Biology's tooling layer levels up: nanoparticle shape controls mRNA delivery, RNA structures at scale

Shape is a dial

In a bioRxiv preprint published October 22, 2026, researchers report that the shape of lipid nanoparticles — not just their chemistry — determines how well they deliver mRNA. Increasing DSPC content in LNPs drives a transition from compact spherical to vesicular, non-spherical morphologies, and the non-spherical particles showed higher and more sustained protein expression than spherical ones across SM-102, ALC-0315 and DLin-MC3-DMA formulations. m:…0718

The interesting part for anyone formulating delivery systems: the performance gain is not fully explained by cells simply taking up more particles. The authors report that the enhanced transfection is not fully explained by increased cellular uptake — which suggests something about the geometry itself is doing work, though the mechanism remains to be pinned down. For mRNA-therapeutic developers, morphology just became a design parameter alongside lipid chemistry — structural tuning, the authors' analysis suggests, is as important as the lipid recipe itself. m:…0718

RNA structures, precomputed

A second bioRxiv preprint, also published October 22, 2026, releases the NuFold RNA Structure Database: large-scale, precomputed full-atom RNA tertiary structures, each with confidence metrics for the predicted models. The program's source code is also available for local implementation. m:…0348

The pitch is infrastructure, not discovery: precomputed structures plus confidence scores remove the need for every group to burn compute on the same predictions — the download-the-database model instead of the run-the-cluster model. The preprint's analysis frames the database as an answer to the computational bottleneck in RNA structural biology, giving researchers access to structural data without local high-performance computing. For teams building RNA therapeutics or functional-genomics pipelines, the question to ask is how those confidence metrics behave on their own sequences, where the ground truth is thinnest. m:…0348

Antibody search on a budget

A third preprint proposes an active-learning framework for antibody binding affinity prediction, aimed at campaigns with limited assay budgets and evaluated across diverse influenza and SARS-CoV-2 datasets. The headline number: a Tanimoto-kernel Gaussian process with a UCB acquisition strategy achieves 57.8% mean Recall@2% — against 9.6% for random selection — a direct measure of how much better the active learner is at finding top binders without assaying everything. m:…3720

The embedding choice matters more than the fancier models: ESM-2 650M embeddings outperformed ProtBert, AntiBERTy, AbLang2 and ProGen2 in the Gaussian-process active-learning setup. But the interpretability story is sobering. Structure-mapped Integrated Gradients and KernelSHAP analyses showed qualitative consistency with paratope geometry yet failed to provide strong residue-level discrimination — the attribution tools can gesture at where binding happens, but not yet explain which residues do the work. m:…3720

The benchmarks get audited

A fourth preprint, MutaStruct, turns the microscope on the benchmarks themselves. The authors find that existing benchmarks for mutation-induced structural change prediction are inflated by duplicate mutation events and highly similar structural pairs — and that a simple Copy-WT diagnostic control remains highly competitive on backbone-based metrics. In plain terms: the tests may be flattering the models. m:…5028

The proposed fix is a stricter curation pipeline: residue-level sequence validation, coordinate-level quality control, context-aware filtering, event-level deduplication, and leakage-aware data splitting. Step back and the through-line across all four preprints is measurement itself: the field is upgrading its tools — a new design knob, a new database, a smarter search strategy, and the tests that keep everyone honest. m:…5028