Novacore

Learning the biological structure behind cellular images

Understanding biology in motion doesn’t start with time - it starts with relationships.
How genes interact.
How pathways connect.
How perturbations reshape cells in ways that reflect deeper molecular structure.

NOVA is our representation-learning framework that bridges cellular morphology with biological connectivity. It learns embeddings from microscopy images that are not just visually descriptive, but aligned with real gene-gene relationships, enabling downstream models to reason with biology rather than appearance alone.

Where FORM simulates treatment effects and PointMIL explains morphological evidence, NOVA provides the foundation: a latent space where cells belonging to related perturbations naturally cluster together. It captures the shared structure behind how different genes, pathways, and interventions manifest in morphology - giving us a biologically meaningful map from images to mechanism.

NOVA is where morphological intelligence begins.

What NOVA is

NOVA combines two key ideas:

1. A vector-quantised transformer
   that learns stable, discrete prototypes of cellular morphology directly from multi-channel fluorescence images.
   
   

2. Graph-regularised relational learning
   that aligns the embedding space with curated biological interaction networks such as STRING and CORUM.

The result is a latent space where visual structure and biological structure reinforce one another.
Cells look similar because they are similar — connected by pathways, complexes, or functions — not merely by pixel-level likeness.

Why this matters

Most self-supervised microscopy models optimise for reconstruction or contrastive alignment. They learn images, not biology. As a result, they may cluster cells by noise, stain intensity, or imaging artefacts - missing the underlying relationships that matter for mechanism.

NOVA solves this by explicitly incorporating biological connectivity into representation learning. It does not use labels, supervision, or hand-crafted rules. Instead, it nudges related perturbations toward shared regions of latent space, ensuring the model:

• respects known gene–gene relationships,
  
  

• discovers new morphological connections,
  
  

• stabilises learning under limited data, and
  
  

• generalises across perturbations and experimental conditions.

This gives us a representation that reflects mechanism, not just morphology.

Inside NOVA

1. A discrete vocabulary of morphological patterns

NOVA encodes each image using a vector-quantised transformer that learns a codebook of morphological prototypes. These prototypes capture recurring shapes, textures, and structural arrangements across channels.

This discrete bottleneck:

• stabilises training,
  
  

• encourages semantically meaningful embeddings, and
  
  

• forces the model to represent cells using combinations of learned morphological building blocks.
  
  

It’s a structured way of saying:
“These are the visual elements biology reuses.”

2. Biological relational regularisation

NOVA doesn’t assume visually similar cells are biologically similar.
Instead, it aligns latent space with curated interaction graphs, so cells belonging to related perturbations are gently encouraged to cluster.

This introduces a lightweight biological prior:

• perturbations that share pathways → closer embeddings
  
  

• perturbations in the same complex → closer embeddings
  
  

• perturbations with known genetic interactions → closer embeddings

This builds biological structure directly into the representation.

What NOVA enables

Biology-aware representation learning

Embeddings reflect real gene and pathway relationships, not just pixel-level similarity.

Improved retrieval of related perturbations

Cells affected by related genes cluster naturally, enabling mechanism-of-action inference from morphology.

Stronger downstream models

FORM, MorphoSense, and other temporal or generative frameworks benefit from NOVA’s biologically structured latent space.

Signal amplification under limited data

NOVA extracts relationships that might be invisible to purely visual models.

Cross-context generalisation

Biological alignment improves robustness across perturbations, plates, and imaging conditions.

How NOVA supports SentiCORE

NOVA is the representation backbone of the SentiCORE platform. It provides:

A biologically structured embedding space

giving every downstream model a foundation grounded in connectivity and mechanism.

Shared latent geometry across perturbation datasets

making cross-study and cross-lab analysis more coherent.

Better interpretability

by organising morphology according to known biological relationships.

Stronger simulations

because FORM and future 4D models can build on embeddings that already respect biological structure.

NOVA ensures that our platform doesn’t just see cells - it understands how they relate.

What comes next

NOVA lays the groundwork for all higher-level reasoning in Sentinal4D.
Our next steps extend these embeddings into:

• 4D morphodynamic models,
  
  

• perturbation graphs enriched with time,
  
  

• multi-phenotype integration, and
  
  

• large-scale unified morphology representations across cell types and treatments.

NOVA teaches us the relationships behind morphology. Everything else builds on that.

Closing thought

Biology is a network long before it is an image. NOVA captures that structure, learning embeddings that reflect not only how cells look, but how their underlying perturbations connect.

If your work depends on understanding relationships between genes, pathways, or treatments directly from microscopy data, NOVA offers a biologically aligned, representation-first foundation. We’d be happy to talk.