TL;DR - AlignIF is an RNA inverse-design framework (Nature Computational Science, 2026) that borrows the MSA idea from structure prediction and applies it in reverse: it aligns multiple structurally similar RNAs into corresponding geometric graphs and uses cross-graph attention to extract conserved 3D patterns, addressing the scarcity of experimental RNA structures.

• Architecture: per-structure geometric graphs (nodes = nucleotides; features include sugar-ring conformation, backbone/base orientation, dihedrals, inter-atomic distances) with intra-graph message passing plus cross-graph node and edge attention over aligned structures, decoded by a random-order autoregressive generator that allows fixing known functional motifs first.
• Benchmarks: 0.595 recovery / 2.221 perplexity on the benchmark set, 0.487 / 2.723 on a time-independent set, and 0.552 / 2.551 on AlphaFold3-predicted structures; even without alignment inputs it reaches 0.574, and gains grow with alignment depth. Redesigned sequences also re-fold closer to targets under AlphaFold3 self-consistency metrics.
• Ablations: removing cross-graph node updates, edge updates, or random-order decoding all degrade performance, with combined node+edge removal worst; substituting Gaussian-perturbed backbones for real homologs gives no gain, indicating the benefit comes from evolutionarily plausible conformational variation, not data augmentation.
• Wet-lab validation: 23 fluorescent aptamers and 7 self-cleaving ribozymes tested — all iMango-III/Mango-I designs fluoresced (two Mango-I designs at ~1.6x and ~1.4x wild-type, attributed to ~2x fluorophore binding capacity and more stable G-quadruplex formation by CD), and all seven pistol ribozymes cleaved (best ~11% of wild-type activity).

TL;DR - MIRACLE (published in Nature Computational Science) is a continual-learning framework for integrating single-cell multimodal data, letting reference atlases be updated incrementally with new batches instead of re-integrating all historical data. It matters because it cuts the compute/storage cost of maintaining ever-growing single-cell atlases while avoiding catastrophic forgetting.

• Built on the multimodal VAE MIDAS, it adds dynamic network expansion (new encoders/decoders inherit existing parameters when new omics/features appear) plus experience replay to retain prior knowledge.
• A distribution-preserving reservoir sampling scheme based on ball-tree partitioning of the low-dimensional embedding retains representative cells; performance plateaued at ~20k stored cells (~1/26 of the data) on a 520k-cell cardiomyopathy scRNA-seq set, and scaled to a 2.3M-cell human lung atlas with near-constant per-update time.
• Across batch/cell-type holdout CITE-seq PBMC data and "mosaic" datasets (DOTEA: ATAC+RNA, RNA+protein, ATAC+protein, tri-omic), MIRACLE matched offline integration and beat online variants of 11 methods plus Multigrate/scVAEIT; authors attribute a slight edge over offline to a curriculum-learning effect.
• Applied to COVID-19, influenza A, and tuberculosis PBMC data, continual updating expanded 13 to 23 immune cell types and surfaced CD4⁺ cytotoxic T cells mislabeled as CD8⁺, an HERC1⁺ T subset with PD-1 pathway activation in influenza, and shared MAIT-cell depletion across all three infections.