Garrett Goh, Scientist, Pacific Northwest National Lab at MLconf Seattle 2017
summary
Convolutional nets trained on molecule images derived from SMILES→image representations (using open-source decoders) predict HIV activity, Tox21 toxicity (12 tasks) and solvation free energies with architectures like Inception-ResNet-v2, AlexNet and ResNet; a 45‑layer, 16‑filter design and 3–4 channel augmented images (adding simple chemical info) produce higher AUCs than descriptor‑based logistic/RandomForest/MLP models and push solvation RMSE to about ~1 kcal/mol—near physics‑based accuracy—while noting labeled data scarcity and one‑model‑per‑property limits.
AI-written summary · about
chapters (22)
- 0:00Intro
- 0:34Recent Trends in Deep Learning
- 1:35What is Deep Learning?
- 2:15Why Deep Learning today?
- 2:43What makes Deep Learning better than traditionaVshallow Machine Learning?
- 3:26A Case Study of Deep Learning Success in Computer Vision
- 4:01A Short History on Feature Engineering in Chemistry
- 5:29Deep Learning for Computational Chemistry
- 5:53Deep Learning as a Machine Learning Tool in Scientific (Chemistry) Research
- 6:45Deep Learning as "Machine Intelligence" in Scientific (Chemistry) Research
- 7:28Designing a Deep Learning Framework with Minimal Chemistry knowledge
- 8:29Deep Learning predicts Physiological, Biochemical & Physical Properties
- 8:59Experiments with Different Deep Neural Network Architectures
- 9:59Chemception Deep Neural Network
- 11:31Tweaking Chemception (Depth & Width)
- 12:03Benchmarking Chemception Performance
- 12:41Activity Prediction Results
- 13:08Toxicity Prediction Results
- 14:22Improving Chemception Performance
- 17:17Conclusion
- 19:45Future Challenges
- 22:23Acknowledgements
related talks
31mML methods to accelerate the drug discovery cycle
36mKeynote: Scientific Discovery: From the Lab Bench to the GPU
37mRecent Solutions Using Deep Learning Models/ The How and Why
03h 51mMachine Learning Summit
41mDeep Learning to Solve Challenging Problems
40mDeep Learning for QSAR
01h 23mGraph Neural Networks in Drug Discovery
23mUsing Graph Nets (GNs) to predict molecular properties