We’re launching AlphaGenome Atlas: an AI-powered searchable database mapping the predicted impact of all 9 billion possible single-letter DNA changes.
AlphaGenome Atlas is over 30 times larger than the AlphaFold Database.
It gives scientists an intuitive way to explore this vast 1-petabyte dataset - unifying interconnected resources so researchers can link genetic variants directly to the molecular mechanisms they disrupt.

We believe foundational biology tools should be accessible to everyone.
AlphaGenome Atlas resources are available to the global scientific community via the Atlas website, our AlphaGenome API, as a skill in Google Antigravity, and coming to @GoogleCloud soon

A new Nature Communications study introduces MitoMatch, repurposing AlphaFold-Multimer to screen 630,003 protein pairs across the human mitochondrial proteome.

Instead of asking only:
“What does this protein look like?”
AI can increasingly ask:
“Who does it interact with — and what does it do?”

David Baker’s team is pushing AI protein design into its next phase.
AI models can now generate millions of protein designs — but the bottleneck has shifted from creation to validation.

In their latest Nature Communications paper, the Baker Lab develops a scalable experimental characterization platform to rapidly test AI-designed proteins and close the loop between Design → Build → Test → Learn.

The future of protein engineering may not just be bigger models — it may be faster feedback loops between AI and the wet lab.

Google is shutting down AlphaFold. It be interesting to see if that means UofWA will need Rosetta@Home for research again.

https://www.engadget.com/2225849/google-shuts-down-alphafold/

Rosetta Foundry has his new docker image (13.9 Gb)

Version 3.0.4 makes AlphaFold 3 runnable on CPU-only machines and on Apple Silicon. Moreover, there are some performance improvements and small bug fixes. We thank everyone who reported issues, proposed new features, and submitted PRs!
AlphaFold 3 now runs on CPU-only machines! It is slow without a GPU, but you can run it anywhere.

That's interesting. Which language, CUDA? OpenCl??

CUDA only so far. I did get a proof of concept working with the Modular/LibTorch custom operator integration to target AMD and Apple Metal (https://docs.modular.com/max/develop/custom-kernels-pytorch/), but the dependencies are already quite restrictive on compute/driver version just for CUDA on Linux/Windows, and Mojo/Modular excludes even more even newer cards so I don't think we'll end up going that way for this application.

Everything said is far above my level of understanding

I'm scanning through the forum to find some info. First this one from 13 July (Monday)

Currently, I have been working on the MAM1 GPU build by implementing a population-based optimizer from the ensmallen library called "Differential Evolution" to join the existing Simulated Annealing and Particle Swarm Optimization execution modes, and launching multiple folds of cross-validation at once. GPU utilization now provides a decent speedup.

OpenDDE is an open-source, all-atom biomolecular foundation model that turns co-folding into a scalable engine for structure prediction, design, and optimization in drug discovery.

The researchers do not retrain AlphaFold3. Instead, they guide the structure generation process during inference using experimental measurements from NMR, X-ray crystallography, and cryo-EM. During the diffusion process, the generated structures are continuously compared against the experimental measurements. This guidance gradually steers AlphaFold3 toward conformations that better agree with the observed data. Instead of predicting a single protein structure, the framework generates a protein ensemble consisting of multiple conformations that collectively satisfy the experimental measurements.

Rather than replacing AlphaFold3, this work extends it with experimental guidance. It represents an important step toward AI models capable of generating experimentally consistent protein ensembles.

IF they wanted to keep work on the PC's they would. But they determined a machine is better.
Faster throughput and elimination and routine work. Less resources, less hassle and less people.
For what they want to do, PC's are not needed as much.

Sorry to say that, but that is the price of advances in automation.

ProtGPT3, a new open-source family of protein language models, from 112M to 10B parameters.
We're releasing models:
• ProtGPT3-112M, 1.3B, 10B (single-sequence) + the RL-aligned versions to reduce low-complexity generations while preserving diversity.
• ProtGPT3-MSA (112M, homolog-conditioned)

The big finding: a 112M MSA model prompted with as few as 15 homologs outperforms supervised fine-tuning (SFT) of models up to 10B parameters on target family generation — across sequence, structure, AND family-level consistency metrics.

OpenBind was created to address a bottleneck in structure-based AI: data. We believe that the next generation of structure-based machine learning methods requires better experimental data, not just new architectures. To address this, we are generating dense, high-quality protein–ligand datasets that link structures with binding measurements at a scale rarely available in public resources, supporting model training, fine-tuning, benchmarking, and error analysis.

Today, we are releasing the first public dataset from the OpenBind consortium: a structure–affinity dataset containing 925 crystallographic binding events from 699 compounds, and associated affinity measurements for 601 compounds.

We are excited to announce that Dr. Green Ahn will join the University of Washington Department of Biochemistry and the Institute for Protein Design as an Assistant Professor in January 2027. Green is a chemical biologist and protein designer whose work sits at the intersection of molecular engineering, functional genomics, and cancer biology.
As an independent investigator, Green’s lab will push into largely uncharted territory: using designed proteins to develop precise biological agents for uncovering complex cell biology and addressing unmet needs in cancer and immune response.
The Ahn Lab is now accepting applications from postdoctoral candidates, University of Washington graduate students, and research technicians in the Seattle area

Hope that, before or later, the code of Rosetta@Home will be merged with this new release...

Diffusion models are an amazing tool for cofolding, they allow us to predict a protein and the molecule bound to it at once. But they are not exactly fast and require a lot of denoising steps to get accurate predictions.

So we distilled ours. Meet DeCAF-Pearl: the first flow map model for all-atom cofolding.

Instead of inching along the denoising trajectory, a flow map learns to jump across it. DeCAF-Pearl runs structure generation ~5x faster than Pearl, our SOTA model, while still maintaining the performance of the teacher model.

That speed up allows us to run larger experiments and generate more synthetic data to improve our models.

Getting there meant reparameterizing into noise-level space to stabilize gradients, committing to clean-structure prediction to keep the rigid-alignment loss biomolecules needed, and building DeCAF-Search, one steering algorithm for every compute budget.

don’t type pymol commands urself, now there’s co-pymol
available for claude code and cursor