ZeUS lets existing 3D-aware AI models design large symmetric protein complexes without retraining by focusing on one repeating unit, lowering computing cost. Symmetric proteins are fundamental to biology and critical for developing new vaccines and advanced materials.
Efficient and Accurate Zero Shot Generation of Symmetric Protein Complexes · Rory Gao, Yuanzhou Chen, Prajit Rajkumar et al. · arXiv:2611.10424
Destructive sequencing never lets you watch the same cell twice, so stMIND tries to fill in the movie between snapshots by splitting cell velocity into three parts: an intrinsic program, a local graph-attention term for nearby cells, and a nonlocal mean-field term for whatever the rest of the tissue is doing (the paper). The authors test the framework on axolotl and zebrafish data, where it beats baselines on expression prediction and spatial coherence, and ablations show both the local and nonlocal pieces are needed. The method flags a regeneration window in axolotl telencephalon, ten to fifteen days after injury, where drift energy spikes before nonlocal signals take over and local effects push back depending on cell type.
stMIND: Learning multiscale interaction dynamics from spatial-temporal transcriptomics · Longquan Li, Lin Wan · arXiv:2611.12228
Protein structures rarely sit still, and OriDynamics tries to generate a broad range of conformations a protein and its partners can adopt, instead of a single frozen pose (the paper). The authors split the generation into a compact latent space for broad conformational exploration, with local all-atom sampling layered on top, and add supervision on mean position, fluctuation amplitude and coordinate covariance. Against a set of existing baselines it reportedly wins on conformational coverage and interaction geometry, and the authors show it can recover alternate functional states along with configurations associated with ligand dissociation. The paper cites a 2026 work on physically grounded generative modeling of all-atom dynamics, with a neutral stance.
OriDynamics: Towards a World Model for Biomolecular Dynamics · Jiahao Yang, Dong Liu, Hongliang Li et al. · arXiv:2611.11958
Existing trajectory-inference tools mostly track one molecular layer at a time, and COATI tries to track RNA and ATAC states together even though sequencing destroys each cell before a second measurement can happen (the paper). The authors extend unbalanced synchronized optimal transport into cross-modal, time-dependent maps, letting the method reconstruct held-out developmental stages and recover chromatin heterogeneity within a single cell type. The unbalanced setup also tracks population mass changes, which the authors tie to DNA-replication programs early in development and lineage-maturation programs later on. Lineage responses from COATI reportedly agreed with independent CRISPR perturbation results, and inferred regulatory patterns were consistent with separate epigenetic data, which lends some outside support.
Synchronized Optimal Transport Reconstructs Coupled Transcriptomic and Chromatin Dynamics · Jingfeng Wang, Yanxiang Zhao, Qing Nie et al. · arXiv:2611.08853
A method extracts precise tissue properties, such as light absorption, directly from heart scans made with Optical Coherence Tomography, reversing image physics instead of needing much training data. Doctors could get objective, per-pixel measurements of disease progression from existing scans, beyond visual assessment.
Differentiable Inverse Rendering Recovers Intrinsic Properties from Intravascular Optical Coherence Tomography · Ajay Manicka, Karim Kadry, Sebastian Ruiz-Lopera et al. · arXiv:2611.12686
A protein scaffold can nail its static refold and still flop once you let it move. That's the gist of this new benchmark, which argues past motif-scaffolding tests leaned on small hand-picked motif sets and single-structure recovery as the finish line. The authors built BILBY-PRO, 220 cases pulled from 174 PROSITE motif patterns, and ran six scaffold generators through it: RFDiffusion and Proteina solved more targets, ESM3 produced more unique working scaffolds on what it did solve. The real gut punch is BILBY-MD: of 55 statically solved model-problem pairs, only 31 held up under molecular dynamics checks on motif drift, scaffold drift, motif-contact preservation, and motif-dihedral free-energy landscapes, and just 8 passed every check. Compact motifs often survived; interface-dependent ones failed more frequently. The Hugging Face listing shows barely a ripple of attention so far, while the BILBY-MD code confirms it picks one scaffold per problem for these MD runs.
Static Recovery Is Not Dynamic Stability: Dynamics-Aware Benchmarking of Protein Motif Scaffolding · Emile de Bruyn, Moritz Schäffler, Jannik Schneider et al. · arXiv:2611.07723
scBIG++ predicts how individual cells change gene activity after genetic or chemical interventions, considering all genes and varied cell responses. It could speed gene-function discovery and therapy development by giving a clearer picture of cellular responses.
scBIG++: Learning Across Genes and Cells to Predict Single-Cell Perturbation Responses · Jiafa Ruan, Zican Qiao, Zongxin Yang et al. · arXiv:2611.11217
Gene text descriptions fed to models can contain answers to the tests, making models look better at predicting gene function, especially for less-studied genes. Current gene-learning benchmarks may overstate true understanding, so their reliability needs re-evaluation.
When gene summaries reveal the answers: Gene Ontology circularity in gene embedding benchmarks · Jiawei Wu, Huishu Wang, Junlin Chen et al. · arXiv:2611.11118
AI can detect sex, age, or self-reported race from routine microscopic tissue images, with accuracy varying by organ and trait. This helps show whether medical AI might use demographic shortcuts instead of purely disease-related features.
Health-System-Wide Mapping of Demographic Information in Routine Clinical Histopathology · Shengjia Chen, Ruchika Verma, Kevin Clare et al. · arXiv:2611.09851
Choosing the right kinetic model for a PET voxel used to mean grinding through nonlinear least squares and hoping noise doesn't trick the Akaike criterion into picking the simplest option. This paper swaps that for a consistency model that samples kinetic parameters and model choice together from one posterior, conditioned on the measured time-activity curve, in seconds per slice on a four-A100 workstation. On simulated whole-body [18F]FDG total-body PET across seven compartment models, it hit 97.97 percent selection accuracy versus 58.67 percent for the old NLS+AIC approach; in simulated brain [11C]raclopride data distinguishing smoking-induced dopamine release, it scored 91.4 percent against 73.5 percent for NLS+BIC. Clinically, it produced whole-body model-selection maps matching literature-expected models per organ and flagged brain voxels where smoking triggered dopamine release.
Bayesian Kinetic Model Selection in Dynamic PET via Consistency Models · Yun Zhao, Qinlin Gu, Andrew J. Reader et al. · arXiv:2611.12235