Hi! I am Andrii Buvailo, and this is my weekly newsletter, ‘Where Tech Meets Bio,’ where I talk about technologies, breakthroughs, and great companies moving the biopharma industry forward.

If you've received it, then you either subscribed or someone forwarded it to you. If the latter is the case, subscribe by pressing this button:

Subscribe now

Let’s get to this week’s topic!

In a new provocative paper, "The Genomic Code: The genome instantiates a generative model of the organism," the authors, Dr. Kevin Mitchell, a Neurogeneticist at Trinity College Dublin and Dr. Nick Cheney, a Computer Scientist at the University of Vermont, propose a new way of thinking about genome. Unlike a traditional notion of a “blueprint” of the organism, the authors suggest that it may be a closer to how a deep neural network operates. Evolutionary forces play a role of a “training“ mechanism, while the embryo plays a role of a “decoder.”

I recommend reading this 30+ page article as it entertains imagination at the very least, and potentially, opens avenues for new ways of modeling evolutionary processes.

But if you don’t have time to go through the whole thing, let’s dissect the key aspects of the proposed theory:

According to authors, traditional metaphors like "blueprint" or "program" are inadequate for explaining how the genome encodes an organism's form because they don't capture the complexity of developmental processes.

A better analogy is to think of the genome as encoding a generative model, similar to those used in machine learning.

In this view, the genome contains "latent variables" – sequences that influence biochemical properties and regulatory interactions. These variables form a network shaped by evolution, which guides development through an "energy landscape." This landscape directs cells toward specific fates, much like how Waddington’s epigenetic landscape visualizes cell differentiation.

Image source: https://arxiv.org/abs/2407.15908

Gene expression is regulated by complex interactions, making development like a decoder that unpacks the genome's compressed information over time. Evolution acts as the encoder, refining the network through natural selection to ensure both robustness and the ability to evolve.

This model explains how traits can evolve independently even though many genes influence them. This is due to emergent modularity in the latent variable space, where different traits can be separately manipulated.

Formalizing this concept mathematically allows for better modeling of developmental processes and their evolution, providing a clearer framework for understanding how genotypes translate into phenotypes.

Introduction to the New Theory: Genome as a Generative Model

Traditional Metaphors and Their Limitations

Historically, the genome has been described using metaphors like "blueprint" or "program." These metaphors suggest that the genome directly encodes the organism's form or development in a straightforward, linear manner. For example:

Blueprint: Implies a detailed, direct map of the organism, where each part of the genome corresponds to a specific part of the organism.
Program: Suggests the genome contains a linear set of instructions to be executed step-by-step to build the organism.

However, these metaphors fail to capture the complexity, indirectness, and dynamic nature of how the genome actually influences development. They overlook the interactions, context-dependence, and emergent properties of gene regulation and organismal development.

Introducing the Generative Model Analogy

To better understand the genome's role, we propose a new analogy inspired by machine learning: the genome as a generative model. In machine learning, generative models are used to generate new data instances (like images or text) from a set of underlying parameters, or "latent variables." These models don't store exact replicas but rather capture the essence or structure of the data in a compressed form.

Key Concepts of the Generative Model Analogy

Latent Variables:
- Definition: In this context, latent variables are sequences in the genome that influence the biochemical properties and regulatory interactions of proteins.
- Function: These variables form a complex, interconnected network that determines how genes are expressed and interact during development.
Energy Landscape:
- Definition: The energy landscape is a conceptual model that describes how developmental processes are guided. Imagine a landscape with valleys and peaks, where valleys represent stable developmental states and peaks represent unstable states.
- Function: The genome shapes this landscape, directing cells towards stable states (valleys) through developmental processes.
Development as Decoding:
- Process: Development is akin to decoding the compressed information in the genome. Each stage of development unpacks different parts of the generative model, progressively leading to the organism's form.
- Self-Organization: Development relies on the self-organizing properties of cells and tissues, influenced but not entirely dictated by the genome. This allows for variability and adaptability in outcomes.
Evolution as Encoding:
- Process: Evolution acts as the encoder, shaping the network through natural selection. Genetic variations that enhance survival and reproduction are retained, adjusting the network's "weights" over generations.
- Robustness and Evolvability: The distributed nature of the generative model allows for robustness against mutations and the capacity for evolutionary change. Genetic variation can accumulate and be acted upon by natural selection.

Emergent Properties of the Generative Model

Modularity:
- Concept: Different traits can evolve independently because they are encoded in separate, orthogonal subspaces within the latent variable space.
- Implication: This modularity allows for the independent manipulation and evolution of traits, providing a clearer explanation for how specific traits can evolve without affecting others.
Robustness and Variability:
- Robustness: The network's distributed nature ensures that development is stable and resilient to mutations and noise.
- Variability: The inherent noise in biological processes and the probabilistic nature of gene regulation introduce variability, which is essential for evolution and adaptation.

Formalisability and Practical Applications

One of the significant advantages of the generative model analogy is its ability to be formalized mathematically. Unlike traditional metaphors, this model can be used to create mathematical and computational simulations of developmental processes. This formalization helps in:

Modeling Development: Providing a detailed and accurate representation of how development unfolds from genomic information.
Understanding Evolution: Simulating evolutionary processes and understanding how genetic variation translates into phenotypic diversity.

Evolutionary Examples Better Explained by the New Approach

Polygenic Traits and Omnigenic Model:
- Older Approach: Struggles to explain how traits influenced by many genes can evolve independently.
- New Approach: The generative model analogy explains that traits are encoded in separate, orthogonal subspaces within the latent variable space, allowing polygenic traits to evolve independently without interfering with each other.
Pleiotropy and Independent Trait Evolution:
- Older Approach: Difficult to account for the independent evolution of traits influenced by the same genes.
- New Approach: The new theory explains how combinations of genetic variants can result in emergent modularity. This disentangled representation in latent variable space allows different traits to be selected independently, despite pleiotropy.
Robustness and Evolvability:
- Older Approach: Does not provide a clear mechanism for how organisms can be both robust to mutations and capable of evolutionary change.
- New Approach: The generative model shows that distributed representations and probabilistic gene regulation lead to inherent robustness. This allows for the accumulation of genetic variation, which can be acted upon by natural selection, enhancing evolvability.
Cancer and Maladaptive Phenotypes:
- Older Approach: Offers limited explanations for the emergence of qualitatively novel phenotypes, like cancer.
- New Approach: The generative model’s framework accommodates the idea that changes in the energy landscape can create new attractor states, leading to novel and sometimes maladaptive phenotypes, providing a more comprehensive understanding of such evolutionary phenomena.
Modular and Disentangled Representations:
- Older Approach: Faces challenges in explaining how distinct traits can be manipulated or evolve separately.
- New Approach: By showing how latent variables are encoded in a modular and disentangled fashion, the new model demonstrates how specific traits can be independently modified and selected for during evolution.

Conclusion

The new theory of the genome as a generative model, if accepted as viable, offers a robust framework for understanding the intricate genotype-phenotype relationship.

By conceptualizing the genome as encoding latent variables within a complex regulatory network, this model provides a detailed explanation of how genetic information is progressively decoded during development.

This approach elucidates the non-linear, context-dependent nature of gene regulation and the emergent properties of developmental processes.

For example, consider the development of limb structures in vertebrates. Traditional metaphors would struggle to explain the precise yet variable nature of limb formation across species. The generative model analogy, however, can accommodate this complexity.
Latent variables in the genome influence the expression and interaction of key developmental genes like Hox genes, which in turn shape the morphogenetic energy landscape.
This landscape guides the spatial and temporal patterns of cell differentiation and tissue morphogenesis, resulting in the formation of limbs.

Moreover, the generative model's formalizability allows for the creation of mathematical simulations to predict and analyze these developmental outcomes. By adjusting the latent variables and network weights, researchers can simulate evolutionary scenarios to study how limb structures could evolve under different selection pressures. This level of precision and adaptability makes the generative model a superior framework compared to traditional metaphors, providing deeper insights into the mechanisms of development and evolution.

Where Tech Meets Bio

Is Genome a Generative Model of the Organism?