Machine Learning For Formulations Containing Proteins

1. Double-click the Materials Science icon
   • (No icon? See Starting Maestro)

Figure 2-1. Change Working Directory option.

2. Go to File > Change Working Directory
3. Find your directory, and click Choose
4. Pre-generated files are included for running jobs or examining output. Download the zip file here: schrodinger.com/sites/default/files/s3/release/current/Tutorials/zip/ml_formulation_proteins.zip
5. After downloading the zip file, unzip the contents in your Working Directorythe location where files are saved for ease of access throughout the tutorial

Figure 2-2. Save Project panel.

6. Go to File > Save Project As
7. Change the File name to ml_formulation_proteins_tutorial, click Save
   • The project is now named ml_formulation_proteins_tutorial.prj

Figure 3-1. Loading the training data into the Formulation Machine Learning panel.

1. Go to Tasks > Materials > Informatics > Formulation Machine Learning
   • The Formulation Machine Learning panel opens
2. For Formulation type, choose Simple
   • We want to train a model for predicting a single mixture instead of a blend of mixtures
3. Click Load CSV, navigate to the provided files and select Section_03 > yogurt_formulations_input.csv, and click Open 
   • The training data set is loaded into the panel
   • You can find references for the data in the last column of the imported csv file

Figure 3-2. Visualizing the firmness distribution of the data set.

The formulations in this data set contain 1-2 proteins as well as one additive for a Calcium source (columns 1 to 3). The following columns display some properties of the formulations (pH, Temperature, etc).

The first two columns of the formulations , which contain the proteins, show missing SMILES for most formulations. Proteins are not easily described via SMILES strings.

4. For View ingredient as, choose Labels
   • Now we can see the labels as defined in the data set
5. For the plot on the right side, choose Firmness (mN) as the first of the  Properties to be displayed
   • The distribution of the firmness is shown for the imported training data set
   • We will generate a model for the prediction of the firmness of our yogurt formulations

Figure 3-3. Settings for building the model.

6. Switch to the Build tab
7. For Featurizers, choose MACCS Keys, Fingerprint, and User Input Features Only (3 items)
8. For Machine learning models, choose Elastic Net, Random Forest, XGBoost, Support Vector Machine, and Dense Neural Network (5 items)
9. For Target property, choose Firmness (mN) (1 item)
10. For Formulation descriptors, choose pH and Temp (oC) (2 items)
11. For Protein descriptors, click Define Proteins
    • The Link Missing SMILES to Proteins dialog box opens

Figure 3-4. Linking the protein sequences to the labels.

The proteins used in the data set are: (Seq_A) Alpha-Casein, (Seq_B) Beta-Casein, (Seq_K) Kappa-Casein, (Seq_W) Beta-Lactoglobulin. Now we have to link the ingredients to either 3D structures or sequences.

12. For Protein Input Type, choose Sequences
    • We are going to link the amino acid sequences for the proteins to their ingredients. It is also possible to link 3D structures from the Project Tabledisplays the contents of a project and is also an interface for performing operations on selected entries, viewing properties, and organizing structures and data instead.
13. For Seq_A click Enter Sequence in the Action row
    • A new window opens
14. Copy the sequence for protein A (Seq_A) from the provided file Section_03 > yogurt_protein_seq.csv and paste it into the the field for Sequence for Seq_A
15. Click OK
16. Repeat steps 13 to 15 for Seq_B, Seq_K, and Seq_W
17. Click OK
    • A window informs you that 4 missing SMILES ingredients have been linked to protein sequences

Figure 3-5. Calculating protein descriptors and loading them as ingredient descriptors.

18. Change the Job name to formulation_ml_calc_protein_descr
19. Click Calculate Protein Descriptors
    • This job should take about 1 minute on a single CPU

Note: When the job is complete, no new entry group is added to the entry lista simplified view of the Project Table that allows you to perform basic operations such as selection and inclusion.

20. A popup window appears, asking if you would like to load the calculated protein descriptors as ingredient descriptors. Click OK.

Figure 3-6. Running the job.

21. Change the Job name to formulation_ml_train
22. Adjust the job settings () as needed
    • The job can be run on a single CPU and takes approximately 5 minutes to complete
23. If you would like to run the job yourself, click Run. Otherwise, proceed with the provided results from the zip file.

Note: When the job is complete, no new entry group is added to the entry lista simplified view of the Project Table that allows you to perform basic operations such as selection and inclusion.

Figure 3-7. Loading the ML model.

24. Upon completion of the calculation, click OK in the popup window. The ML Model Manager panel opens automatically and the trained model appears in the list.

If you want to load the provided model, switch to the Performance tab and click Load Model to open the ML Model Manager panel. Click Add Model, browse to Section_03 > formulation_ml_train > formulation_ml_train.mlform, and click Open.

25. Select the formulation_ml_train entry and click OK
    • If you get a warning saying that the uploaded ML model was trained on an older version, feel free to click OK and proceed. If you wish to retrain the model, you can use the ML Model Manager panel.

Figure 3-8. Visualizing the results.

The plot contains predicted versus actual values from the train and test sets, along with the corresponding distributions of the data sets for both predicted and actual values. The table below shows corresponding R² and RMSE values for both data sets.

In this case, the model generalized well on the test set, even though both the training and test sets were quite small (90 and 10 data points, respectively). This results in lower correlation (R²) and higher errors (RMSE). A larger data set would very likely improve the performance.

Figure 3-9. Calculating feature importance.

26. Switch to the Feature Importance subtab
27. Change the Job name to formulation_ml_feature_importance
28. If you would like to run the job yourself, click Run. Otherwise, proceed with the provided results from the zip file.
    • This job should take about 1 minute on a single CPU
    • Feature importance is not available for Set2Set and Graph-based ML models

Note: When the job is complete, no new entry group is added to the entry lista simplified view of the Project Table that allows you to perform basic operations such as selection and inclusion.

Figure 3-10. Loading the model with the feature importance results.

Click Add Model, browse to Section_03 > formulation_ml_train > formulation_ml_train.mlform, and click Open.

29. Click Load Model
    • The ML Model Manager panel opens
30. Click Add Model
31. For loading the provided model, browse to Section_03 > formulation_ml_feature_importance > formulation_ml_train_xai.mlform, and click Open
32. For loading your own model, browse to formulation_ml_feature_importance > formulation_ml_train_xai.mlform, and click Open
33. Select the formulation_ml_train_xai entry and click OK
    • If you get a warning saying that the uploaded ML model was trained on an older version, feel free to click OK and proceed. If you wish to retrain the model, you can use the ML Model Manager panel.

Figure 3-11. Visualizing the feature importance.

34. You can click on a descriptor name to view a description at the right
35. You can use the arrow buttons below the plot to view more descriptors

Figure 3-12. Visualizing and comparing featurizer-model type pairs.

36. Switch to the Model Comparison subtab

This histogram shows the 5-fold cross-validation R² score, or the CV score, for all featurizer-model type pairs. The error bars show the minimum and maximum CV scores for pairs that appear more than once in the trained model ensemble.