FDB-17 contains 10 million fragment-like molecules selected from GDB-17’s 166.4 billion entries. Fragment-likeness filters reduce GDB-17 by 36x to 4.6 billion molecules, then even sampling across (HAC, heteroatoms, stereocenters) triplets produces a 460x further reduction to a manageable, diverse library enriched in 3D-shaped molecules.
GDBMedChem applies medicinal chemistry-inspired functional group and structural complexity filters to GDB-17, reducing 166.4 billion molecules to 17.8 billion, then evenly samples across molecular size, stereochemistry, and polarity to produce 10 million drug-like molecules. 97% of its substructures are absent from known molecule databases.
LSTNet is a deep learning framework for multivariate time series forecasting that uses convolutional layers for local dependencies, a recurrent-skip component for periodic long-term patterns, and an autoregressive component for scale robustness.
Buehler and Reymond introduce two molecular complexity measures, MC1 (fraction of non-divalent nodes) and MC2 (count of non-divalent nodes excluding carboxyl groups), derived from analyzing synthesizability patterns in GDB-enumerated molecules. They compare these measures against existing complexity scores across GDB-13s, ZINC, ChEMBL, and COCONUT.
An updated description of nauty and introduction of Traces, two programs for graph isomorphism testing and canonical labeling using the individualization-refinement paradigm.
QM9 provides B3LYP/6-31G(2df,p)-level geometric, energetic, electronic, and thermodynamic properties for 133,885 small organic molecules (up to 9 heavy atoms of C, N, O, F) drawn from the GDB-17 chemical universe. It is one of the most widely used benchmarks in molecular machine learning.
SpeechT5 proposes a unified encoder-decoder pre-training framework that jointly learns from unlabeled speech and text data, achieving strong results on ASR, TTS, speech translation, voice conversion, speech enhancement, and speaker identification.
Surge is a constitutional isomer generator based on the canonical generation path method, using nauty for graph automorphism computation. Its three-stage pipeline (simple graph generation, vertex coloring for atom assignment, edge multiplicity for bond orders) generates 7-22 million molecules per second, outperforming MOLGEN 5.0 by 42-161x on natural product molecular formulas.
VEHICLe (Virtual Exploratory Heterocyclic Library) is a complete enumeration of 24,867 mono- and bicyclic heteroaromatic ring systems built from C, N, O, S, and H. Of these, only 1,701 have ever appeared in published compounds. A random forest classifier trained on known vs. unknown ring systems predicts that over 3,000 additional ring systems are synthetically tractable.
VQM24 exhaustively enumerates all neutral closed-shell molecules with up to 5 heavy atoms from C, N, O, F, Si, P, S, Cl, Br, yielding 258k constitutional isomers and 578k conformers (836k total). Properties are computed at the wB97X-D3/cc-pVDZ level, with diffusion QMC energies for 10,793 molecules up to 4 heavy atoms. ML models show up to 8x higher errors than on QM9, making VQM24 a more challenging benchmark.
Ye et al. find that language model loss on each domain follows an exponential function of training mixture proportions. By nesting data mixing laws with scaling laws for steps and model size, small-scale experiments can predict and optimize mixtures for large models, achieving 48% training efficiency gains.
Xie et al. propose DoReMi, which trains a 280M proxy model using Group DRO to find optimal domain mixture weights, then uses those weights to train an 8B model 2.6x faster with 6.5% better downstream accuracy.