Automated Discovery of Cancer Types from Genes
summary
Applies a variational autoencoder (Keras/Tybalt) to TCGA gene-expression from ~11,000 tumors across ~32 cancers, compressing 5,000 selected genes into a 36-dimensional latent space, then projects that space with tSNE into an interactive tSNE plot (Highcharts) to reveal clusters that align with known types and cross-organ similarities; uses a Random Forest on the latent features to validate structure and reports ~98% classification accuracy, all run in Jupyter/Google Colab.
AI-written summary · about
chapters (21)
- 0:00Intro
- 1:20CURRENT METHODS
- 2:01BIOLOGICAL EXAMPLE
- 4:10CANCER CELLS
- 5:16GENE EXPRESSION PROFILES
- 5:54MACHINE LEARNING
- 6:10LIMITATIONS OF SUPERVISED LEARNING
- 7:14UNSUPERVISED LEARNING
- 8:22AUTOENCODER: A TYPE OF DEEP NEURAL NETWORK
- 11:03VARIATIONAL AUTOENCODING
- 12:03THE GREENE LAB AT UPENN
- 13:30INFORMATION WITHIN THE DATASET
- 15:08STEP 2: BUILD AND TRAIN VAE USING KERAS
- 16:23EVALUATING THE VAE'S LATENT REPRESENTATION
- 19:08CREATING
- 20:27OBSERVATIONS
- 22:02CONFUSION
- 22:56IDEAS I'M PURSUING: GENESET-36 CANCER PANEL
- 24:29HYPOTHESIS
- 24:54IDEAS I'M PURSUING: LATENT REPRESENTATION
- 25:48POP: THE POWER OF PYTHON
related talks
12mHow data science can make your genome actionable
35mFrom cells to drug responses - machine learning in cancer research
27mKubernetes + AI Joining Forces in the Battle Against Cancer
33mUnderstanding Molecular Complexity for Precision Medicine
39mIdentifying Disease Mechanisms Using Gene Networks as Priors
32mHow machine learning helps cancer research
29mMachine Learning Techniques for Omics Data Analysis
29mCancer diagnostics enabled by machines: finding and classifying single tumor....