EternaBrain

Using Eterna 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

Author

Rohan Koodli

Notes

• Link to paper on PLoS Computational Biology, link to cover 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, and Rhiju 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 see license.

Benchmarks

Eterna100

61/100

Dependencies

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

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

RNAfold version 1.8.5 from ViennaRNA (see config-ViennaRNA.md for installation instructions)

R: ggplot2, reshape2

Using a pretrained model

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

Generating your own data and CNNs

Step 1: Generate the training data

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

Selecting expert solutions

Go to experts.py and modify the variables content and uidList. content is the puzzle IDs of the puzzles you want movesets on, and uidList 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, and experience will retrieve all players with an experience above a certain threshold.

Example:

content = getPid() # all single-state puzzles
uidList = 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 in teaching-puzzle-ids.txt, which contains 92 key puzzles:

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

Selecting the fastest solutions

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

Example:

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

Step 2: Training the convolutional neural network (CNN)

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

for pid in content:
    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'))
        for i in range(len(feats)):
            feats[i].append(yloc[i])
        real_X.extend(feats)
        real_y.extend(ybase)
        pids.append(feats)
    except IOError:
        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')

Step 3: Predicting

Load your model into the appropriate locations for the base predictor and location predictor in predict_pm.py. Specify the RNA secondary structure, starting nucleotide sequence, and path to Vienna in DOT_BRACKET, NUCLEOTIDES, and path. Also specify the natural energy and target energy in current_energy and target_energy (default is 0 kcal).

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

Key Puzzles

Multi-state puzzles

6892343 - 6892348, 7254756 - 7254761

Key Players

8627, 55836, 231387, 42833