Preprint scanner project / Issue 2026-09-23
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 2026-09-23
Preprint—not peer reviewed
The authors introduce a programmable generative framework for docking, fragment design, and de novo drug design. They report experimental validation of designed or optimized ligands for five protein targets. The framework may make targeted small-molecule and non-natural-peptide design more flexible.
Preprint—not peer reviewed
The authors introduce a latent generative search framework that applies reward-guided inference-time search to steer Proteina-Complexa. They report more validated binders than other tested methods in a multiplexed phage-display screen exceeding a million designs. The authors suggest the approach can reach polar, flexible targets that have resisted previous de novo binder methods.
Preprint—not peer reviewed
The authors present ALIGNN 2.0, a dependency-free pure-PyTorch reimplementation that unifies property and force-field prediction in a single graph. They report ALIGNN 2.0 led the original ALIGNN on most single-property JARVIS-Leaderboard benchmarks. The authors suggest a common graph pipeline could improve portability and reproducibility across materials predictions.
Preprint—not peer reviewed
The authors introduce ECENet, an edge-centered line-graph neural network built on edge cluster expansion. They report state-of-the-art performance on MD22 and an accuracy-cost Pareto-frontier position after training on SPICE-MACE-OFF. The authors position edge-centered representations as an efficient foundation for physically expressive interatomic potentials.
Preprint—not peer reviewed
The authors developed Rachel, a stateful environment that executes and checks LLM-directed chemistry without setting a search policy or stopping rule. They report strict route closure across both evaluated cohorts without supplied reference routes. The authors suggest recorded revisions can carry a changed strategy into subsequent planning steps.
Preprint—not peer reviewed
The authors introduce a two-stage framework that learns a continuous reaction path before refining a transition-state candidate. They report frequency-confirmed first-order saddle-point candidates for 405 of 410 predictions. Treating paths as reusable representations could connect prediction with mechanistic interpretation and quantum refinement.
Preprint—not peer reviewed
The authors use language-model agents to run automated experiments and maintain revisable synthesis understanding. They report 18 autonomous deposition experiments that identified a narrow optimal growth window. This approach makes testable explanations, rather than optimized samples alone, an output of autonomous synthesis.
Preprint—not peer reviewed
The authors combine models for valence density, augmentation occupancies, and spin initialization into a reference-free electronic initializer. They report reducing end-to-end density-functional calculation time on unseen structures while preserving converged energies. Complete electronic initialization could remove a practical bottleneck in accelerated materials calculations.
Preprint—not peer reviewed
The authors predict catalyst endpoints by combining a Tafel term with a residual learned from labelled catalysts. They report lower error for the Ni-Ru library than the data-driven models tested. This approach could support catalyst selection beyond the measured portion of polarization curves.
Preprint—not peer reviewed
The authors trained a MACE MLIP with EMLE embedding for simulations of enzymatic, base-catalysed Kemp elimination. They report their EMLE ML/MM scheme captured the experimental barrier difference between a designed Kemp eliminase and its evolved counterpart. The authors suggest transfer from a single enzyme variant can support barrier assessment during computational screening.
Preprint—not peer reviewed
The authors present UFO-MGen, a flow-based model that uses topological information from Wyckoff representations to generate crystals. They report the highest generation-success and SUN rates under a multi-stability evaluation. The authors suggest its property-constrained generation could aid inverse materials design.
Preprint—not peer reviewed
The authors demonstrate physics-constrained agentic development of a meta-GGA functional using an LLM to assist an r2SCAN revision. They report fewer downstream rejections than uniform random search and stronger high-response candidates. The authors conclude that agentic search can support density-functional development when coupled to automated physical verification.
Preprint—not peer reviewed
The authors present miRstring, a biogenesis-aware RNA language framework for decoding the boundaries of an miRNA duplex. They report designed precursor scaffolds repressed target mRNAs. The authors suggest miRstring could connect cross-species annotation with artificial miRNA design.
Preprint—not peer reviewed
The authors introduce an agentic workflow that translates interatomic models into executable LAMMPS pair styles. They report matching reference energies and forces for all seven tested material-potential pairs. The workflow may move effort from routine implementation toward model definition and physical validation.
Preprint—not peer reviewed
The authors introduce PolyBench26, an open dataset for polymer property prediction spanning homopolymers and copolymers. They report graph-based models had the lowest prediction errors and retained their advantage as training-set size varied. The authors position the benchmark as a reproducible basis for models addressing complex polymer design.
Preprint—not peer reviewed
The authors couple high-throughput experiments with Bayesian optimization to identify thermostable messenger-RNA nanoparticle formulations. They report retained bioactivity after storage at body temperature for over two months. This could help develop vaccine formulations that reduce dependence on ultra-cold storage.
Preprint—not peer reviewed
The authors introduce an autonomous scanning-electron-microscopy pipeline that embeds micrographs and selects regions of interest. They report that ambiguity-based sampling captured more diverse and rare morphologies than random selection. This can shift automated microscopy toward general morphological surveys of unknown materials.
Preprint—not peer reviewed
The authors present SPIBER, joining SPIB with GFlowNets to reconstruct free-energy landscapes from short trajectories. They report free-energy differences recovered within a thermal-energy unit of reference values for a particle, alanine dipeptide, and AIB9. The authors suggest the method can estimate free energies without converged state populations or added molecular-dynamics simulations.
Preprint—not peer reviewed
The authors introduce a physics-informed generative framework for locating internal voids from surface displacement measurements. They demonstrate inference from both synthetic data and experimentally measured displacement fields. Its uncertainty-aware reconstructions could help design measurements for non-invasive defect characterization.
Preprint—not peer reviewed
The authors introduce OmicsBench to test reasoning about multi-omics sequences using traceable evidence chains. They report that TA-OPD strengthened biological evidence grounding while improving prediction on most tasks across the tested Qwen3.5 models. The authors suggest the benchmark and post-training approach could help identify reasoning failures in multi-omics models.