The combination of deep learning model and wet experiment is expected to be used for metabolic flux analysis

background

On the road of protein structure transformation , Need powerful computer programming assistance , In the field of deep learning , Found a powerful tool , So study how to combine with wet experiment .

Tools

DeeplearningApproach Anconda python And various dependent packages Computer Wet experimental data

Realize the idea

1. First , Good configuration DeeplearningApproach, How to configure see another article 《 Deep learning predicts enzyme activity parameters, improves enzyme constraint model construction, and builds ab initio environment 》.
2. Use the mutation sites with existing data to predict , For example, I want to predict the sequence of a protein L59,W60,Y153,R416 Site mutation to alanine （A）, So I need to configure my tsv file , I wrote a script for the mutation here , The code is as follows ：

import copy

def sort(a,b,c,d):
    for n,i in enumerate(a):
        if n+1 == b:   #  Determine whether it is the site I want 
            d = a[n]   #  Assign this site of the original sequence to d
            a = copy.deepcopy(a)   #  Here you need a deep copy , Otherwise it will change WT Sequence , Of course, if it is the same site, it can be predicted as others 19 Amino acids can be used without 
            a[n] = c   #  This site of the original sequence is replaced by the site I need ,
            i = a[n]   #  Extract the changed amino acids 
            break
        else:
            a[n] = i
    a = ''.join(a)
    # print(' Has already put the first %d Amino acids %s Change it to %s, The output sequence is as follows :%s'%((n+1),d,i,a))

    print(' Enter the substrate name  '+' Enter the chemical formula of the substrate  '+a) #  The difference between a plus sign and a comma 
    
def main():
    #  Manual input 
    # while True:
    # a = list(input(' Please enter the amino acid sequence ：'))
    # b = int(input(' Please select the amino acid site to be modified ：'))
    # c = input(' Please select the single letter abbreviation of amino acid to be changed ：')
        # sort(a,b,c,d=None)

    aa_list = ['A','G','V','L','I','P','F','Y','W','S','T','C','M','N','Q','D','E','K','R','H']
    pro = ' Input amino acid sequence '
    pro_list = []
    for p in pro:
        pro_list.append(p)
    for i in aa_list:
        sort(pro_list,231,i,d=None)


if __name__ == '__main__':
    main()

3. By extraction WT,L59,W60,Y153,R416 Of Kcat Value to draw a column chart （ Of course, it's OK to look without drawing , Just a few data ）, Compare with the existing experimental data ：

experimental data

Deep learning prediction data

4. From the results , The fitting is not very good , But there is no problem in the general direction ,Kcat Less than WT The actual is also small , Greater than WT The reality is also big .（ Of course, the amount of data is very small , It doesn't explain the problem , The predicted results can only be tried ） 5. By way of A231 Site mutation to other 19 Amino acids （ It is still configured with the above code tsv file ）, The prediction results show that A231G Of Kcat Value ratio WT Big , Maybe we can try , Or mutate other sites to find Kcat Greater than R416A Try the mutation of . 

forecast A231 Site mutation to other 19 Amino acids

Deep excavation

1） How to use it DeeplearningApproach characteristic ？
Make one tsv library , This library contains unit point mutations and double site mutations of all amino acids near the active site of the protein , And WT Of Kcat It's worth comparing , Screen beneficial mutations for experiments .
2） Can you make DeeplearningApproach More suitable for a certain exact experiment , Make it Kcat The value is more accurate , And then use it Kcat Do metabolic flux analysis , Maybe we can start from two aspects ：

• Adjust the parameters , To optimize the fitting / Under fitting phenomenon , Do some super parameter optimization , Make it more appropriate to the existing experimental results .
• Using a large amount of experimental data （ At least one million ） Retraining the model , Get a model with strong specificity （ It is almost impossible to achieve , Just a guess ）

Be careful

The experimental data are fictitious , To introduce the method .