The challenge
A pharmaceutical manufacturer needed an unbiased way to characterise disease states and drug effects at scale from mitochondrial microscopy data. Predefined class labels impose the researcher's assumptions on the data. The business driver was acceleration of the drug-discovery process.
At a glance
StackPython, scikit-learn, TensorFlow/Keras, Azure ML
MethodsK-means, DBSCAN, GMM, t-SNE / UMAP
WorkstreamsUnsupervised clustering + supervised classification
What we did
SWICON built an image-processing and feature-embedding pipeline over fluorescence microscopy features and segmented object property files, producing fixed-dimension embeddings (mean, median, covariance) from variable-length inputs. Three unsupervised approaches were evaluated: K-means with elbow-method cluster selection and multi-initialisation stability checks, density-based DBSCAN and Gaussian mixture models, with t-SNE and UMAP for visualisation and silhouette scoring for validation. The pipeline was shared with a parallel supervised classification workstream in the client's Azure ML environment.
The result
The natural morphology groupings surfaced by the clustering were validated for stability across initialisations and interpreted in biological context together with the client's domain experts, giving the research team a data-driven categorisation to work from rather than a hand-defined one.
