Thrilled to share our paper in @NatureComms
RNA-binding proteins can enhance transcription largely independently of condensation
Nascent RNA recruits FUS to transcription site, boosting RNAP loading and preventing R-loops to enhance elongation.
https://t.co/6KC1knQYkR
1/3 🚀 We’re excited to introduce CryoPilot!
Two AI agents take cryo-EM data from raw images to refined atomic models—making decisions, testing strategies, and keeping the analysis moving while you focus on your science!
https://t.co/vIvF0fgXRg
#CryoEM#AIAgents
How does human NuA4/TIP60 recognize and modify nucleosomes? Our work reveals cooperative recognition involving the acidic patch and linker DNA and uncovers an interaction between the Tudor domain and catalytic core that regulates HAT activity.
Preprint: https://t.co/h91vtCpXFr
We’re sharing how our biology research team worked with Claude to map the first 3D structure of a vault particle from bacterial life.
Vault particles are a longstanding mystery. They are barrel-shaped objects, among the largest protein structures found in our cells. They were first discovered by Leonard Rome and Nancy Kedersha in 1986, but their function remains unknown.
With Claude, we’ve found results that might bring us closer to solving that mystery.
Until now, vaults were only known in the cells of complex things like animals and plants. We have shown that similar proteins from bacteria (and the viruses that infect them) can also assemble into vaults. Finding the structures in these simpler forms of life may make it easier to investigate their function in our own cells.
For our life sciences team, the project was also a demonstration of the potential (and the limitations) of using AI to accelerate science.
In our full piece, we discuss how Claude was able to dramatically speed up some stages—manually determining the vault particle’s structure would have taken weeks to months of painstaking work, but here it took days. We also discuss how Claude made some simple mistakes along the way, and how we collaborated with our machine learning researchers to address them.
Read more: https://t.co/KowVVCSPp8
In Science, researchers present RFdiffusion3, a new generative AI system that designs proteins and nucleic acids atom by atom, which could enable the ability to engineer novel biomolecules with precise, complex functions.
Learn more: https://t.co/ARUjcNoODV
Dyna-1, developed by @DorotheeKern and colleagues, predicts micro-to-millisecond protein dynamics in roughly 30 seconds. Published in @Nature, the model was trained on missing NMR assignments to identify protein regions likely to move.
More: https://t.co/Ww4tRdcczE
CryoLigATE: Enhancing the resolvability of cryo-EM maps in protein-ligand complexes using deep learning
1 CryoLigATE is a deep-learning post-processing framework tailored to a persistent cryo-EM bottleneck: ligand-binding sites often look substantially worse than the surrounding protein, limiting reliable atomic interpretation for structure-based drug discovery.
2 Unlike general cryo-EM map enhancers that are mostly protein-trained (and can even worsen ligand-site metrics), CryoLigATE is trained specifically to recover ligand density features (functional groups, continuity, topology) while preserving already high-quality experimental signal.
3 The training set is built for chemical diversity: 6,511 unique cryo-EM protein–ligand pairs (global resolution ≤4 Å) spanning drug-like molecules (59.3%), lipids (12.2%), nucleotides (9.7%), plus steroids, carbohydrates, peptides, and more; the dataset also covers diverse protein families including many membrane proteins.
4 Method overview: the model takes an experimental map and automatically extracts a 32 Å cube around a specified ligand site using a preliminary atomic model only for localization; it then predicts an enhanced local map in seconds on a desktop GPU, with no manual map preparation.
5 Architecture: a 3D Swin-Conv U-Net (hybrid convolution + shifted-window transformer) to combine local texture sharpening with non-local context modeling, operating on standardized 64×64×64 voxel sub-volumes at 0.5 Å sampling.
6 Supervision strategy: “ground truth” is a forward-simulated EM density generated from deposited atomic models; training uses paired normalization and a loss designed to prioritize ligand regions (including a 30× penalty inside the ligand mask) plus ligand-focused concordance and shape/continuity terms.
7 Quantitative evaluation uses Q-score atom resolvability on an independent test set (n=649). Overall distributions change modestly, but CryoLigATE shows clear gains where it matters most: for poor ligand sites (ligand-only Q-score <0.5; n=81), predicted maps significantly improve, with ΔQ typically positive (up to ~0.3).
8 The method is designed to avoid harming good maps: for cases with already well-resolved ligands (Q-score >0.5; n=558), there is no significant reduction after processing, suggesting CryoLigATE generally preserves high-quality experimental density.
9 Case studies across ligand classes show improved interpretability: NAD+ in MapSPARTA (0.19→0.41), sialic acid in influenza hemagglutinin (0.44→0.65), allopregnanolone in GABAA receptor (0.34→0.56), siponimod in S1P receptor complex (0.22→0.46), and a tethered lipid in respiratory complex I (0.38→0.56), often clarifying functional groups and restoring density continuity.
10 Robustness tests indicate the model can recover expected density even with perturbed ligand coordinates (rotation/translation or even substituting the ligand model), but a key caveat is demonstrated: if directed to an empty region with no experimental signal, CryoLigATE can produce amorphous “hallucinated” density—so it should refine putative ligand density, not invent binding where none is visible.
💻Code: https://t.co/OeHuzWFhwP
📜Paper: https://t.co/TpY8jPC5L8
#cryoEM #DeepLearning #StructuralBiology #DrugDiscovery #ProteinLigand #ComputationalBiology #Bioinformatics #CryoEMMaps