An abstract model of nonrandom, non-Lamarckian mutation in evolution using a multivariate estimation-of-distribution algorithm

This paper presents a simulation model based on estimation-of-distribution algorithms that demonstrates how nonrandom, non-Lamarckian mutations, driven by internally accumulated genomic information, interact with selection and recombination to advance evolution, thereby offering a computational framework that aligns with Interaction-based Evolution theory and connects to Darwinian observations and computational learning theory.

The Gist

The Big Question: How Do We Evolve?

For over a century, scientists have believed there are only two ways life changes over time:

1. The "Random Dice" Theory (Standard Evolution): Mutations are like rolling dice. They happen by pure accident, with no connection to what the animal needs. If a rabbit happens to grow a faster leg by accident, it survives. If not, it doesn't. Nature just picks the winners.
2. The "Lamarckian" Theory (The Old Idea): An animal senses a problem (like a cold winter) and decides to change its genes to grow a thicker coat, passing that new coat to its babies. This idea was largely rejected because it seems too magical—like a computer rewriting its own code just because it's cold outside.

This paper proposes a third option: A "Goldilocks" theory called Interaction-Based Evolution (IBE). It suggests mutations aren't random dice rolls, but they also aren't conscious decisions. Instead, they are smart reactions to the organism's own internal history.

---

The Analogy: The "Smart Chef" vs. The "Blind Cook"

Imagine you are trying to invent the perfect recipe for a cake.

• The Random Cook (Standard Evolution): You throw ingredients into a bowl completely at random. Sometimes you get flour and eggs; sometimes you get sand and ketchup. You taste them, throw away the bad ones, and keep the good ones. You have to try millions of random combinations to find a cake.
• The Lamarckian Cook: You taste the batter, realize it's too sweet, and magically change the recipe in your head to add less sugar before you bake the next batch. This is impossible in biology.
• The Smart Chef (This Paper's Model): You look at the cakes that worked well yesterday. You notice a pattern: "Every time I used vanilla and cinnamon together, the cake was great." You don't just guess randomly; you learn from the successful cakes of the past. You then use that knowledge to mix the ingredients for today's batch. The new ingredients aren't random, but they aren't magic either—they are based on the history of what worked.

The Computer Experiment: The "Restricted Boltzmann Machine"

The authors didn't use real animals; they used a computer simulation. They set up a puzzle (a math problem called MAX-SAT) where a population of "digital organisms" had to find the best solution.

They compared two groups:

1. Group A (Random Mutation): They took the best digital organisms, copied them, and randomly flipped a few bits (like flipping a coin to change a 0 to a 1).
2. Group B (The "Smart" Model): They took the best organisms and fed their data into a special AI brain (called a Restricted Boltzmann Machine). This AI brain looked at the winners and asked, "What patterns do these winners share?" It then used those patterns to generate the next generation.

The Result:
The "Smart" group (Group B) solved the puzzle much faster and better than the "Random" group.

• Why? Because the AI brain noticed that certain genes (bits of code) worked well together. It didn't just flip bits randomly; it flipped them in a way that respected those partnerships.
• The Metaphor: Imagine a dance troupe. The Random group keeps swapping dancers randomly. The Smart group watches who dances well together and ensures those pairs stay together in the next show.

Key Concepts Made Simple

1. The "Used-Together-Fused-Together" Rule

In the real world, if two genes are used together constantly (like genes for running and genes for breathing), they might physically stick together or mutate in a way that keeps them linked.

• Analogy: Think of a toolbox. If you always use a hammer and a nail together, you might eventually glue them into a single "Hammer-Nail" tool. You don't need to invent a new tool from scratch; you just simplify the process by combining what you already use.
• The Paper's Point: Mutations aren't random accidents; they are often the result of the genome "gluing together" things that have been working well together for a long time.

2. Simplicity Creates Complexity

Usually, we think "simple" means "less complex." But this paper argues that simplification is the engine of complexity.

• Analogy: Think of learning to drive. At first, you have to consciously think about every gear shift, mirror check, and pedal press (complex). Eventually, your brain "chunks" these actions into one smooth motion: "Drive." You simplified the process, which freed up your brain to learn how to race or drive in snow (complex new skills).
• The Paper's Point: Evolution simplifies the internal rules of the organism. Once the organism is "simplified" and efficient, it can handle much more complex challenges.

3. The Bell Curve Mystery

In nature, traits like height form a "Bell Curve" (most people are average, few are very tall or very short). Scientists usually say this happens because many genes add up like numbers (1 + 1 + 1 = 3).

• The Paper's Twist: The simulation showed you can get a Bell Curve even when genes interact in complex, non-additive ways. It's not just a math sum; it's a complex dance where the whole is greater than the sum of its parts.

Why Does This Matter?

This paper suggests that evolution is a form of learning.

• The Population is the Student: The group of organisms learns from its collective history.
• The Environment is the Teacher: It tells the group what works (survival) and what doesn't (death).
• The Mutation is the Homework: Instead of guessing randomly, the organism uses its internal "notes" (accumulated genetic history) to figure out the next step.

The Bottom Line

This paper argues that we have been thinking about evolution wrong. We assumed mutations were either random accidents or magical wishes.

Instead, mutations are intelligent summaries of the past. They are the genome saying, "We've tried a million things. These specific combinations worked. Let's build the next generation based on those successes, not on a coin flip."

It bridges the gap between biology and computer science, showing that life evolves not just by surviving the fittest, but by learning from the winners.

Technical Summary

1. Problem Statement

The paper addresses a fundamental conceptual dichotomy in evolutionary biology that has persisted for over a century: the choice between random mutation (where mutations are accidental and unrelated to their utility) and Lamarckism (where organisms sense the environment and directly induce beneficial heritable changes).

• The Limitation of Current Theory: Traditional evolutionary theory relies on "random mutation" combined with natural selection. However, recent empirical evidence (e.g., regarding the HbS malaria-resistance mutation and APOL1 mutations) suggests that specific mutations arise more frequently in contexts where they are adaptive, contradicting the strict "randomness" of mutation origination.
• The Gap: Existing models cannot explain how mutations can be non-random (dependent on internal genomic information) without being Lamarckian (dependent on immediate environmental sensing).
• The Goal: The authors aim to demonstrate that a third alternative exists: nonrandom, non-Lamarckian mutation. They propose that mutations are influenced by information accumulated internally in the genome over generations, a concept central to their Interaction-based Evolution (IBE) theory.

2. Methodology

The authors utilize a computational simulation framework based on Estimation-of-Distribution Algorithms (EDAs), specifically employing a Multivariate EDA powered by a Restricted Boltzmann Machine (RBM).

• The Core Mechanism (RBM): Unlike standard Genetic Algorithms (GAs) that use random point mutations, the authors use an RBM to model the generation of new genomes.
  • Training: At each generation, the top-performing individuals (survivors) are used to train the RBM. The RBM learns the statistical distribution and dependencies (interactions) between the bits (genes) of these survivors.
  • Generation: The trained RBM then generates the next population by sampling from the learned distribution. This process mimics "learning" the rules of successful combinations rather than random guessing.
• The Analogy: The RBM's weight adjustment mechanism (Hebbian learning: "units that fire together wire together") is analogized to biological mechanisms like gene fusion, where genes used together are more likely to become physically fused or linked over generations.
• Experimental Setup:
  • Problem: The MAX-SAT problem (specifically MAX-3-SAT), an NP-hard Boolean satisfiability problem used to create complex adaptive landscapes.
  • Comparisons: The RBM-based model (IBE-like) is compared against:
    1. RM (Random Mutation): A standard evolutionary model using random bit-flip mutations.
    2. BRG (Best Random Guess): A baseline tracking the best solution found by pure random guessing.
  • Parameters: Simulations varied population sizes ($N$), genome lengths ($n$), and survival rates. The RBM parameters included hidden units, learning rates, and iteration counts.
  • Specific Tests:
    • Genome Completion: Testing if the RBM could infer missing gene values based on known values, proving it learned inter-gene dependencies.
    • Parity Functions: Using even-parity functions to verify the model relies on interactions between genes rather than independent additive effects.
    • Selection Pressure Changes: Observing how variation production changes when the selection environment shifts.

3. Key Contributions

• Demonstration of a Third Category: The paper provides a concrete, mathematical demonstration that heritable change can be neither random nor Lamarckian. The mutation mechanism depends on internal genomic information (learned from previous generations) but does not sense the external environment.
• Reconciliation of Darwinism and Mendelism: The model offers a new framework for reconciling gradual evolution with discrete genetic changes. It suggests that sexual recombination breaks down combinations, but nonrandom mutation (driven by internal fan-in of information) integrates successful combinations back into the genome for future generations.
• Parsimony and Fit: The study illustrates that evolution is driven by the interaction of two forces: Natural Selection (external fit) and Natural Simplification (internal parsimony). The RBM simplifies complex correlations into local connections, mirroring biological regulatory simplification.
• Reinterpretation of Randomness: The paper argues that random bits do not directly encode improvement; instead, randomness enables generalization. The RBM learns the structure of success, and random sampling explores the space defined by that structure.

4. Key Results

• Superior Performance: The RBM-based model consistently outperformed the Random Mutation (RM) model, particularly as problem size and population size increased. The RBM model showed a rapid "transient" phase where fitness increased sharply, whereas the RM model improved more slowly.
• Gene Interactions (Fan-in): In the "Genome Completion" test, the RBM successfully inferred missing gene values based on the context of other genes. This proved the model was not treating genes as independent variables but was capturing complex interactions (epistasis).
• Bell-Shaped Distributions: Contrary to the standard view that bell-shaped trait distributions require additive genetic effects (Central Limit Theorem), the simulation showed that complex, non-additive functions (where genes interact) can also produce bell-shaped fitness distributions that shift over time under selection.
• Variation and Selection: The model demonstrated that the amount of genetic variation produced is not fixed. When selection pressure changes, variation production increases temporarily to explore new solutions, then decreases as the population converges. This aligns with Darwin's observation that changed conditions increase heritable variation, but explains it via internal information accumulation rather than stress-sensing.
• Learning vs. Evolution: The simulation reframes evolution as a learning process where the population is the hypothesis and individuals are examples. This contrasts with Valiant's "evolvability" model, which treated the individual as the hypothesis.

5. Significance

• Theoretical Shift: The paper challenges the dogma that mutation must be random to avoid Lamarckism. It proposes that nonrandom, non-Lamarckian mutation is a logically coherent and biologically plausible mechanism.
• Bridge Between Fields: It creates a bridge between evolutionary biology, evolutionary computation, and computational learning theory. It suggests that evolutionary algorithms (like EDAs) are not just optimization tools but abstract models of a specific type of biological evolution.
• Implications for Complexity: The results suggest that biological complexity arises not just from random accumulation of small changes, but from the integration of information and regulatory simplification. The "breaking down" of genetic combinations by sex is not a bug but a feature that allows the system to learn and integrate successful patterns.
• Future Directions: The authors call for future research to develop more realistic biological models that incorporate "internal fan-in" of information, moving beyond the current dichotomy of random mutation vs. Lamarckism. This has implications for understanding the origins of complex adaptations, the role of sexual recombination, and the development of next-generation machine learning algorithms.

In summary, the paper uses a multivariate EDA with an RBM to simulate a form of evolution where mutations are guided by the internal history of the genome. This model successfully solves complex optimization problems more efficiently than random mutation models and provides a theoretical basis for understanding how complex life can evolve without violating the principles of natural selection or resorting to Lamarckism.