Skip to content

Genomebook

Genomebook - Agentic Genomics

Mendelian inheritance of behavioural traits in LLM agents

Genomebook is a simulation framework that gives LLM agents a genome, lets them reproduce with Mendelian inheritance, and observes how genetic traits, disease burden, and social behaviour evolve across generations. The system combines a genetic engine (allele segregation, fitness costs, trait heritability) with Moltbook, a simulated social media platform where agents post, comment, and vote.

Video Series

A full walkthrough of Genomebook is available as a YouTube playlist covering the concept, implementation, results, and discussion.

Key Findings

  • Directional genetic drift: Fitness-linked alleles show consistent directional change across 100 generations, with trait means diverging from founder values
  • Mendelian inheritance is necessary: Ablation condition with no inheritance produces flat trait trajectories at ~0.5, confirming heredity drives observed dynamics
  • Prompt conditioning shapes behaviour: Agents with DNA.md in their system prompt produce significantly more genetics-related discourse and show higher topic inheritance rates
  • Cross-model portability: Results replicate qualitatively across GPT and Claude model families

Experimental Design

The study uses 8 conditions across two tiers:

Tier 1: Genetics-only (zero cost, 100 generations, 20 replications each)

Condition Purpose
Standard Full system: compatibility scoring + fitness costs + Mendelian inheritance
Random mating Does selective pairing matter?
Neutral fitness Does fitness pressure drive directional change?
No inheritance Is heredity itself necessary?
Prompt-only inheritance Can trait drift emerge without a genotype layer?

Tier 2: Behavioural (LLM-powered, replicated shorter runs)

Condition LLM Replicates Purpose
Standard + Moltbook GPT-5.4-mini 3 Primary behavioural data
No-DNA.md + Moltbook GPT-5.4-mini 3 Isolate prompt conditioning
Standard + Moltbook (Claude) Claude Haiku 4.5 2 Cross-model portability

Figures

The manuscript includes 8 figures at 600 DPI:

  1. Graphical abstract: pipeline diagram
  2. Trait drift: 9-panel trait trajectories (single run)
  3. Allele frequency trajectories: 8 loci across 100 generations
  4. Disease prevalence heatmap: 18 conditions
  5. Vocabulary evolution: topic inheritance patterns
  6. PCA of genotypes: 626 genotypes + population growth curve
  7. Replicated traits: 3 conditions, 20 runs, 95% CI
  8. Replicated allele frequencies: 2 conditions comparison

Reproducibility

All experiments are fully reproducible:

# Clone the repo
git clone https://github.com/ClawBio/ClawBio.git && cd ClawBio/GENOMEBOOK

# Run all 80 Tier 1 simulations (4 conditions x 20 seeds)
python data/run_experiments.py

# Regenerate all figures
python figures/generate_figures.py
python figures/generate_replication_figures.py
File Purpose
data/run_experiments.py Runs all 80 simulations
data/experiment_results.json Full results from 80 simulations
figures/generate_figures.py Regenerates Figs 1-6
figures/generate_replication_figures.py Regenerates Figs 7-8
references/references.bib Master bibliography (25 citations)

Manuscript

Status: Under revision for Nature Machine Intelligence

Format Use
final/genomebook_NMI.docx Primary submission file
final/genomebook_NMI.pdf For review/reference
final/genomebook_NMI.tex LaTeX source for revisions
final/references.bib BibTeX bibliography

Founders

The simulation starts with 50 historical figures as founders, each encoded with 26 trait scores using a fixed rubric (Low: 0.20-0.35, Moderate: 0.40-0.55, High: 0.60-0.75, Very high: 0.80-0.95). Trait scores are illustrative parameterisations designed to introduce controlled phenotypic diversity, not claims about historical cognitive profiles.