Solutions

CNN model trained with UMAP feature transformed scatter plotted images on CLL samples

Data Augmentation + CNN Model Training

1. Solution to the limited patients sample size: Data Augmentation with multiple templates embedding (Figure.3a)

• Enumerate for pairs of samples with one positive and one negative labels.

• Sample UMAP template embedding: embed transformed sample to the templates pairs that marked in the previous steps, and

• Saved as inputs for CNN model training.

2. Hyperparameter optimization with Gridsearch

3. CNN model training

• Split training testing, and validation datasets

• Train model with augmented datasets and the optimal parameters

• Obtain accuracy and predictions on testing datasets

4. Optional ensemble learning algorithm XGBoost was utilized to combine multiple models for the optimal performance.

1. UMAP Construction and feature transformation, Multiple templates + Single Positive Voting

ALL patients and non-patients with labels, 60 samples in total for training/cross-validation, 29 samples without labels for testing

a. To accommodate the varying endotypes among ALL samples, we construct 5 templates (based on 2 different samples from ALL patients and 2 samples from non-patients) as a single training set (Figure 4), then apply HDBSCAN for cluster classification

b. Prediction, Training and testing via cross validation. Get the prediction results for each template, check if there is any outliers

c. Single-positive voting: Consider the predictions of the 5 templates collaboratively, mark the sample as “ALL” as long as one of the model gives a true prediction

d. Repeat the above steps for a new training set, following the 4-folds cross-validation procedure