Repository files navigation

UsingEterna data to understand and predict how players solve RNA folding puzzles.

• These data are move sets graciously donated by Eterna players to accelerate scientific research into RNA design
• Neural networks to learn how top players solve Eterna puzzles and to predict solutions to RNA folding puzzles
• Unsupervised learning to group Eterna players based on their style of solving RNA folding puzzles

Rohan Koodli

• Link to paper onPLoS Computational Biology, link tocover art.
• EternaBrain 1.2 is the version referenced in the above paper, and supports Python 2. EternaBrain 2.0 supports both Python 2 and 3.
• Many thanks to the contributing authors who provided guidance, ran benchmarks, and did testing:Benjamin Keep,Katherine Coppess,Fernando Portela, andRhiju Das.
• This software is freely available for non-commercial use. Royalties for commercial use will come back to fund research on Eterna, administered by Stanford University. Please seelicense.

61/100

Python:numpy, tensorflow, pandas, seaborn, matplotlib, scikit-learn

Conda:viennarna (runconda install -c bioconda viennarna, you should be able to runpython -c "import RNA" without any errors)

RNAfold version1.8.5 from ViennaRNA (seeconfig-ViennaRNA.md for installation instructions)

R:ggplot2, reshape2

Configure RNAfold, and enter the correct path to Vienna 1.8.5 in thepath inpredict_pm.py field, then runpython predict_pm.py "<valid structure in dot-bracket notation>".

Following curates a subset of training data "eternamoves-select" which trains an effective CNN move predictor with reasonable test accuracy.

Go toexperts.py and modify the variablescontent anduidList.content is the puzzle IDs of the puzzles you want movesets on, anduidList is the user ID's of the players you want movesets from. You can either specify these manually, or you can use functions to get them for you.getPid() will retrieve all the single state puzzles, andexperience will retrieve all players with an experience above a certain threshold.

Example:

content=getPid()# all single-state puzzlesuidList=experience(3000)# the top 70 experts, or the top 1 percent of all players

or, if you want less puzzles and more experts, you can read inteaching-puzzle-ids.txt, which contains 92 key puzzles:

withopen(os.getcwd()+'/movesets/teaching-puzzle-ids.txt')asf:progression=f.readlines()progression= [x.strip()forxincontent]progression= [int(x)forxincontent]progression.extend([6502966,6502968,6502973,6502976,6502984,6502985,6502993, \6502994,6502995,6502996,6502997,6502998,6502999,6503000])content=progressionuidList=experience(1000)

Go tofastest.py and modifycontent andmax_moves.content requires the same inputs as above, andmax_moves is an integer specifying the maximum number of moves you want the data to have.

Example:

content=getPid()# all the single state puzzlesmax_moves=30# all solutions in under 30 moves

EternaBrain uses a convolutional neural net (CNN). Run bothbaseCNN.py andlocationCNN.py. Just specify the path and name of your pickled data files here:

forpidincontent:try:feats=pickle.load(open(os.getcwd()+'/pickles/X-exp-loc-'+str(pid),'rb'))ybase=pickle.load(open(os.getcwd()+'/pickles/y-exp-base-'+str(pid),'rb'))yloc=pickle.load(open(os.getcwd()+'/pickles/y-exp-loc-'+str(pid),'rb'))foriinrange(len(feats)):feats[i].append(yloc[i])real_X.extend(feats)real_y.extend(ybase)pids.append(feats)exceptIOError:continue

Specify the name and directory of where you want the model to be saved here:

saver.save(sess,os.getcwd()+'/models/base/baseCNN')saver.export_meta_graph(os.getcwd()+'/models/base/baseCNN.meta')

Load your model into the appropriate locations for the base predictor and location predictor inpredict_pm.py. Specify the RNA secondary structure, starting nucleotide sequence, and path to Vienna inDOT_BRACKET,NUCLEOTIDES, andpath. Also specify the natural energy and target energy incurrent_energy andtarget_energy (default is 0 kcal).

DOT_BRACKET='((((....))))'path=os.getcwd()+'./RNAfold'len_puzzle=len(dot_bracket)NUCLEOTIDES='A'*len_puzzlece=0.0# current energyte=0.0# target energy

You can specify the minimum amount of the puzzle you want the CNN to solve (on its own, it generally cannot solve long puzzles). The amount is calculated by how much of the current structure matches the target structure. Once it reaches the threshold specified or completes the maximum number of moves, the sequence moves to the Single Action Playout (SAP), which runs a Monte Carlo Tree Search to determine what mutations bring the RNA molecule closer to the target secondary structure.

MIN_THRESHOLD=0.65

Now, you can run the model and it will attempt to find a nucleotide sequence that will fold into the secondary structure provided.

6892343 - 6892348, 7254756 - 7254761

8627, 55836, 231387, 42833