Novozymes Enzyme Stability Prediction

Link: https://www.kaggle.com/c/novozymes-enzyme-stability-prediction

Problem Type: ordering objects in list

Input:

• you also get alphafold2 predictions (wildtype) of the enzyme variants
  • these variants are found via “point amino acid mutation and deletion”
    • This is why they are called mutations
  • you can see how the mutations were made in this notebook: https://www.kaggle.com/code/gehallak/nesp-3d-geometry-0-32-lb

Output: the predicted rank thermal stability of enzyme variants

• officially, you want to output n tuples (seq_id, tm) - tm is the melting temperature of the enzyme
  • Higher tm means the protein variant is more stable.
• Since only the Spearman Correlation Coefficient will be used for the evaluation, the correct prediction of the relative order is more important than the absolute tm values.
  • Basically, the tm value that you predict is just used to order your predictions (before running spearman’s correlation coefficient with the t_true)

Eval Metric: Spearman’s correlation Coefficient

Summary

• Enzymes are proteins that act as catalysts in the chemical reactions of living organisms.

• Novozymes finds enzymes in nature and optimizes them for use in industry.

• why do they care about the tm levels?

  • cause a difference in the melting point (and pH) can affect the efficiency of the enzyme
    • not sure why they care about ranking though

• glossary:

  • “residue” are amino acids and nucleic acids

• Note: there were training data issues: https://www.kaggle.com/competitions/novozymes-enzyme-stability-prediction/discussion/356251

Important notebooks/discussions

• EDA and explaining what this competition is

  • https://www.kaggle.com/code/dschettler8845/novo-esp-eda-baseline
  • The frequency of some amino acids respective to the length of the AA sequence has a low, but non-zero correlation with the target tm�� (melting point)
  • The fractional frequency appears slightly more informative
  • Some amino acid sequences are REALLY long… especially considering all of our test AA sequences are relatively short ~240 vs more than 32,000
  • While some amino acids are more common than others, all are relatively common (Sequences on average contain 1.5% to 10% of each amino acid respectively).
  • ph is from 0.0-11.0
  • there is a negative tm (target) value. why?

• highest public kernel before the contest end

  • https://www.kaggle.com/code/seyered/eda-novozymes-enzyme-stability
  • He basically used a bunch of techniques from many different papers
    • these techniques tended to look at the sequence of the acid to predict the tm
    • then he ensembled the tm predictions of the techniques using this
      • code

         Global ensemble:
        testDf['tm'] = (
            4 * rank_nrom('rosetta') + 2*rank_nrom('rmsd') + 2*rank_nrom('thermonet') + 2*rank_nrom('plddtdiff') +\
            rank_nrom('sasaf') + rank_nrom('plddt') + rank_nrom('demask') + rank_nrom('ddG') + rank_nrom('blosum')
        ) / 14

    • the way he found the coefficients in the ensemble is by increasing 1 to the weight until it led to a lower public score
      • yep they were definitely overfitting.
  • interesting to note: ppl said that CV and the public LB were unstable

• Difference features

  • https://www.kaggle.com/code/cdeotte/difference-features-lb-0-600
    • PDB files are protein data bank files. they describe the three-dimensional structures of molecules
    • he took a dataset of PDB files to make features

• Showing that there is a high correlation between the distance between the mutation’s (xyz) and the enzyme’s center (xyz)

  • https://www.kaggle.com/code/gehallak/nesp-3d-geometry-0-32-lb
    • by using the sole feature mutation coord - enzyme center coord as the predicted value of tm, he got a public score of 0.32

Solutions

• (1st) Protein as a Graph

  • https://www.kaggle.com/competitions/novozymes-enzyme-stability-prediction/discussion/376371
  • “there was a big shakeup, so this isn’t the best solution”
  • Nodes represent residues, and if two residues are closer than a certain distance, the nodes are connected by edges.
  • The features of each node are the embedding vectors of wildtype sequence and mutant sequence generated by HuggingFace’s facebook/esm2_t33_650M_UR50D
  • graph centrality -  measure of the influence of a node in a network
  • they mention dTm, but nobody else mentioned this.
    • I assume it’s the difference between tm and (mutation coord - enzym center cord) in the train data
  • this was the code he used to create the graph from the pdb (3D coords of protein) files:
    • https://www.kaggle.com/code/gyozzza/create-graph-data-from-pdb-files-for-gnn
    • how it works:
      • each node in the graph is an amino acid in the pdb file (which tells us its xyz)
      • 1. for each protein, create the distance matrix, which is the euclidian distance between any two amino acids in the protein
        • 2. get the adjacency mat for amino acids if their distance is < a threshold:
        • adj = distance_matrix < distance_threshold
        • u, v = np.nonzero(adj)
    • uses https://www.dgl.ai/
  • I don’t understand what they mean when they say that they are using a graph isomorphism network.
    • are they comparing the mutation of the enzyme to the original enzyme, and using the original enzyme as a reference for the mutation, so that the tm value is just an offset of the tm of the original enzyme?
    • I guess training an IGN outputs an embedding that represents the graph. now that they have an embedding, they can use it as a feature vector for downstream models

Takeaways

• TODO