Google has specifically prevented its AI tool from telling users who won the 2020 presidential election, even though it happily tells us that Trump won in 2016 and 2024. This sort of political intervention into LLM responses is deeply problematic. (h/t to Charles Mann)

Two stages of research: 1) I think I’m onto something 2) oh no

This is super cool result for AI & protein design! Claude wrote CUDA kernels that beat NVIDIA's own cuEquivariance on triangle attention by 2.7–2.9x, and triangle multiplication by 1.7–3.2x. Two engineers supervised it. Neither had prior kernel engineering experience. 30+ models such as AlphaFold3, Boltz-2, OpenFold3, protein LMs, genomics, done in 4 weeks. Triangle ops are the cubic bottleneck in Pairformer-based structure prediction. Fix them and the whole stack gets cheaper: 4x average speedup, 100x fewer GPU hours for comparable protein design results. FlashPairformer is now open-sourced: anthropic.com/research/claud…

Mathematics has now fully morphed into biology: Project costs millions $ ✅ Put out paper that nobody has read ✅ Massive paper supplement nobody will read ✅ Advertise with pretty art ✅ Vicious authorship fight ✅ Say you're curing cancer ✅

Pete Hegseth: We need our military to be more like the Spartans! *immediately loses fight with Persia over an important chokepoint*

Generative models can’t discover what they can’t reach. We’re excited to introduce ActFlow: a continued pre-training scheme that actively expands the valid design space reachable by flow and diffusion models. We call this generable set expansion — a new learning principle for out-of-distribution generative modeling, and a step toward evolvable search spaces for scientific discovery. (1/5)

UMAP makes useful pictures, but what structure does it destroy? Daan H. de Groot and coauthors introduce Bonsai, a Bayesian method that represents high-dimensional data as a tree rather than forcing it into a conventional two-dimensional embedding. The motivation comes from single-cell biology. Methods such as UMAP and t-SNE are extraordinarily useful for visual exploration, but their geometry is not a faithful representation of the original high-dimensional space. Local and global relationships can be distorted, and changing visualization parameters can change the apparent biological story. Bonsai takes a different approach. Each cell is treated as a noisy point in high-dimensional gene-expression space, complete with uncertainty estimates. The method then reconstructs the maximum-likelihood tree connecting those points under a probabilistic model of how expression states can change. Because a tree can always be drawn in two dimensions, visualization no longer requires projecting the data onto two artificial coordinates. On synthetic benchmarks, Bonsai almost perfectly reconstructs structures that PCA and UMAP largely fail to recover. It also preserves cell-to-cell distances much more faithfully. And the biological payoff is not only prettier visualization. Applied to cord-blood single-cell data, Bonsai recovers established differentiation relationships and identifies an apparently previously undescribed subset of natural-killer cells associated with the myeloid rather than lymphoid lineage. The method also scales through a backbone implementation to datasets exceeding one million cells. There is a broader scientific lesson here. Dimensionality reduction is not a neutral preprocessing step. When ML changes the representation of scientific data, it can also change which hypotheses become visually plausible. Paper: de Groot et al., Nature Biotechnology (2026), CC BY 4.0 | DOI: 10.1038/s41587-026-03220-2

HASHTAG JUSTICE FOR THE ARMATIMONADETES

Genuinely feel like I should start a blog on all the cool shit i see in random genomes that I can't be bothered to ever publish on. Somebody needs to be out here advocating for the Armatimonadetes

Who is out here trying to use protein design to make novel active sites and then go back and find matching ones in nature? Anybody?

I’m really excited to share our new preprint, out today, where we built a fully autonomous system for enzyme engineering — integrating a self-driving lab with a generative protein language model that worked together over a month to engineer substrate specificity in glycoside hydrolase (GH1) enzymes. biorxiv.org/content/10.64898… [1/n]

Silico can also improve drug design and lots of other AI x bio work via interpretability! Last week, I found that by reading the internal representation of protein-folding model, we can better rank and generate ligand-binding poses, up to 4x! 🧵 1/

Do we even need to fold proteins at all to find drugs? Or do we just do that to satisfy our 3D brains? Meet Ptarmigan-1 🧵 - Structure-free drug discovery - 5000x less compute than co-folding - Finds known and cryptic pockets - Actually works on poorly structured targets

metaSMASH: Scalable Biosynthetic Gene Cluster Detection for Large Metagenomic Assemblies biorxiv.org/content/10.64898…

A way for AI to design proteins better than natural ones nature.com/articles/s41586-0… @davidrliu @broadinstitute @NicholasKrasnow

De Novo Design of Ligand Binding and Sensing With a Physics-Based Generative Approach 1. The study introduces a physics-based generative framework that designs ligand-binding proteins directly around target molecules while simultaneously optimizing protein folding. Rather than searching existing scaffold libraries, it builds customized helical architectures with binding pockets tailored to small molecules or metal ions. 2. Its central innovation is the joint optimization of scaffold topology, internal protein stability, and protein–ligand interactions. This enables the authors to design not only static binders, but also proteins whose ligand recognition drives fragment reassociation or global protein folding—properties that are directly useful for biosensor construction. 3. The pipeline combines Rotamer Interaction Fields with Monte Carlo fragment assembly. Favorable ligand-contacting side chains are first enumerated as spatial anchors, after which protein backbones are generated around those anchors using ligand-interaction scores and residue-pair packing scores to guide folding. 4. For metal ions, the framework explicitly encodes coordination chemistry. It generated tetrahedral zinc-binding sites and octahedral nickel-binding sites by controlling coordinating residue identities, positions, and geometries during backbone construction. 5. The method produced experimentally validated binders for five chemically diverse small molecules: dolutegravir, cianidanol, DFHBI, serotonin, and dopamine. It successfully targeted serotonin and dopamine despite their unusually small size and polarity, which provide relatively few opportunities for favorable protein contacts. 6. Initial screening identified 14 binders among 96 dolutegravir designs, 20 among 94 cianidanol designs, six among 38 serotonin designs, and one among 33 dopamine designs. Subsequent structural or sequence optimization improved the best dolutegravir affinity to 971 nM and the optimized serotonin binder affinity to 29.9 μM. 7. High-resolution crystal structures strongly validated the computational models. The serotonin binder SRO_30, dopamine binder DPA_9, and zinc binder ZK2 matched their designed backbones with Cα RMSDs of approximately 0.97 Å, 0.79 Å, and 0.45 Å, respectively, while the designed ligand poses and metal-coordination geometries were largely reproduced. 8. The zinc binder ZK2 coupled metal recognition to protein folding, binding zinc with a reported Kd of 150 nM while showing no detectable binding to copper, cobalt, or nickel. This selectivity is notable because copper would ordinarily be thermodynamically favored over zinc according to the Irving–Williams series. 9. The authors converted the serotonin and dopamine binders into fluorescent sensors by splitting their helical bundles and fusing the fragments to circularly permuted GFP. Repeat helices, interface-weakening mutations, fragment truncation, and linker optimization reduced ligand-independent fragment association and improved signal responsiveness. 10. The optimized serotonin sensor wcSRO30_1.0 reached a ΔF/F₀ of approximately 340% during development and displayed a reported maximal response of 411%, an EC₅₀ of 539.9 μM, strong selectivity over related neurotransmitters, reversibility in E. coli, and measurable responses in HEK293T cells. A dopamine construct produced a 37.5% fluorescence response and was described as the first reported soluble dopamine biosensor. 11. The zinc-induced folding mechanism was also connected to split NanoLuc complementation. Two resulting zinc sensors achieved EC₅₀ values of 72.8 and 78.7 nM, illustrating how designed folding transitions can directly generate sensitive luminescent outputs without requiring a naturally evolved sensing domain. 12. Overall, the work presents ligand-binding protein design as an integrated problem of molecular recognition, scaffold generation, conformational energetics, and signal transduction. The current neurotransmitter sensors still require improvements in affinity and physiological sensitivity, but the framework provides a modular route from a specified ligand to an experimentally validated binder and then to a functional biosensor. 📜Paper: doi.org/10.64898/2026.07.13.… #ProteinDesign #ComputationalBiology #Biosensors #DeNovoDesign #GenerativeBiology #SmallMolecules #MetalBinding #Serotonin #Dopamine #SyntheticBiology

Congratulations to the Michelle Chang Lab, publishing “Targeted enzyme discovery using metal-coordination mining” today in @Nature. Great work from Yanni and Michelle. Full paper: nature.com/articles/s41586-0…

Tensor GNNs for protein–ligand affinity prediction Lead optimization is often a problem of small differences. A fluorine atom, a stereocenter or a single pocket mutation can shift affinity enough to make or break a molecular probe. FEP can capture these effects, but its cost limits use across design cycles. Jie Yu and coauthors present PBCNet2.0, a Siamese neural network for relative protein–ligand affinity prediction. It compares two complexes and predicts the activity change, rather than scoring each molecule in isolation. This matches medicinal chemistry: asking whether a new analog improves over a reference. The ML advance is scale plus geometry. PBCNet2.0 is trained on 8.6 million protein–ligand complex pairs from BindingDB, 14× more than the original PBCNet. It replaces hand-crafted encodings with a Cartesian tensor-based equivariant module, letting the model learn distance and angular constraints from 3D structures. In zero-shot tests on the FEP benchmark, PBCNet2.0 reaches Spearman ρ = 0.67, close to Schrödinger’s FEP+ at ρ = 0.70, while remaining far more efficient. In retrospective prioritization, it improves lead optimization efficiency 7.18-fold and reduces resource use by about 41%. The scientific part matters as much as the benchmark. The model recognizes subtle motifs, including fluorine orthogonal multipolar interactions, where geometry is essential. In an ENPP1 case, replacing fluorine with hydrogen reduced binding strongly, and the experimental ΔpKd closely matched the prediction. Perhaps the most interesting result is emergent mutation reasoning. PBCNet2.0 was not trained on mutation data, yet it predicts affinity changes caused by binding-pocket substitutions. In ALDH1B1, it helped prioritize residues for validation, with five of six selected mutations impairing binding. This is structure-aware ML moving from screening toward decision support in lead optimization. In drug discovery and chemical biology, models like this could help prioritize analogs, anticipate resistance liabilities and reduce synthesis before the next design cycle. Paper: Yu et al., Nature Chemical Biology (2026), journal license | doi.org/10.1038/s41589-026-0…

Protein Engineering Textbook betterenzyme.com

死海の生物はどうやって泳ぐのか oist.jp/ja/news-center/news/… 死海は塩分濃度が高く水の粘度も高いため、微生物が泳ぐのが困難です。 今回、死海にすむ古細菌Haloarcula marismortuiで、泳ぐためのフィラメントの外側に、鞘のような構造があることを発見。 極限環境で泳ぐためのめっちゃ面白い構造です