This project is focused on exploring modeling techniques to predict the labs-of-origin for DNA constructs called plasmids. Plasmids have been used for decades in molecular cloning applications and are critically important to both research activities and industrial production. However, the increased availability of advanced methods and tools for genetic engineering raises the specter of potential harm due to unintended or malicious activities by a broader range of actors. The development of tools that can correctly identify the lab of origin for a given plasmid is becoming ever more important and urgent.

Source: https://www.nlm.nih.gov/exhibition/fromdnatobeer/img/exhibition-recombinantDNA.jpg

The Genetic Engineering Attribution Challenge (GEAC), a data science competition sponsored by altLabs and hosted by DrivenData, was created to crowdsource potential solutions to this problem. (The competition page can be viewed at https://www.drivendata.org/competitions/63/genetic-engineering-attribution/.) DrivenData and altlabs published a blog with some starter code and guidance to help participants get started and to successfully format competition submissions. Partipicants downloaded training data and test data on which to make predictions for competition submission. The training data set included over 63,000 plasmids submitted by a total of 1,314 labs. Plasmid sequence lengths ranged from a few dozen DNA 'letters' to over 60,000, making this a fairly unusual sequential analysis relative to, say, natural language processing or time series analysis.

For my Flatiron data science capstone project, I chose to use a dataset from the GEAC competition. It was a fascinating topic that allowed me to revisit and update the molecular biology knowledge I had gained in college. It also pushed me to learn much more about deep learning and AI than I otherwise would have at the end of an already-rigorous data science program!

Source: http://clipart-library.com/clipart/479704.htm

While I did peek at a few abstracts of scientific papers by altlabs and others on this topic, I started off by thinking through the problem, applying what I had learned in the program, asking for guidance from my instructors, and beginning with the the guidance and starter code from the DataDriven/altlabs blog post.

• The data sets provided by the competition organizers are too large to be hosted on github. The required .csv files at https://www.drivendata.org/competitions/63/genetic-engineering-attribution/page/164/. Login is required, but accounts are free and easy to set up.

• The DrivenData website for this competition features a blog post that provides guidance and starter code to participants, so that everyone is able to access the data and format submissions properly. The blog is available at https://www.drivendata.co/blog/genetic-attribution-benchmark/.

There are 41 columns in this dataset. Each row corresponds to a plasmid DNA sequence, which is uniquely identified by sequence_id, a 5-character alphanumeric string. In addition to the DNA sequences provided in sequence, there are 39 binary features that provide metadata about the plasmids. All variables are described below.

• sequence (type: string): A plasmid DNA sequence. Any Us were changed to Ts and letters other than A, T, G, C, or N were changed to Ns. Possible values: A, T, G, C, or N
• bacterial_resistance_ampicillin, bacterial_resistance_chloramphenicol, bacterial_resistance_kanamycin, bacterial_resistance_other, bacterial_resistance_spectinomycin (type: binary): One-hot encoded columns that indicate the antibiotic resistance of the plasmid used for selecting during bacterial growth and cloning.
• copy_number_high_copy, copy_number_low_copy, copy_number_unknown (type: binary): One-hot encoded columns that indicate the number of plasmids per bacterial cell.
• growth_strain_ccdb_survival, growth_strain_dh10b, growth_strain_dh5alpha, growth_strain_neb_stable, growth_strain_other, growth_strain_stbl3, growth_strain_top10, growth_strain_xl1_blue (type: binary): One-hot encoded columns that indicate the strain used to clone the plasmid.
• growth_temp_30, growth_temp_37, growth_temp_other (type: binary): One-hot encoded columns that indicate the temperature the plasmid should be grown at.
• selectable_markers_blasticidin, selectable_markers_his3, selectable_markers_hygromycin, selectable_markers_leu2, selectable_markers_neomycin, selectable_markers_other, selectable_markers_puromycin, selectable_markers_trp1, selectable_markers_ura3, selectable_markers_zeocin (type: binary): One-hot encoded columns that indicate genes that allow non-bacterial selection (for a plasmid used outside of the cloning organism).
• species_budding_yeast, species_fly, species_human, species_mouse, species_mustard_weed, species_nematode, species_other, species_rat, species_synthetic, species_zebrafish (type: binary): One-hot encoded columns that indicate the species the plasmid is used in, after cloning.

DrivenData and altlabs published a blog (https://www.drivendata.co/blog/genetic-attribution-benchmark/) providing ideas for how to approach the project and some starter code to explore the data and properly format model predictions for submission. To generate predictions to feed into the function, they constructed a fairly simple random forest model with DNA n-grams ("bag of words"). They then took the model predictions and ran them through the function to generate the predictions in the appropriate format for competition submission.

Note: What follows below is excerpts of code, visualizations, and results. For full technical details, please see the technical notebook in the repo for this project.

sequence_lengths = train_values.sequence.apply(len)
sequence_lengths.describe()

count    63017.000000
mean      4839.025501
std       3883.148431
min         20.000000
25%        909.000000
50%       4741.000000
75%       7490.000000
max      60099.000000
Name: sequence, dtype: float64

We can see that the vast majority of plasmids are less than about 10,000 base pairs (bp), with a large spike of plasmids of length ~1,000 bp. However, the scale of this graph can be misleading: there are still thousands of plasmids beyond 10,000 bp in length; this can be seen in the following graphs.

Looking just at the distribution of plasmid lengths between 8,000 and 25000 bp, we see there are still many hundreds just in this bandwidth of plasmid length:

(First 5 rows of dataframe)