This is the code for the simulation-based inference of the paper:

Julie N. Chuong, Nadav Ben Nun, Ina Suresh, Julia Matthews, Titir De, Grace Avecilla, Farah Abdul-Rahman, Nathan Brandt, Yoav Ram, David Gresham (2024) DNA replication errors are a major source of adaptive gene amplification.

It shares some components with the code of Avecilla et al. (2022) (graceave/cnv_sims_inference), mostly regarding network training.

To run the inference and see its results:

1. Install sbi library: pip install sbi
2. Create a folder named presimulated_data, and generate 100k simulations: python generate_presimulated_data.py -p 100000 -m WF -n reproduced -g Chuong_116_gens.txt
3. Train a neural density estimator on the simulations: python infer_sbi_initial_beneficial.py -m WF -pd WF_presimulated_data_100000_reproduced.csv -pt WF_presimulated_theta_100000_reproduced.csv -g Chuong_116shares_gens.txt -s 42 -n reproduced Once done, posterior should be at posteriors/posterior_reproduced.pkl
4. Go to Empirical Analysis.ipynb, change the posterior path, and run the notebook.

To run the collective posterior inference, run the Overall Posterior.ipynb notebook.

• The normalizing constants are saved in posteriors/log_Cs, but can be re-generated using OverallPosterior.get_log_C method.
• Overall posterior MAPs and samples are in the maps folder, but can be re-generated using get_map and sample methods.

To review your trained network and simulations, you can go to Simulation Analysis.ipynb. These simulations were generated from a narrow parameter range (details in the notebook). If you wish to validate your trained network on a different simulation set, you can simply generate other simulations by following step 2 above, and change the relevant cells in the notebook.

For more information, you can contact Nadav Ben Nun or the Ram lab at Tel Aviv University.

Source code: MIT License. Other content: Creative Commons Attribution 4.0 International License.