Latent Manifold Navigation and Evolutionary Priors for Novel Biopharma

For sixty years, drug discovery has chased what evolution has already sampled. We argue that the most therapeutically valuable molecular states lie beyond natural selection's walk — mathematically probable, structurally viable, yet entirely unsampled by biological evolution.

Directed evolution inspired this journey — temporally advancing the selective process. However, a better shovel is still a shovel. At Mesengenic, we utilise evolutionary datasets to form Bayesian priors, enabling us to build upon rather than search within.

Evolution as a Constrained Random Walk

When viewing evolution as an agent which exerts selective pressure, the process can be effectively conceptualised as a constrained random walk across a high-dimensional fitness manifold. Agnostic to domain and scale, Evolutionary Density operates as a continuous probability field — biasing mutational flow toward epistatic peaks that are mathematically probable yet remain entirely unsampled by natural biological evolution.

The Unsampled Frontier

Natural selection traverses only a fraction of the viable sequence space. Deep generative models, such as Hierarchical Variational Autoencoders, offer a robust mechanism to bypass the traditional bottleneck of high-throughput labeling by treating natural evolutionary density as an explicit prior for fitness.

If you could navigate any point on the fitness manifold — including states evolution never reached — and predict structural viability before experimental validation, you would compress R&D timelines by orders of magnitude. Evolutionary density gives you the prior. VAEs give you the compass. Causal inference gives you the direction.

Two Goals

Mesengenic exists with two explicit objectives:

1. Shorten R&D timelines — by skipping non-functional sequence spaces and focusing computational search budgets on fine-tuning complex macromolecular functions.
2. Expand out-of-distribution capabilities — by identifying Stable Abiological Manifolds: structural geometries evolution never sampled, yet the underlying scaffold is mathematically robust enough to physically sustain.

The Probability Field

Evolutionary Density of a given sequence family operates as a continuous probability field that inherently biases mutational flow toward epistatic peaks. By moving beyond traditional discrete one-hot encodings toward a continuous Latent Kernel approach, a Gaussian Process can map and measure topological distances directly across the VAE's learned manifold.

VAEs as Structural Compass

Deep generative models treat natural evolutionary density as an explicit prior for fitness. This mathematical integration equips the engineering pipeline with an algorithmic structural compass — allowing researchers to skip non-functional sequence spaces entirely and focus search budgets on fine-tuning complex macromolecular functions.

Key capabilities:

• Latent Kernel Navigation — Gaussian Process distances measured directly on the VAE manifold, not in sequence space
• Prior-Guided Exploration — evolutionary density biases sampling toward viable regions without exhaustive labeling
• Structural Compass — algorithmic direction through high-dimensional fitness landscapes

Case Study: P411-HF Scaffolds

The practical utility of this framework is illustrated in engineering P411-HF scaffolds for new-to-nature enzyme catalysis. A VAE-informed generative model identifies structural tunnels within the latent space — mathematically bridging standard P450 structural motifs with novel geometric arrangements required for specialized chemistry.

Increased steric bulk of a chemical precursor establishes a rigid physical boundary condition, collapsing a chaotic high-dimensional reaction space into a single, highly enantioselective output. Enantioselectivity and substrate specificity are enforced as non-negotiable mathematical constraints rather than random byproducts of empirical screening.

Biochemical Firewalls

Extending this methodology yields an architecture for designing Biochemical Firewalls within targeted prodrug therapies — engineering specialized variants that remain entirely silent until triggered by specific synthetic cues, functioning orthogonally to native metabolism.

Beyond Natural History

When engineering systems with no natural history, the primary bottleneck is defined by the severe ruggedness of the underlying fitness manifold. This introduces an acute tension between driving functional innovation and preserving the structural integrity of the underlying scaffold.

Defining the Manifold

Stable Abiological Manifolds represent precise structural geometries that natural evolution has never sampled, yet the underlying scaffold is mathematically robust enough to physically sustain. By navigating these unmapped topological domains, the generative model can quantitatively predict how far a viable configuration can be structurally deformed to accommodate novel function before triggering catastrophic collapse of the global fold.

Structural Guardrails

Traditional screening relies on trial-and-error library creation — high risk of non-functional sequences and fold collapse. Generative latent design resolves this friction by establishing mathematical Structural Guardrails directly within the learned representation space.

Mutational Budget Optimization

Advanced generative architectures stretch the boundaries of the traditional mutational budget by shifting focus from purely natural evolutionary distributions toward identification of Stable Abiological Manifolds. This circumvents the need for direct evolutionary precedents, allowing predictive design of highly altered viable states with scaffold survival guaranteed.

Bridging stochastic dynamics and applied macromolecular design allows engineers to navigate highly rugged trade-offs systematically — unlocking entirely new-to-nature abiological functions while embedding structural guardrails directly into the design pipeline.

Hierarchical VAE Tiers

The Mesengenic AI framework deploys a specialized suite of Hierarchical Variational Autoencoders as the core inference engine — operating across high-dimensional frustrated spin frameworks for algorithmic tissue engineering.

VAE v1 — The Developmental Prior

The Seed VAE builds Gaussian Mixture Models of healthy developmental populations. These compressed representations serve as a Developmental Prior — anchoring the latent space to biophysical constraints and mitigating catastrophic forgetting in static neural architectures.

VAE v2 — Latent Library Design

The second tier maps exact boundaries of structural viability in sequence space. By navigating Stable Abiological Manifolds, VAE v2 optimises sample efficiency — avoiding non-functional space while unlocking new-to-nature functions.

VAE v3 — The Causal Engine

The Causal Engine introduces directional asymmetry to resolve frustrated regulatory couplings. By analysing Jᵢⱼ coupling directionality in massive perturbation datasets, the engine triggers the growth of specialised latent branches through Symmetry-Breaking events.

Symmetry-Breaking Logic

Unlike standard autoencoders, the Mesengenic architecture is dynamic. By monitoring reconstruction error across the manifold, the network initiates Symmetry-Breaking events — branching new latent layers to mirror differentiation into specialised lineages.

Computational Light Patterns

A developmental hierarchy serves as an explicit structural template. Target perturbations identified by the Causal Engine function as computational light patterns — driving an initially simple, low-parameter neural unit to self-organize into a highly specialized, maximum-efficiency network architecture.

By treating the Causal Regulatory Network as the artificial DNA of the system, this methodology permits the autonomous growth of complex neural networks from a single computational unit — bypassing the arbitrary limitations of hand-coded deep architectures.

From Seed to Structure

The model initiates as a low-parameter Seed VAE and undergoes directed Symmetry Breaking orchestrated by the VAE v3 Causal Engine. The name Mesengenic — derived from embryological mechanics denoting growth from the middle layer — functions as a guiding parameter when exploring principles of macro-scale collective organisation.

The Markov Random Field

A defining frontier in computational biology involves leveraging generative machine learning to decode the Causal Intelligence inherent within self-organising biological systems. These systems can be mathematically modeled as a high-dimensional Markov Random Field — a topology wherein component spins explicitly map out and define the continuous energy landscape governing system state.

The energy of a system state is defined by component interactions and external perturbations:

E(s) = − Σ Jᵢⱼ sᵢ sⱼ − Σ hᵢ sᵢ

Frustrated Spins

A primary challenge lies in resolving Frustrated Spins embedded within the systemic Jᵢⱼ coupling matrix. Within complex regulatory landscapes, frustrated antiferromagnetic couplings (Jᵢⱼ < 0) are not mathematical anomalies — they represent the critical, high-entropy decision-forks of system state transitions.

The paradigm shift: moving from static, undirected couplings toward dynamic Causal-Developmental AI that structurally grows its own computational topology to resolve localised network tensions through a completely differentiable lineage.

The do-Operator Logic

Using the Pearl framework for causal inference, we distinguish between observation P(B|A) and intervention P(B | do(A)). By applying Sparse Mechanism Shift (SMS) to perturbation datasets, we measure the directional asymmetry in component-to-component interactions.

If the impact of forcing Gene A on Gene B — P(B | do(A)) — is significantly greater than the inverse, the model orients the undirected coupling as a directed edge (A → B). This transforms the undirected spin-glass into a Directed Acyclic Graph, identifying Causal Pivot Genes that act as master regulators of state transitions.

Resolving Epistasis

Recent milestones include solving for epistasis — identifying causal relationships in frustrated Jᵢⱼ couplings. This moves beyond correlation-based network inference toward intervention-valid causal structure, enabling precise prediction of which perturbations will flip system fate decisions.

Manifold Parameterisation

The framework addresses the transcriptomic measurement revolution by utilising a Variational Autoencoder to project high-dimensional single-cell RNA sequencing counts (~20,000 dimensions) onto a parsimonious latent manifold (z with roughly 20 dimensions). Gaussian Mixture Models of healthy hematopoietic populations serve as Developmental Priors — anchoring the latent space to biophysical constraints.

Scale of Validation

Recent experimental validation at haematopoietic stem cell scale:

• 90K+ cells screened
• 237 gene knockout permutations explored
• 508M+ interrogated data points

Using a Conditional VAE executing Sparse Mechanism Shift, we identified the causal pivot genes responsible for lineage commitment — regulators that act as master decision points in the HSC differentiation hierarchy.

Inverse Manifold Simulation

The model acts as a scalable tool for Algorithmic Tissue Engineering, targeting causal rewiring found in Myelodysplastic Syndromes and Clonal Hematopoiesis.

Localisation — Manifold comparison between healthy priors and diseased states localises the coordinate of divergence (z-fail) where the HSC trajectory is biased toward pathological branching.

The Causal Nudge — The generative engine performs inverse in-silico simulation, predicting the minimal perturbation do(X) required to flip identified Causal Pivot Genes — effectively nudging the system back across the Waddington ridge toward a healthy regulatory attractor.

Target-Agnostic Architecture

The architecture has been validated as target-agnostic — scaling from new-to-nature proteins to single-cell transcriptomic landscapes without hand-coded domain-specific modules. The same hierarchical VAE framework that navigates P411-HF enzyme manifolds resolves frustrated gene couplings in hematopoietic differentiation.

Synthetic Cell Fates

By combining Conditional VAE inference with Sparse Mechanism Shift, the platform explores Synthetic Cell Fates — predicted cell states that exist on the manifold but have not been observed in nature, enabling rational design of reprogramming strategies.

Algorithmic Tissue Engineering

The Mesengenic framework represents a scalable approach to Algorithmic Tissue Engineering — moving from empirical perturbation screening toward causal, manifold-guided intervention design. By localising divergence points between healthy and diseased regulatory states, the platform enables precision reprogramming rather than brute-force gene editing.

Clinical Trajectory

The hematopoietic validation pipeline directly addresses medically urgent landscapes:

• Myelodysplastic Syndromes (MDS) — causal rewiring of dysplastic lineage bias
• Clonal Hematopoiesis — early detection and intervention at the causal pivot
• Lineage Reprogramming — minimal perturbation strategies derived from inverse manifold simulation

R&D Timeline Compression

The convergence of evolutionary density priors, abiological manifold navigation, and causal inference compresses the traditional drug discovery funnel:

1. In silico target identification replaces exhaustive combinatorial screening
2. Structural guardrails eliminate non-viable candidates before synthesis
3. Causal pivot genes focus experimental validation on high-confidence interventions

Out-of-Distribution Capabilities

Mesengenic's second core goal — expanding out-of-distribution capabilities — is not incremental improvement on existing methods. It represents access to an entirely new region of sequence and regulatory space: mathematically probable states that evolution's constrained walk never reached.

The Path Forward

Today, Mesengenic AI is building on hierarchical VAE architectures, causal spin frameworks, perturbation atlases, and the mathematical foundations of stochastic dynamics applied to biological self-organisation.

Biology for decades operated within the boundaries of what evolution sampled. The tools have changed. Evolutionary density priors, abiological manifold navigation, and causal developmental AI constitute the architectural address at which novel biopharma lives.

The most valuable therapeutic and catalytic states are not hidden in nature's archive — they are mathematically adjacent to it, waiting to be navigated. This space has not been systematically explored before. It is now.