Janet Matsen - Programming microbes using Python

The genome of a typical microbe contains roughly 5 million base pairs of DNA including > 4000 genes, which provide the instructions for cellular replication, energy metabolism, and other biological processes. At Zymergen, we edit DNA to design microbes with improved ability to produce valuable materials and molecules. Microbes with these edits are built and tested in high throughput by our fleet of robots. Genomes are far too large for exhaustive search, so identifying which edits to make requires machine learning on non-standard features. Our task to extract information from trees, networks, and graphs of independently representable knowledge bases (metabolism, genomics, regulation), in ways that respect the strongly causal relationships between systems. In this talk, I will describe how we use Python’s biological packages (e.g. BioPython, CobraPy, Escher, goatools) and other packages (NetworkX, TensorFlow, PyStan, AirFlow) to extract machine learning features and predict which genetic edits will produce high-performance microbes.

https://us.pycon.org/2018/schedule/presentation/242/

Transcript

1. Programming Microbes Using Python @JanetMatsen Data Scientist, Zymergen PyCon 2018
2. Microbes can produce chemicals drugs alcohol flavors novel materials https://en.hdbuzz.net/038
3. Microbes are grown in HUGE tanks
4. Zymergen’s high-throughput testing 105 liters 10-3 liters 101 liters
5. Zymergen’s high-throughput testing 105 liters 10-3 liters 101 liters design large #s of variants automate data collection identify statistically significant improvements predict tank performance from plates
6. Robotic automation drives throughput
7. How does genetic engineering work?
9. Challenges
10. The DNA search space is huge. Genome optimization is hard 1/2 of genes have unknown function thousands of genes millions of characters
11. perturb each of 4,000 genes: 4,000 experiments
12. perturb 6-gene combinations: 4,000 choose 6 scipy.special.com(4000, 6) ≈ 1017
13. exhaustive exploration of 6-gene combinations would take > 1010 years
14. exhaustive exploration of 6-gene combinations would take > 1010 years … that’s the age of the universe ...
15. Genome optimization is hard We can’t just simulate. vs ...more complex! futurism.com
16. Epistasis: combinations have unpredictable effects Genome optimization is hard https://en.hdbuzz.net/038 increase the activity of gene X
17. Epistasis: combinations have unpredictable effects Genome optimization is hard www.indiamart.com https://youtu.be/96A2lGSOOZM
18. Balancing objective functions: survival and product production. Genome optimization is hard product 90% of theoretical maximum!
19. Web Analytics Zymergen Experiments are expensive & time consuming
20. Challenges Recap product
21. Our solution
22. Python increases Zymergen’s efficiency.
23. TensorFlow AirFlow PyStan
24. Metabolism describes chemical reactions “metabolites”
25. Metabolism describes chemical reactions
26. Metabolism describes chemical reactions
27. Metabolism is a complex network!
28. Let’s make MSG! (1909–1962) Initially MSG was extracted History of glutamate production. The American Journal of Clinical Nutrition, 2009. Fermentation established in 1957 (glutamate)
29. Let’s make MSG! “on-pathway” “off-pathway” Let’s make MSG!
30. Extracting metabolic features from CobraPy
31. Extracting metabolic features using CobraPy “fluxes”
32. Extracting metabolic features using CobraPy >>> model = \ cobra.io.read_sbml_model('e_coli.sbml')
33. Extracting metabolic features using CobraPy >>> msg_reaction = \ model.reactions.get_by_id('msg_production')
34. Extracting metabolic features using CobraPy >>> model.objective = msg_production
35. Extracting metabolic features using CobraPy >>> model.optimize()
36. Extracting metabolic features using CobraPy >>> def summarize_fluxes(model): ... return pd.DataFrame(results)
37. Extracting metabolic features using CobraPy We can loop over multiple objectives.
38. Extracting metabolic features using CobraPy
39. Extracting metabolic features using CobraPy
40. Wait… I thought you told me you can’t use simulations for self-play. futurism.com
41. CobraPy: a subset of reactions are modeled
42. CobraPy: a subset of reactions are modeled We’ve only observed one corner of the board game. models only include a fraction of the genes
43. CobraPy: linear steady-state modeling ∴ & ∴ Not for simulations over time, or varying external conditions.
44. But the features still work!
45. Maps were made with using + escher.github.io
46. goatools
47. THANK YOU! @JanetMatsen P.S. Zymergen is hiring!