Welcome to Levitating Labs!
How we operate:
- Open research
- Molecular AI
- Honest metrics
- Zero shortcuts evaluation
Featured
Calibrated Uncertainty for DFT HOMO-LUMO Gap Prediction, and What Active Learning Did (and Didn't) Add
A 6-model PaiNN ensemble with Δ-learning and conformal intervals predicts HOMO-LUMO gaps at 0.131 eV MAE. We also report two negative results: uncertainty-guided active learning and model-based dataset curation did not beat random selection in our setup.
Physics-Guided Lattice Energy Prediction with Target-Quantile Scaffold Cross-Validation
We benchmark a 2,243-dimensional Pure Physics Neural Network against random and Bemis-Murcko scaffold splits on 23,496 organic compounds, demonstrating a 14.1% R² generalization drop under OOD scaffold shift and establishing scale-invariant nMAE evaluation.
From 29% to 96.7% ROC-AUC: How ESM-2 Language Model Fixes Topological Memorization in Drug-Target Graph Networks
We benchmark an ESM-2 + Inductive KNN protein language model architecture against traditional 3-mer counting baselines across 4 rigorous OOD splits, demonstrating +39.6% ROC-AUC gains on unseen target proteins and +67.1% on holdout disease categories.