Morphosense

Before a cell metastasises, responds to a drug, or fails a therapy, it does one simple thing: it changes shape.

Those changes do not happen in a single frame. Cells round, spread, send out protrusions, retract them, and cycle through distinct behaviours as they adapt to their environment. MorphoSense is our way of turning those continuous 3D shape changes into interpretable, time-resolved data – so R&D teams can see how a cell responds, not just whether a label changes at the end.

MorphoSense is an interpretable pipeline for 3D cellular morphodynamics – built on multiple instance learning (MIL) and graph transformers – designed to classify variable-length 4D trajectories and highlight the key time points and shape states driving each prediction.

It is how we start turning motion into meaning.

Why 4D cell shape matters

Cell morphodynamics – a cell’s time-dependent “shape-shifting ability” – is central to division, migration, and differentiation. When this shape control goes wrong, cancer cells gain the ability to escape a primary tumour, migrate through tissue, and seed metastases elsewhere in the body.

Historically, most analysis has looked at:

• 2D static images, with readouts like area, eccentricity, or number of protrusions.
  
  

• Static 3D shapes, which add volume and surface geometry but still freeze the cell in time.
  
  

These approaches have been powerful for profiling drugs and signalling states, but they miss a key dimension: how the shape evolves after a perturbation.

For example:

• A drug may cause cells to transiently spread before retracting.
  
  

• Others may trigger oscillations between rounded and protrusive states.
  
  

• Immune cells may switch between fast, exploratory morphologies and slower, more adherent modes as they navigate tissue.
  
  

Those patterns are time series of 3D shapes. To learn from them, we need methods that:

1. Handle variable-length trajectories (because cells are not all tracked for the same duration).
   
   

2. Work with multivariate features derived from 3D point clouds.
   
   

3. Provide inherent interpretability – telling us which time points and which shape states were decisive.
   
   

This is exactly the problem that MorphoSense is built to solve.

What is MorphoSense?

At its core, MorphoSense reframes time series classification as a multiple instance learning problem.

• Each cell trajectory is treated as a bag of instances – one instance per time point.
  
  

• Each instance is a multivariate feature vector (the 3D shape embedding for that time point).
  
  

• The bag has a single label: the drug treatment, tissue environment, or dynamic class.
  
  

Traditional MIL assumes instances in a bag are independent. That is not true for time series: shape states are temporally correlated. MorphoSense addresses this by building a graph over time and then applying a graph transformer with MIL pooling.

Step 1: Temporal + similarity graph

For each cell:

• Nodes = time points.
  
  

• Temporal edges link t → t+1 to capture progression over time.
  
  

• Similarity edges connect non-adjacent time points with similar features, based on cosine similarity above a threshold.
  
  

This gives a cycle-aware temporal graph that encodes both local progression and recurring shape states (e.g. cycles between round and protrusive morphologies).

Step 2: Interpretable graph transformer

The temporal graph is passed through graph transformer layers, which use edge-aware attention to propagate information across time and shape cycles. This allows MorphoSense to capture:

• sustained responses,
  
  

• delayed effects of treatment,
  
  

• recurrent transitions between shape states.

Step 3: Conjunctive pooling for explanations

On top of the transformer, MorphoSense uses conjunctive pooling, a dual-stream MIL head:

• A classification stream outputs logits for each time point.
  
  

• An attention stream learns an importance weight for each time point.
  
  

The final prediction for the trajectory is a weighted sum of instance logits, with weights given by the attention scores.

Crucially, the interpretation for each time point is class-specific and defined directly as:

importance = attention × instance logit

That means:

• The same quantities that produce the final prediction are the explanation.
  
  

• There is no separate explainer network or post-hoc saliency step.
  
  

• You get a ranked list of time points and can see exactly which phases of the trajectory were most important for each class.
  
  

This is fully aligned with our communication strategy: interpretations are sparse, direct, and grounded in the model’s own decision rule.

How MorphoSense fits into Sentinal4D

MorphoSense is one of the core engines inside our platform for 4D phenotypic profiling.

It allows us – and our partners – to:

• See when the biology changes
  For each cell, MorphoSense highlights the specific time points that drive a classification: the onset of a protrusive phase, a collapse into a rounded shape, a sustained spreading state. Scientists can directly inspect these intervals and relate them to known mechanisms.
  
  

• Link dynamic patterns to treatments
  Different drugs often have distinct temporal signatures, not just distinct endpoints. MorphoSense turns these signatures into measurable, interpretable objects at the level of time-resolved 3D shape.
  
  

• Support earlier, defensible decisions
  Because every prediction is accompanied by timepoint-level importance scores, results are transparent to interdisciplinary teams. Instead of asking stakeholders to trust an opaque model, we can show when the model believes the drug has its defining effect – and how that aligns with other assays.
  
  

• Build towards continuous 4D behaviour profiling
  MorphoSense extends the interpretability we achieved with PointMIL on static 3D shapes to full 4D trajectories. Together, they create a consistent language for understanding both single frames and entire sequences.

This is exactly the type of capability we aim to offer as a partner: evidence you can see, test, and trust.

Closing thought

Life happens in motion. With MorphoSense, we are starting to read that motion: time point by time point, trajectory by trajectory, treatment by treatment.

If your work involves 4D cellular imaging – whether in drug discovery, immunology, or beyond – and you want interpretable insight into how cells change shape over time under your treatments, we would be glad to explore how MorphoSense can support your R&D pipeline.