Marker Imputation for Spatial Proteomics (MISP) based imaging data
Proteomics based multiplexed imaging can be limited due to the i) availability/limit of the markers in the panel, ii) imaging artefacts, and iii) high experimental costs. An example is the BRUCE GBM MIBI dataset, where consecutive tissue sections were imaged with two marker panels; immune and tumor panels. While both panels have an overlap of around 20 markers, each contain specific markers towards immune cell types and tumor cell types respectively. Inspired from 7UP, we implement a simple explainable XGBoost based imputation of unseen markers from a overlap of existing antibody markers between two different panels.
This implementation can work as a proof of concept or a baseline to build better generative models.
On a test split of the immune panel with over 29 different markers, the model is able to achieve the following metrices
| Avg. MSE | Avg. PCC |
|---|---|
| 0.0003 | 0.643 |
Implemented over the Owkin GBM AI Hackathon 2025