Dissecting AIDO.Cell
A project journal on mechanistic interpretability for single-cell foundation models
This is the running log of my thesis project: using Sparse Autoencoders (SAEs) to open up the internal representations of AIDO.Cell, a single-cell foundation model, and then using the extracted features to steer cell identity in interpretable ways.
Each post below covers one slice of the work. The goal is for you to be able to follow the project without me having to explain it from scratch each time we meet. Whenever a result has a long-form write-up in reports/, the post links to it.
Posts
- Project overview — the motivation, the big picture, and what SAEs buy us here.
- SAEs on AIDO.Cell: biological footprint — 64% of SAE features receive significant GO annotations, and they are more specific than raw neurons.
- Steering single features — amplifying or suppressing individual features produces symmetric, biologically coherent expression changes.
- Steering cell identity: CD4→CD8 — a contrastive objective recovers sparse feature combinations that reprogram CD4 T cells toward CD8.
- Latent vs. expression objective — a surprise: the two objectives find nearly orthogonal solutions. The scFM’s internal geometry and its output manifold are not interchangeable.
- Probing regulatory logic — a negative result: AIDO.Cell does not encode TF→target directionality, at least by a propagation test on TBX21.
- Roadmap — what’s next and open questions.
Repo: mech_interp_bio/ · Last updated: April 21, 2026