Neural tensor factorization cuts ADMET predictions from ~340 years to under 10 minutes
CatAstro_Piyush · x · 2026-11-15
A researcher needed ADMET predictions over a 20+ trillion make-on-demand molecule library; Chemeleon GNN estimated 340 years on one GPU. Using neural synthon factorization (SynFactor), distilling a full-molecule property model into a cached rank-16 CP factorization, runtime dropped to under 10 minutes.
Benchmarks show strong agreement with the teacher model across 18 outputs (LogD, hERG, clearance, Caco-2, CYP enzymes), with Spearman ρ mostly 0.7-0.96. The project is open-sourced on GitHub under MIT license.
More from Infra
- Quanta's AI server shipments dip as GB300-to-Rubin transition slows, NVIDIA earnings due Tuesday — tengyanAI · 2026-11-15
- Investor pushes back on WSJ: Southeast Asia GPU clouds are America's legal AI backdoor — GavinSBaker · 2026-11-15
- iceberg-mcp-server: an open MCP server to query and manipulate Apache Iceberg tables — dragonejt · 2026-11-15
- ARK analyst: future AI bottleneck won't be power but reusable launch capacity — DMaguireARK · 2026-11-15
- Wyoming could fund its entire $1.7B state budget from data center revenues — panickssery · 2026-11-15
- FreeToken open-source engine runs 284B DeepSeek locally at 25 tok/s on RTX 5090 — aigclink · 2026-11-15