GPU-accelerated virtual drug screening
summary
GPU-accelerated pipeline turns 2.9M compounds into production predictions: feature engineering with RDKit to Morgan fingerprints and scaffold splits, training a random forest with cuDF and cuML for 20x-45x training speedups, packing the model into a GPU-backed endpoint serving subsecond (97 ms) inference, logging to BigQuery and running scheduled drift checks; demo deploys successfully and supports traffic-split model upgrades and autoscaling.
AI-written summary · about
chapters (31)
- 0:00GPU drug discovery intro
- 1:09Speakers overview
- 2:13Live demo caffeine binding
- 3:59EGFR drug binding science
- 5:23Why GPUs accelerate traditional machine learning workflows
- 7:31Drug discovery basics
- 9:43Live GPU coding session
- 11:38Molecular fingerprints features
- 17:07AlphaFold protein folding
- 18:10Lipinski rules scaffold splitting
- 21:42RAPIDS CUDA feature engineering
- 24:38GPU ML model training
- 26:29Model performance evaluation
- 28:59Notebook to production MLOps
- 30:07Deploy to GPU endpoints
- 32:20Vertex AI model registry
- 33:46GPU config for low latency
- 35:10Monitor model drift
- 36:14Log prediction requests
- 38:44Auto model monitoring Vertex AI
- 40:51Retrain traffic split models
- 42:17Custom GPU serving containers
- 44:04GPU screening pipeline summary
- 44:27Q&A scaffold splitting
- 45:55Physical validation results
- 46:36Historical data training
- 46:56GPU autoscaling costs
- 47:58Model versioning endpoints
- 49:05PCA failure molecular data
- 50:11Minimal GPU code changes
- 51:41Closing resources
related talks
01h 12mGPU-Accelerated Data Analytics in Python
01h 08mGPU Data Science with Rapids
38mRAPIDS- Open Source GPU Data Science
39mAccelerating Data Science with RAPIDS
41mAccelerating Data Science with RAPIDS
34mSpeeding up Machine Learning tasks using GPUs in Python
01h 22mHigh-Performance Data Science at Scale with RAPIDS, Dask, and GPUs
34mRAPIDS: How to accelerate your data science pipeline by orders of magnitude