Genomics Gets New Tools for Designing RNA, Reading Variants, and Sequencing Tissue

Designing RNA that folds right

In a preprint, RiboDFM approaches RNA sequence design through backbone-conditioned discrete flow matching, and the authors report higher sequence recovery and tertiary-structure self-consistency than existing methods. The practical question behind that result is whether a designed sequence will actually fold into the intended shape, and self-consistency is the check that the sequence and the structure agree with each other rather than drifting apart. m:…4208

The same preprint adds a second useful property: nucleotide probabilities derived from the flow trajectory serve as an intrinsic measure of sequence-structure compatibility. In other words, the design process itself produces a read on how well each candidate fits its target structure, instead of requiring a separate scoring pass after the fact. The authors also report that supplementing experimental structures with models predicted by trRosettaRNA2 expands the training data available to RNA design frameworks, a workaround for the scarcity of experimentally solved RNA structures. m:…4208

Interpreting variants without a bioinformatics team

Design is only one bottleneck. A second preprint, BUDDY, targets the interpretation step: it integrates bioinformatic tools, protein structure viewers, and AI-assisted experimental design to streamline functional modeling of human missense variants in baker's yeast, Saccharomyces cerevisiae. The platform significantly reduces the bioinformatic preparation time required for ortholog variant modeling compared with traditional fragmented workflows, and it is designed so researchers without prior bioinformatics training can design and execute functional studies for clinically relevant variants, such as those associated with Lynch syndrome. m:…4639

A separate preprint tackles the scoring side of interpretation. AlphaGenome variant-impact scores and SpliceAI2 scores can be calibrated to serve as graded PP3 and BP4 evidence for missense, non-coding, and synonymous variants, giving variant classifiers a way to use model outputs as evidence tiers rather than as raw numbers. SpliceAI2 shows better discrimination than AlphaGenome scores in splicing-specific assays and higher sensitivity at 99 percent specificity across four variant classes, so the choice of model matters most exactly where splicing is the question. m:…1629

The calibration preprint also carries a caution about how classifications can move without biology moving. ClinVar likely benign classifications for variants near exons, within 20 nucleotides, show a 60-fold change in number without corresponding changes in measured splicing. That gap is a reminder that reclassification counts and laboratory measurements are different things, and that calibrated scores still need to be checked against measured splicing before they change a clinical call. m:…1629

Sequencing and counting, with less hands-on time

On the measurement side, Seq-o-Matic, an open-source automated sequencer for in situ sequencing, reduces hands-on time by 6- to 17.5-fold compared with manual methods, according to its preprint. The quality check matters as much as the time saving: the preprint reports sequencing quality, including signal retention, channel specificity, and phasing, comparable to manual sequencing. Technical variation between runs using the automated system is small relative to biological strain differences in mouse brain tissue, which is the comparison that determines whether automation is hiding the biology it is meant to reveal. m:…8341

The fifth tool closes the loop on structural variation. SLCNV automates the full copy-number-variation analysis workflow from raw low-pass whole-genome sequencing data to gene-annotated outputs, and it is portable across Unix workstations, high-performance computing clusters running Slurm, and cloud environments through Docker and Singularity. It supports both bulk and single-cell sequencing data across any species and genome assembly, so the same pipeline can serve a human cohort and a non-model organism without being rebuilt for each. m:…2049

Taken together, the five preprints describe a toolchain rather than a single breakthrough: design RNA with a compatibility signal built in, model variants in yeast without assembling a fragmented workflow, calibrate AI scores into evidence tiers, automate the sequencing bench, and count copy-number changes end to end. Each step still carries its own check — measured splicing against reclassification, manual-comparable sequencing quality, and strain-level biological signal against run-to-run variation — and those checks, not the automation alone, are what would make the tools trustworthy in practice. m:…1629 m:…8341