Fully open source and available now, explore Winnow today:

💻 Read the blog here: bit.ly/3Lq335a

📃 See the paper: bit.ly/43go0FG

📕 Explore the codebase: bit.ly/3JLWqcM

We hope Winnow can become a standard tool to make de novo sequencing results easier to interpret. Feedback is very welcome! We’d love to hear from researchers and practitioners who might want to try Winnow in their own pipelines.

A great collaboration with @tpjenkins.bsky.social and Konstantinos at DTU Bioengineering.

🐍 Download the code from PyPI:
pypi.org/project/winn...

🗞️ Download the pretrained model:
huggingface.co/collections/...

💻 Explore our repo:
github.com/instadeepai/...

Winnow is available today:
📃 Check out our preprint:
arxiv.org/abs/2509.24952

✨ Key features of Winnow:
- DNS score calibration and rescoring
- Lightweight, non-parametric false discovery rate estimation
- Common metrics such as PEP and q-value
- Pretrained general model for zero-shot estimation with InstaNovo
- Extensible framework

Winnow supports these discoveries by improving peptide recall while simultaneously providing accurate error estimates.

That could mean identifying peptides that drive immune responses in diseases like cancer, tracing protein interactions in complex microbial communities, or even sequencing ancient proteins from fossil samples.

Why does this matter? De novo sequencing tools like InstaNovo don’t just recover known biology. They can uncover what’s new and hidden.

Over the past 9 months we developed Winnow. It provides a principled and robust method to estimate false discovery rate in de novo peptide sequencing, where the lack of reference databases makes traditional error control strategies impossible.

A few years ago we set out to build a cutting-edge de novo peptide sequencing model. That became InstaNovo, which set a new benchmark for peptide sequencing.
bsky.app/profile/tpje...