🤞 We hope these results motivate the community to rethink narrow-task benchmarking and narrow-task applications, and to pivot towards open-ended biological inference in zero-shot settings.
@saigenci.bsky.social
We call this pattern #finetuning_masking: supervised adaptation on narrow tasks can compensate for weaker pretraining and mask representational quality.
This matters because many current benchmarks may not be measuring the properties that cell foundation models are intended to provide.
Reproducibility is key to science. In computational biology, we routinely manipulate high-dimensional data (spatial, single-cell, bulk) through filtering, normalisation and transformation. Capturing those steps clearly improves reproducibility and transparency.
tidyomics.github.io/tidyomicsBlo...
📢 📃 Finetuning masking challenges narrow-task evaluation of cell foundation models
shorturl.at/dcz3l
We tested foundation model architectures across extreme reductions in pretraining scale. We found that standard narrow tasks were insensitive to pretraining scale, up tp 128-fold reduction!
New #tidyomics blog post! 🎉
#tidySummarizedExperiment greatly simplifies your omics workflows, giving tidyverse capabilities.
We have improved its performance 30x on large datasets with tens of thousands of samples!
tidyomics.github.io/tidyomicsBlo...
Plz share!
@mpibiochem.bsky.social
This work started as an internal resource and has now become public with helpful APIs. It was 3+ years in the making, with a massive Thanks to Mengyuan Shen and the whole Team!
We hope you find it useful 🙏
@saigenci.bsky.social
@wehi-research.bsky.social
@bioconductor.bsky.social
#cellNexus is out! -> "Quality control, annotation, aggregation and analytical layers for the Human Cell Atlas data"
- QC layers
- Consensus cell annotation layers
- Pseudobulk aggregation of the whole human atlas
- R/Python APIs cellNexus.org
Preprint: www.biorxiv.org/content/10.6...
Juan Henao has published a post on the #tidyomics blog introducing *omicslog*, which tracks and stores data manipulation directly from tidy R (and base R) workflows, in the data object itself!
Inspired by `tidylog`.
Feedback welcome!
@bioconductor.bsky.social
First #tidyomics hackathon at #EuroBioC2026 in Turku!
tidyomics.github.io/tidyomicsBlo...
Join us at Euro Bioc (eurobioc2026.bioconductor.org) for a dive into our ecosystem, where you can learn about it and, hands-on, improve and extend it.
Plz share! 🙏
@bioconductor.bsky.social
Another Bioc 3.23 update: new dplyr function support for a number of tidyomics packages @stemang.bsky.social
Adding slice_sample, *_head, *_tail, *_min, and *_max to:
- tidySingleCellExperiment
- tidySeurat
- plyranges
The dplyr descriptions for these:
dplyr.tidyverse.org/reference/sl...
