Keynote: Scientific Discovery: From the Lab Bench to the GPU
summary
A production platform called BioNeMo combines a training framework (based on NVIDIA NeMo) and an inference service accelerated on GPUs. A small‑molecule language model MIM uses a perceiver encoder and mutual information training to cluster chemical space and supports control generation with CMA‑ES oracle optimization (four oracles: two property, two binding). The model is productized and ships soon via the API. Team size grows from two to 40 developers and the approach is working though not yet state of the art.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:04PData 2013 to drug discovery
- 2:28AI in early drug discovery
- 4:12Language models in science
- 6:01Protein folding with deep learning
- 7:49Introducing Bimo
- 12:06Bimo for GPU deep learning
- 14:54Molecule language model MIM
- 25:04Career journey lab to AI
- 34:15Team growth and future
- 35:04Q&A
related talks
45mKeynote: Panel(...)
23mKeynote: AI, Quantum, and AI + Quantum
29mBuild Customize and Deploy LLMs At-Scale on Azure with NVIDIA NeMo
23mNeMo: Open Source Toolkit for Conversational AI
12mKeynote: Build Accelerated LLM Applications with Open Source Tools from NVIDIA
19mThe Small Model Infrastructure Nobody Built (So We Did)
31mScientific Machine Learning With NVIDIA PhysicsNeMo
20mWhat's new in Machine Learning Keynote