Morphomil

MorphoMIL: reading treatment response in the shape of cell populations

In every well of a plate, cells share the same treatment and genetics – but they rarely share the same response. Some round up, some spread, some barely change at all. That heterogeneity is where a lot of the biology – and the risk in drug development – actually sits.

MorphoMIL is our way of turning that variation in 3D cell shape into structured, interpretable data at the population level. Developed in collaboration with the Institute of Cancer Research and recently published in Cell Systems, MorphoMIL combines geometric deep learning with attention-based multiple instance learning to profile how treatments reshape entire distributions of cell morphologies, not just their average.

What MorphoMIL is and what it does

MorphoMIL starts from 3D point-cloud representations of single cells (and nuclei) and learns how a whole population of shapes reflects a given treatment or pathway state.

Instead of assuming all cells in a well behave the same, MorphoMIL:

• Treats each field of view as a bag of cells, each with its own 3D shape signature.
  
  

• Learns which subsets of cells carry the strongest treatment signal, and which are bystanders.
  
  

• Uses an attention mechanism to assign each cell an importance weight, making it clear which morphologies drive a particular prediction.

This means MorphoMIL can:

• Distinguish compounds that act on similar pathways by how they reshape the mix of cell shapes in a well.
  
  

• Capture heterogeneous responses in genetically identical cells – including rare responder or non-responder subpopulations.
  
  

• Link those population-level shape patterns back to underlying signalling states and protein networks, providing hypotheses about mechanism of action.

All of this happens in a weakly supervised way: the model sees only treatment labels at the well level and discovers the relevant cell-level phenotypes automatically.

How MorphoMIL strengthens the Sentinal4D platform

Within Sentinal4D’s SentiCORE platform, MorphoMIL is the engine that answers a specific question:

“How does this treatment reshape the population of cell states – and which subpopulations matter most?”

Concretely, MorphoMIL helps our partners to:

• See beyond the average
  Instead of a single summary feature per well, MorphoMIL reveals which fractions of the population change, in what way, and how strongly they contribute to a treatment signature.
  
  

• Quantify heterogeneity and risk
  We can flag treatments where only a minority of cells adopt the desired morphology, or where a distinct resistant subpopulation emerges – critical information for de-risking a programme early.
  
  

• Connect morphology to pathways and targets
  By scoring gene knockdowns or new compounds against learned “shape signatures” of reference drugs, MorphoMIL helps infer pathway involvement and potential protein interactions from imaging alone.
  
  

• Provide interpretable evidence for teams and stakeholders
  Because every prediction comes with cell-level attention scores, scientists can inspect the actual 3D shapes that define a phenotype, rather than trusting an opaque classification.
  
  

MorphoMIL complements our other models:

• PointMIL – interpretable classification on single 3D shapes.
  
  

• MorphoSense – interpretable analysis of time-resolved 4D shape trajectories.

Together, they allow us to read treatment effects at three levels: per cell, over time, and across the full population.

What comes next

We are now extending MorphoMIL to:

• Larger chemical and genetic screens, where population-level heterogeneity becomes a primary signal rather than a nuisance.
  
  

• Combination treatments, to see how dual or triple perturbations reshape morphology distributions compared with single agents.
  
  

• Patient-derived and primary cell systems, where understanding diversity within genetically similar populations is essential for translational relevance.

Longer term, MorphoMIL-style population modelling will underpin how SentiCORE scores risk and opportunity in new programmes: from early hit triage through to mechanistic follow-up.

Closing thought

Treatments do not affect all cells equally. Even when cells are genetically identical, they diverge – and those divergences often decide whether a candidate succeeds or fails downstream.

MorphoMIL turns that heterogeneity into usable evidence: showing how treatments reshape populations of 3D cell shapes, which subpopulations carry the signal, and how those patterns connect back to pathways and targets.

If you are running 3D phenotypic screens and want to understand not just whether a treatment has an effect, but how consistently it acts across a population of cells, MorphoMIL is built for that question.