Mentor, scientist & engineer. Having fun in @slavovlab.bsky.social and Parallel Squared Technology Institute @parallelsq.bsky.social with biology & single-cell proteomics.
https://nikolai.slavovlab.net
Nikolai Slavov
Loading...
This large new study finds that performance gains from single-cell transcriptomic models plateau well before the largest training set sizes.
Single-cell models show no clear data scaling laws:
www.nature.com/articles/s41...
This provocative perspective challenges the notion that the brain can be usefully modeled like a computer.
Metaphors can be useful, but they can also be vague, incoherent and misleading — failing to capture animal cognition, for example.
1/
I never expect counting RNAs in ever more single cells to scale model performance far.
Now, the evidence is in.
It suggests that scaling RNA counts alone is insufficient.
Scientific progress depends on measuring the variables that regulate biological systems.
1/
It's limited by asking the right questions and making the direct measurements that can constrain the solution space.
Neuroscience needs to stop treating the brain as if it is a computer.
www.nature.com/articles/d41...
It is interesting to see AI enthusiasts discover that simple linear models outperform large models in both computational efficiency and predictive accuracy.
The right model depends on the task and objectives.
Achieving many objectives is not limited by model size or architecture.
1/
"No act of kindness, no matter how small, is ever wasted."
-- Aesop
Be kind
Be generous
Mass spectrometry proteomics loves benchmarks.
But an important one is rare:
- Accuracy of proteome quantification when using short LC gradients.
Fast MS instruments can quantify 7 - 9k proteins from 200ng samples using short separation times affording the analysis of 200 – 500 samples / day.
1/
The program for #SCP2026 is online.
The talks span diverse technologies and biological questions.
Join the discussions, flash talks, and hands-on workshop.
🗓️ July 14-16, 2026 | Boston, USA
🔗 Check out the full program:
single-cell.net/proteomics/s...
⬛ Have you seen accuracy benchmarks for this workflow ?
If such a dataset exists, I'd love to see it.
If not, it's one of the most important benchmarking experiments still missing in high-throughput proteomics.
The figure is from: Nat Methods 20, 375–386 (2023).
www.nature.com/articles/s41...
