Preprint scanner project / Issue 2026-07-08
A weekly review of arXiv, ChemRxiv, and bioRxiv for ai-enabled methodological advances. Up to twenty papers reviewed every week; a quiet week may publish none.
Issue 2026-07-08
Preprint—not peer reviewed
The authors present a multi-agent language-model pipeline that classifies reactions and writes rules under a verification loop. They report that a lightweight fingerprint classifier classifies unseen reactions while resolving chemistry more finely. This could make reaction classification adaptable to new chemistries.
Preprint—not peer reviewed
AquaGen generates all-atom configurations with explicit solvent and periodic boundaries from a Boltzmann distribution. For absolute hydration free energy, the authors report estimates four to ten times faster than standard graphics-processor molecular dynamics with comparable accuracy. Generating ensembles at atomistic resolution keeps free-energy estimates connected to force-field evaluations and subsequent molecular dynamics.
Preprint—not peer reviewed
The authors present SpinGTP, which generalizes scalar functions to spin-weighted spherical harmonics. They report accuracies comparable to full CGTP and superior performance for chiral materials and non-centrosymmetric geometries. This could broaden scalable modeling of three-dimensional atomistic systems.
Preprint—not peer reviewed
EquiFiLM adds an external conditioning signal to equivariant force fields through a per-layer modulation block acting on scalar channels. In charged liquid water, the authors report stable molecular dynamics and a charge-dependent first-shell response across the tested range. This provides a compact way to extend a foundation potential toward driven electronic conditions while preserving geometric symmetry.
Preprint—not peer reviewed
MARLIN infers a molecular fingerprint from spectral peaks and generates candidate structures without supplying a molecular formula. On the NPLIB1 benchmark, the authors report that it was the strongest formula-free method across exact matches, structural distance, and fingerprint similarity. Formula-free structure elucidation could better match discovery settings where the identity of an unknown compound is genuinely open.
Preprint—not peer reviewed
The authors introduce PGLD, which uses a synthesis-aware parametrization of stochastic DNA libraries. They report using PGLD to design a library to explore the mutation space of a broadly neutralizing influenza antibody. This could connect AI-based sequence design with high-throughput experimentation.
Preprint—not peer reviewed
The authors train 3D-PRIMME on two successive time steps to learn three-dimensional grain-growth dynamics. They report that the learned operator retained kinetics and grain topology when applied to much larger domains without retraining. The local rule could make large-scale microstructure prediction more tractable when physical scaling behavior must be retained.
Preprint—not peer reviewed
The multi-minima iterative genetic algorithm combines a learned interatomic potential with penalties intended to move the search beyond explored minima. Given composition alone, the authors report an orthorhombic structure for La5CoPb2 that exactly matched an independently determined diffraction structure. Searching global and competing metastable states could inform phase selection for unfamiliar materials compositions.
Preprint—not peer reviewed
The authors evaluate three active-learning strategies for whole-cell bacterial bioactivity. They report a five-fold increase in experimental screening hit rate. This approach could help address data-acquisition bottlenecks in whole-cell antibiotic discovery.
Preprint—not peer reviewed
Dyna-Mat supplies condensed-phase first-principles molecular-dynamics trajectories for testing foundation interatomic potentials at finite temperature. The authors report individual systems where small force errors nevertheless produced qualitative errors in predicted structure. Evaluating trajectories and observables alongside static errors can expose failures that matter during materials simulations.