eternagame / eternabrain Goto Github PK

• Properly encode resetting or pasting sequences

Came across the recent biorXiv posting of this work - awesome stuff! I should probably comment through the official Disqus comment thread but whatever.

It was interesting to note how you honed down the input data and implemented a couple of regularization strategies to improve the accuracy of the tandem-CNN model. I'm curious if you have messed around with different loss functions? Though the probability vectors at hand that are being predicted are short, some probability theory can still be applied - I think. See this for some insight on choosing a loss function. If you know, or can reason out, the distribution of the noise, a tailored loss function may improve the accuracy.

Some other off the shelf loss functions: https://www.tensorflow.org/api_docs/python/tf/losses.

Best of luck!

Resets base sequence to starting sequence each time rather than adding previous base changes to overall sequence

[[sequence],[current structure],[target structure],[energy]]

• Sequence and target are of length 85 for pid 6892348, but current structure and energy are of length 84 (only occurs for datapoints above 431)
• Most likely due to bug in structure_and_energy_at_current_time

1. read player experience data file for single-state puzzles
2. set a minimum threshold for experience
3. select players above threshold
4. feed uid into function to read movesets from a specific set of puzzle IDs and user IDs

• Python running out of memory when clustering movesets (not on reduced dimension data)
• 8GB
• Too many features for clustering algorithm
• 6892344, 6892345, 7254756, 7254759

Currently predicting only base

Options:

• Predict location using sequence, structures, and energy, then feed location prediction into separate neural net along with sequence, structures, and energy to predict base
• Predict location and base together, in a num_locations x num_bases matrix (or just 85 x 4)

After a certain number of epochs, the Keras model's loss becomes NaN and accuracy drops to 0.2698 (roughly 1/4)

CNN

Use BEAR notation as additional structural feature

SAP

For comparisons, use BEAR instead of dot-bracket notation

Tensorflow

cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=prediction,labels=y))

Logits and labels not of same size.

TFLearn

net = tflearn.input_data(shape=[None, 4716])

Shape size is incorrect for input matrix.

Things DSP should do after CNN and base structure selector:

Can look at EteRNAbot rules for more strategies.

• Correct base pairings
• Make all bases in loop A's
• Change end pairs to G-C pairs
  • for i in pairmap: get index of last pair before -1, get index of paired base, and change bases at those indices to G-C

• Boosting
  • Hairpins - boost with G
    • if num_unpaired_bases_in_a_row >= 3 then boost with G
  • Internal - boost with opposite G's
    • `if number of unpaired on either side is within +-3, then put Gs on either side
  • Single-bond stack - G-G in the loop
    • for every paired base: if '(' has one '.' following it and its complementary ')' has one '.' preceding it, then G-G boost
  • Bulges
    • http://eternagame.wikia.com/wiki/Zigzag
  • U-G-U-G superboost
    • for every paired base: if '(' has two '.'s following it and its complementary ')' has two '.'s preceding it, then UGUG boost
  • http://eternagame.wikia.com/wiki/Boosting
• If bases aren't pairing correctly:
  • Flip orientation of base pairs nearby
  • Change pairs to G-C pairs

Checks if base prediction already matches existing base in that location.
Example:

base_sequence = [1,1,1,1]
Prediction = [1630.97,1630.88,1630.56,1630.30]
argmax(Prediction) == 0 # base A
location = 3
if base_sequence[location] == argmax(Prediction):
   SmartPredictor() # takes 2nd highest probability and uses that as base

Problems

1. Base mutations are only to 4 (C)
2. Previous mutated sequence carries over to next iteration

Example:

Should be:

# this is for only 1 randomly generated sequence
[4 2 4 4 2 2 3 1 2]
[1 2 4 4 2 2 3 1 2]  [2 2 4 4 2 2 3 1 2]  [3 2 4 4 2 2 3 1 2]  [4 2 4 4 2 2 3 1 2]
# calculate reward
[4 2 4 4 2 2 3 1 2]
[4 1 4 4 2 2 3 1 2]  [4 2 4 4 2 2 3 1 2]  [4 3 4 4 2 2 3 1 2]  [4 4 4 4 2 2 3 1 2]
# calculate reward

Currently is:

[4 2 4 4 2 2 3 1 2]
[4 2 4 4 2 2 3 1 2]  [4 2 4 4 2 2 3 1 2]  [4 2 4 4 2 2 3 1 2]  [4 2 4 4 2 2 3 1 2]
# calculate reward
[4 2 4 4 2 2 3 1 2]
[4 4 4 4 2 2 3 1 2]  [4 4 4 4 2 2 3 1 2]  [4 4 4 4 2 2 3 1 2]  [4 4 4 4 2 2 3 1 2]
# calculate reward

Eventually, the entire list becomes 4s.

Eterna uses Vienna1 for calculating folding; EternaBrain is currently using Vienna2

I was wondering if the training, validation, and test datasets for EternaBrain were publicly available? In particular, I am looking for a dataset of RNA sequences and contact maps. I don't really need the specific player moves.

Thank you for your help, and sorry if this is off topic.

Some puzzles have bases which cannot be mutated - this feature might need to be encoded

Where the next consecutive base is paired, the DSP does not change the last base pair in a stack to G-C.

• Precision
• Recall
• F1 score

Puzzle Info

Sequence: GGGAUAACCU Structure: (((....))) Locks: oooxxxxooo

Base sequence

[0,0,1,0],[0,0,1,0],[0,0,1,0],[1,0,0,0],[1,0,0,0],[1,0,0,0],[1,0,0,0],[0,0,0,1],[0,0,0,1],[0,1,0,0]

Structure

[0,1,0,0],[0,1,0,0],[0,1,0,0],[1,0,0,0],[1,0,0,0],[1,0,0,0],[1,0,0,0],[0,0,1,0],[0,0,1,0],[0,0,1,0]

Pairmap

[9,0,0,0],[8,0,0,0],[7,0,0,0],[-1,0,0,0],[-1,0,0,0],[-1,0,0,0],[-1,0,0,0],[2,0,0,0],[1,0,0,0],[0,0,0,0]

Locks

[0,0,0,0],[0,0,0,0],[0,0,0,0],[1,0,0,0],[0,1,0,0],[1,0,0,0],[1,0,0,0],[0,0,0,0],[0,0,0,0],[0,0,0,0]

• Actual length should be 29971
• Features of 6892348 - length 30092
• Labels of 6892348 - length 29830
• Difference of 262

Occurring only on puzzles 6892343, 7254758

• Needs to work for moves with multiple base changes in one move
• Figure out a way to encode number of moves needed to complete the puzzle

Add target energy to list of features

• Use ViennaRNA to get energy given target structure of puzzle

• The encoded moveset data has 4 dimensions, which is too many for sklearn clustering models
• Need to find a way to reduce dimensions/PCA before fitting data to a model

Encode location similar to bases

[0,0,0,0,0,0,0,0,0,1,0,0,0,0,0] # location 10; number of bases = 15

For multi-state puzzles, the features should look like this:

[sequence, current_structure, target_structure, current_energy, 
target_energy, reporter, A, B, C] # A, B, C are other oligos

Instead of copying and pasting structures from Eterna website:

• Copy and paste all structures into a .txt file
• Read in structures and attempt to solve each one
• Record number of puzzles solved

• On 1 GPU, limited to 10 convolutional layers
• When parallelized across multiple GPUs, can add more layers as more memory available

Currently structure_and_energy_at_current_time works only with 1 puzzle ID. Would reduce the number of pickles and the amount of time unpickling when training.

• Change learning rate (below 0.0001)
• Change dropout rate (below 0.5)
• Increase number of epochs
• Change number of layers
• Change number of nodes

Hi Rohan,

Here are the getting started puzzles I was talking about: [6892343, 6892344, 6892345, 6892346, 6892347, 6892348, 7254756, 7254757, 7254758, 7254759, 7254760, 7254761]

I uploaded the problems file to the dropbox folder - github didn't like it's size when I tried to commit it directly. User data will also be there too. Hope everything went well with the SAT tests!

• Currently using Vienna web interface
• Will train faster when using Eterna or Vienna locally