Epigenetic Tracking: Implementation Details

“Epigenetic Tracking” is the name of a model of cellular development that, coupled with an evolutionary technique, becomes an evo-devo (or “artificial embryology”, or “computational development”) method to generate 2d or 3d sets of artificial cells arbitrarily shaped. ‘In silico’ experiments have proved the effectiveness of the method in devo-evolving any kind of shape, of any complexity (in terms of number of cells, number of colours, etc.); being shape complexity a metaphor for organismal complexity, such simulations established its potential to generate the complexity typical of biological systems. Moreover, it has also been shown how the underlying model of cellular development is able to produce the artificial version of key biological phenomena such as embryogenesis, the presence of “junk DNA”, the phenomenon of ageing and the process of carcinogenesis. The objective of this document is not to provide new material (most of the material presented here has already been published elsewhere): rather, it is to provide all details that, for lack of space, could not be provided in the published papers and in particular to give all technical details necessary to re-implement the method.

💡 Research Summary

**
The paper “Epigenetic Tracking: Implementation Details” serves as a comprehensive technical manual for reproducing the Epigenetic Tracking (ET) model, a developmental framework that couples a cellular automaton‑like growth process with an evolutionary algorithm to generate arbitrarily complex 2‑D and 3‑D shapes. The authors begin by positioning ET within the broader field of artificial embryology, highlighting its distinctive use of “driver cells” that carry a Cell Epigenetic Type (CET) and a Meta‑Operator Code (MOC). These two identifiers allow the system to select, at each discrete developmental step, a specific Development Operator (DO) from a genome‑encoded library. When a DO’s (CET, MOC) pair matches a driver cell, the operator executes actions such as inserting new driver cells, changing the colour of existing cells, or modifying the local geometry. By iterating this process over many time steps, the model can sculpt intricate structures with multiple colours and fine‑grained detail.

Evolutionary optimization is performed with a standard genetic algorithm. An individual consists of a genome (the ordered list of DOs) and an initial set of driver cells. Fitness is measured by comparing the final phenotype to a target shape using a voxel‑wise distance metric that accounts for both spatial occupancy and colour matching. Genetic operators include crossover of DO sub‑lists, insertion/deletion of operators, and mutation of operator parameters, CET values, and MOC codes. A notable design choice is the intentional inclusion of “junk DNA” – surplus, non‑functional operators – which expands the search space and improves evolvability, especially for highly complex targets.

From an implementation standpoint, the authors detail data structures that make the system scalable. The cellular space is stored as a dense 2‑D or 3‑D array; each cell record holds a flag indicating driver or normal status, its CET vector, colour, and a pointer to any associated driver data. To achieve constant‑time lookup of applicable operators, a hash table keyed by the concatenation of CET and MOC is employed. A priority matrix resolves conflicts when multiple operators would act on the same voxel in a single step, ensuring deterministic outcomes. The paper also describes how the model can be repurposed to simulate biological phenomena such as ageing (by gradually reducing operator activation probabilities) and carcinogenesis (by artificially boosting the activation of proliferation operators), thereby providing a bridge between computational development and biological theory.

Experimental results cover a wide range of target shapes, from simple geometric primitives to highly detailed multi‑colour objects containing up to a million cells. The authors report convergence curves, genome sizes, and runtime statistics, demonstrating that ET can reliably evolve solutions for both 2‑D and 3‑D tasks. Comparisons between runs with and without junk DNA show that the former consistently achieve higher success rates on complex targets, confirming the hypothesised benefit of a larger neutral space.

The discussion acknowledges current limitations: the number of operators can grow dramatically, leading to increased memory consumption and CPU time, and the present implementation lacks GPU acceleration, which restricts the size of feasible 3‑D simulations. Future work is outlined, including parallelisation strategies, dynamic operator compression, and quantitative validation against empirical biological data to further substantiate the model’s relevance to real embryogenesis, ageing, and tumorigenesis.

In summary, the paper provides a thorough, reproducible blueprint for the ET system, covering algorithmic foundations, evolutionary strategies, low‑level implementation tricks, and experimental validation. It equips researchers with the necessary details to replicate, extend, or adapt the model for investigations into developmental biology, evolutionary computation, and the emergence of complex structures from simple rule‑based processes.