatGPT come with filters to keep certain info from getting out. A new mathematical argument shows that systems like this can never be completely safe.
2
1
5
Hello London! 👋 Our vehicles are now driving in London as we prepare for commercial service in 2026.
240
342
4K
A number of people are talking about implications of AI to schools. I spoke about some of my thoughts to a school board earlier, some highlights: 1. You will never be able to detect the use of AI in homework. Full stop. All "detectors" of AI imo don't really work, can be
Gemini Nano Banana Pro can solve exam questions *in* the exam page image. With doodles, diagrams, all that. ChatGPT thinks these solutions are all correct except Se_2P_2 should be "diselenium diphosphide" and a spelling mistake (should be "thiocyanic acid" not "thoicyanic") :O
945
3K
17K
@elocinationn Great analysis... many of these problems are why I started blogging... for example,
1
2
7
Excited to share Nona: a unifying multimodal masking framework for functional genomics. Models for DNA have evolved along separate paths: sequence-to-function (AlphaGenome), language models (Evo2), and generative models (DDSM). Can these be unified under a single paradigm? 1/15
5
51
228
🔥It takes domain expertise to navigate a field and assess progress. Building (mediocre) AI models is easy—the challenge is rigorous evals. Beware of new benchmarks that are half baked or exclusion of SOTA to make performance seem impressive. THIS IS A PERVASIVE ISSUE IN AI x BIO
3
4
47
Interested in longitudinal spatial (multi)omics applied to an unmet clinical need? Want to work in academia, together with industry to see your research translated into a drug development context? Join us to investigate early signatures of fibrogenesis https://t.co/hddILOK2cX
1
4
6
The bottleneck for deep skill isn't usually intelligence, but boredom tolerance. Learning has an activation energy: below a certain skill threshold, practice is tedious, but above it, it becomes a self-sustaining flow state. The entire battle is persisting until that transition.
122
906
7K
@willyakah @ron_alfa Anyone who thinks ML engineering is the main bottleneck in building better cellular models is going to be sorely disappointed. Better problem formulation & the right data collection strategy will win any day of the year.
5
24
237
🚀 Our new Science paper is out (w/ B DeMeo, D Burkhardt, A Shalek, M Cortes): https://t.co/NSfnCblwHh We show that active learning + transcriptomic perturbations can guide which exps to run next, boosting phenotypic hit rates >13x. AI not just predicting bio, but designing it.🔁
8
78
399
Active learning with DrugReflector beats SotA in phenotypic hit-rate for virtual screening. Includes a sc pert dataset with 10 lines and 104 compounds. Out in @ScienceMagazine! Grateful to @cellaritybio & @fabian_theis for the opportunity to contribute! Link below
1
0
2
Readers responded with both surprise and agreement last week when I wrote that the single biggest predictor of how rapidly a team makes progress building an AI agent lay in their ability to drive a disciplined process for evals (measuring the system’s performance) and error
deeplearning.ai
DeepLearning.AI | Andrew Ng | Join over 7 million people learning how to use and build AI through our online courses. Earn certifications, level up your skills, and stay ahead of the industry.
85
291
2K
I'm going to start adding "humanity's last" to every new paper. "Humanity's last Gaussian process", "Humanity's last batch size"...
6
10
167
This is insufficient verification of perturbation prediction. These are like positive controls. Plenty of correlation in gene expression. High likelihood of extensive false positives if perturbations are done systematically for all genes due to correlation != causality.
Because it is multimodal, Nephrobase goes beyond clustering. In silico gene perturbations simulate biology: CCL2 → immune chemotaxis VCAM1 → endothelial migration GDF15 → growth regulation SOX4 → morphogenesis + energy metabolism
3
8
86
This is a nice dataset and model, but this is an apples & oranges comparison. The Nephrobase model is literally trained on cell labels with a supervised loss but I'm pretty sure the other models are totally unsupervised. Guess which will do better at separating cell labels?
In both human & mouse, Nephrobase produces tight, biologically correct clusters. ARI 0.82 vs ~0.3–0.5 for others Perfect batch mixing This works because it integrates multispecies + multimodal inputs into one latent map.
2
5
38
@growing_daniel “The best minds of my generation are thinking about how to make people click ads.”
7
35
698
Wow this is a disappointingly bad take/comic. To all the students, PhD or earlier: If you spend a week trying out things that don't work, you didn't do nothing! If you ran your experiments properly, you should have confidence in the result, and at least some intuition as to why
37
46
780
Today in @ScienceMagazine, we report a new DNA editing technology to seamlessly write massive changes into the right place in the human genome. The reason gene editing hasn't transformed human health is that current gene editing technologies like CRISPR are very limited. The
What if we could universally recombine, insert, delete, or invert any two pieces of DNA? In back-to-back @Nature papers, we report the discovery of bridge RNAs and 3 atomic structures of the first natural RNA-guided recombinase - a new mechanism for programmable genome design
76
673
3K