To predict the effect of Genetic Variants to enable Personalized Medicine

Every year, there are 7 million patients who suffer from cancer worldwide, and 5 million people die of cancer. 60% of patients can only survive for around 5 years after diagnosis. However, no matter what kind of cancer, there are specific changes in the genome in terms of the specific type of tumor (typing) or the different stages of development.

The genome changes (causes) cell differentiation, Abnormal development, and growth transmission which may be caused by gene mutation or genetic inheritance. In order to avoid a huge amount of manual work for the researchers, machine learning & deep learning plays a very important role in this territory. For example, a cancer tumor may have thousands of genetic mutations in a DNA sequence and the challenge is to distinguish the mutations that contribute to tumor growth from the neutral mutations. Nowadays, a lot of researchers still study how genetic testing can help us to find a way to treat these cancers.

This project is based on text-based clinical literature, annotated mutations, and Genes Where Mutation is Located, I developed some Machine Learning algorithms using the genomics knowledge to automatically classify genetic variations. The data is from Kaggle, and the attribute information includes ID and the Gene where this genetic mutation is located; Variation — The amino acid change for mutations; Text — The clinical evidence used to classify the genetic mutation; The Class column contains the genetic mutation that has been classified on.

• Gene: The gene where this genetic mutation is located)
• Variation: The aminoacid change for this mutations)
• Text: The clinical evidence used to classify the genetic mutation
• Class: 1-9 the class this genetic mutation has been classified on

The bar chart shows the top 30 genes where mutations are located. The first place is BRCA1 and you can see there is a gene named BRCA2 in fifth place. What are they? The name “BRCA” is an abbreviation for the “Breast Cancer Gene.” BRCA1 and BRCA2 are two different genes that have been found to impact a person’s chances of developing breast cancer. Additionally, the changes in the TP53 gene greatly increases the risk of developing breast cancer, bladder cancer, and lung cancer, etc. Besides, there are around eight mutations in the EGFR gene have been associated with lung cancer.

After knowing those genes and the potential risks of their mutations, Let us take look at bar charts for each class. Obviously, one gene could be classified in different classes. For example, the gene BRCA1 is almost in all classes, which means one type of gene could have more than one mutation. The gene like “PTEN” predominantly presents in a single Class. On the other hand, there are a few genes in Class 8 and Class 9.

In our dataset, there are a total of 2993 different variations and each variation matches one or more genes. Not all variations can cause cancers but they may influence cancer risk.

In Class 1, the one has most frequency is Truncating Mutations which is in second place in Class 6. Fusions, Deletion, Amplification have high frequencies in Class 2,4, and 6&7 respectively. On the other hand, the number of each variation’s frequency is one (1) in Class 3,5,8 and 9.

After data engineering, data pre-processing, and data vectorization, we can go ahead to start dive into our modeling part. The picture below is the process of the modeling part. I analyzed each column individually in the first stage using different models — Naive Bayes, Neural Network, and Logistic Regression. In the second stage, all of those outputs are collected as new features for the next stage modeling part — XGboost.

• Train data f1 score:0.6157672054520656
• Test data f1 score:0.5930716695679839

• Train data f1 score:0.9774319139048305
• Test data f1 score:0.8126132108732343

• Train data f1 score:0.798351080032705
• Test data f1 score:0.7846091468870092

• Train data f1 score:0.6157205891605201
• Test data f1 score:0.5939324450221499

• Train data f1 score:0.9780175131194836
• Test data f1 score:0.9038712574030837

• Train data f1 score:0.9014796254055549
• Test data f1 score:0.8585273978514553

• Train data f1 score:0.615936855192663
• Test data f1 score:0.5957018668273237

• Train data f1 score:0.9769987251836152
• Test data f1 score:0.9029228363338574

• Train data f1 score:0.9039234128069726
• Test data f1 score:0.8628257154404384

Let us take a look at the new train and new test data. Each column contains the predicted values from three models.

• F1 Score of train data : 0.99
• F1 Score of test data : 0.92

-Before model stacking, the highest F1 score from neural network is 0.90. After model stacking, the F1 score is 0.92. The final F1 score increased 2% Based on the confusion matrix below, we can see that the model stacking method worked very well.

We may obtain higher accuracy and f1 score after we apply model stacking, however, it is hard to explain and interpret the results. Before model stacking the highest accuracy is from the neural network model and the predictor is Variation. The accuracy and f1-score are 89.61% and 90.29% respectively. After model staking the accuracy and f1-score are 91.48% and 92.04% respectively. Obviously we get better results and the accuracy & f1 score increased around 2%. Based on the confusion matrix below, most classes were predicted very well, but our model did not predict very well for Class 7. If we are able to get a large dataset, the accuracy and f1 score will increase for sure. Besides, the bigger dataset we have, the better results the Neural Network generates.