Notes

The Latent Language of Morphogenesis

Reading the Evoscope Autoencoder as a Mesoscopic Generator

Research Note · June 2026

A conceptual reading of the Evoscope v0.9.1 autoencoder. This note explores how latent representations may be interpreted as mesoscopic descriptions linking morphology and regulation.

Introduction

This note emerged from the development of Evoscope 0.9.1, a computational framework designed to study how multicellular morphologies relate to underlying regulatory states (see the Research page for project details and associated publications). In Evoscope, synthetic tissues are generated by populations of interacting cells carrying simplified regulatory programs ("genomoids"), and a convolutional autoencoder is trained to learn compact representations of these emergent morphologies while simultaneously predicting their regulatory states.

The present note is not a technical description of Evoscope itself. Rather, it explores a conceptual question raised by the model: what kind of information is actually stored within the latent representations that connect morphology and regulation?

Figure 1. Representative Evoscope morphology. Example of a synthetic multicellular morphology generated by Evoscope 0.9.1. While the underlying genomoid regulatory states are not directly visible, they contribute to the emergence of the spatial organization observed here. The latent representations discussed in this note can be viewed as mesoscopic descriptions linking these hidden regulatory processes to their observable morphological outcomes

Autoencoders are commonly described as dimensionality reduction algorithms. Their purpose is usually presented in practical terms: compress a high-dimensional input into a smaller latent representation and then reconstruct the original data as accurately as possible.

The Evoscope v0.9.1 autoencoder can certainly be understood in this way. Given a morphology snapshot represented as a one-hot encoded spatial grid, the model learns a compressed latent representation capable of reconstructing the original morphology.

However, such a description only captures part of what is happening.

When examined in the context of Evoscope and Foldomics, the autoencoder can be interpreted as a mechanism that searches for a mesoscopic description of a biological system: a representation that lies between morphology and regulation.

The architecture suggests a simple but powerful question: can a compressed representation of morphology become informative enough to reconstruct not only form, but also the underlying regulatory state? This question motivates the interpretation developed in the following pages.

The Encoder: From Morphology to Latent Description

The encoder receives a morphology snapshot as input. Through successive convolutional layers, spatial resolution is progressively reduced while the number of feature channels increases. Conceptually, the transformation can be represented as:

Morphology Snapshot

↓

32 feature maps

↓

64 feature maps

↓

128 feature maps

↓

Latent Space z

The encoder therefore performs two operations simultaneously:

First, it reduces the dimensionality of the original morphology.
Second, it constructs increasingly abstract descriptions of the observed spatial organization.

In this view, the encoder is not merely compressing information. It is generating candidate descriptions of the observed system. The resulting latent vector z can be interpreted as a compressed hypothesis about the morphology.

The Latent Space as Functional Compression

The latent space is often described as a compressed version of the input. In Evoscope, however, this interpretation is incomplete.

The latent coordinates are not required only to reconstruct morphology. They must also support the prediction of regulatory variables through a second branch of the network.

Consequently, z is not simply a visual compression:

It becomes a functional compression.
Information that is useful only for visual reconstruction may disappear.
Information that is useful for predicting regulatory state may be preserved.

The latent space therefore evolves toward a representation that balances multiple constraints simultaneously. This is the origin of its mesoscopic character.

Figure 2. Annotated reading of the Evoscope v0.9.1 autoencoder architecture.

Decoder and Gene Head

Two independent observers receive the same latent representation.

The Decoder attempts to reconstruct morphology.
The Gene Head attempts to predict regulatory variables.

Both receive exactly the same input: z. Yet they ask different questions.

The Decoder asks:

"Does z contain enough information to reconstruct the original form?"

The Gene Head asks:

"Does z contain enough information to infer the regulatory state?"

The training objective combines both requirements:

Loss = α · Loss_reconstruction + β · Loss_gene

The latent space is therefore shaped by two simultaneous pressures:

preservation of morphology
predictability of regulation

Neither objective dominates completely.

The final representation emerges as a compromise between them.

Emergence of Mesoscopic Coordinates

An important observation follows from the architecture. Regulatory variables are not directly used as input to the encoder.

The encoder only observes morphology. Nevertheless, the gene-prediction loss propagates backward through the network and modifies the encoder itself.

As training progresses, the encoder is encouraged to preserve those morphological features that are predictive of regulatory state. The latent coordinates therefore become progressively enriched with information that links morphology and regulation.

In this sense, the mesoscopic coordinates are not predefined. They emerge from the negotiation between reconstruction and prediction.

The Poet, the Sculptor, and the Painter

A useful metaphor can help illustrate this process. Imagine three artists. The Encoder is a poet. The Decoder is a sculptor. The Gene Head is a painter. The poet observes a group of models. He studies their shape, posture, appearance, and all visible characteristics.

Instead of creating a sculpture or a painting, he writes a short poem. The poem is compressed. It cannot contain every detail. It must contain only what is essential.

This poem is the latent representation z. The poem is then sent to the sculptor and the painter. The sculptor has never seen the original models.

He reads the poem and attempts to reconstruct their form.

The painter has never seen the original models either. She reads the same poem and attempts to reconstruct their colors and defining traits.

The poet then compares their work to the original models. If important information has been lost, he rewrites the poem.

The process repeats. Gradually, the poem becomes capable of evoking both form and regulation.

The latent space is therefore neither a sculpture nor a painting. It is a language.

A compact description from which multiple aspects of reality can be reconstructed.

Figure 3. The Poet, the Sculptor and the Painter: a metaphorical interpretation of latent-space learning.

Conclusion

From a conventional machine-learning perspective, the Evoscope autoencoder is a convolutional autoencoder equipped with a regulatory prediction head. From a mesoscopic perspective, it can be viewed differently. The encoder becomes a generator of compressed descriptions.

The latent space becomes a language rather than a storage container. The decoder and gene head become independent interpreters of that language. In this interpretation, z is neither morphology nor regulation. It is a mesoscopic description capable of evoking both.

The autoencoder therefore becomes more than a dimensionality reduction algorithm.

It becomes a generator of mesoscopic representations.

Credits

Author: Luca Zammataro

Lunan Foldomics LLC

Prepared with the assistance of generative AI tools for drafting, illustration, and editorial support.

Page updated

Google Sites

Report abuse