Home
What Fed-MaxFuse does¶
Fed-MaxFuse is a federated implementation of MaxFuse, a diagonal integration method for single-cell data. It matches individual cells across two modalities that share no cells and almost no features, and it does so without the two modalities ever being pooled in one place.
The motivating case is spatial proteomics meets single-cell sequencing. A 46-marker CODEX panel and a whole transcriptome have only a handful of genuinely comparable features between them, and in practice the two datasets sit in different institutions that cannot share raw data under GDPR or HIPAA.
The idea in three sentences¶
Only two operations in MaxFuse genuinely need both modalities at once: cell matching and canonical correlation analysis. Everything else (graph construction, meta-cells, fuzzy smoothing, propagation, final embeddings) is local to one dataset. Fed-MaxFuse moves those two operations to a coordinating server that only ever sees fuzzily-smoothed, L2-normalised representations, and keeps the rest on the nodes.
Does it work?¶
Across 2,600 experimental runs, federated Fed-MaxFuse is statistically indistinguishable from centralized MaxFuse on the metrics that measure matching geometry:
| Metric | Mann–Whitney U p | Verdict |
|---|---|---|
| FOSCTTM | 0.35 | no significant difference |
| FOSKNN | 0.88 | no significant difference |
| Steadiness (meta & non-meta) | 0.31 / 0.53 | no significant difference |
| Cohesiveness (meta & non-meta) | 0.85 / 0.47 | no significant difference |
It does lose on label accuracy, most visibly on the heterogeneous Tonsils dataset. The gap is attributable to federated batching not being implemented, rather than to federation itself. See Comparison with Centralized MaxFuse.
Where to start¶
-
New here? Start with Introduction, then Architecture & Data Flow.
-
Want to run it? Go to Installation and the Quick Example.
-
Want the maths? Read Federated Matching and Iterative Federated CCA.
-
Evaluating for privacy? Read Security & Privacy first, including what is not guaranteed.
Status and scope¶
Licensing status is unresolved
No licence is currently granted for this code. A previous MIT declaration has been withdrawn pending review of how the upstream MaxFuse licence applies. See Licensing Status before using any part of this work.
Fed-MaxFuse is research software produced for an MSc thesis at TU Delft. It is complete and tested for the two-modality case on the two benchmark datasets described in these docs. It is not a production federated-learning deployment: it assumes honest-but-curious participants over an assumed-secure channel, and provides no formal privacy guarantees. Read Security & Privacy before drawing conclusions about what it protects.