We have started a project trying to predic the interactions/structures of all yeast protein pairs using an AlphaFold pooling approach. We are making the current dataset open and we welcome collaborations.
www.evocellnet.com/2026/03/mapp...
Predicting protein-protein interactions (PPIs) at proteome scale can take months with co-folding models due to the massive all-vs-all comparisons required.
We are excited to announce FlashPPI, a contrastive learning framework that predicts proteome wide physical interfaces in minutes. 1/🧵
Slava Ukraini.
Musk on Twitter 2/17/2026: You can just take a picture of your medical data or upload the file to get a second opinion from Grok
Elon posted this yesterday. Just for fun I did it with my before/after cancer PET scans that I had already posted online. In the "before" scan Grok missed the massive tumor lighting up my liver, and in both scans it flagged a non-existent "area of concern" in my chest. Other than that, works great!
If you want to implement an AI tutor, you need to study it’s performance in naturalistic settings with the kinds of context windows that student learners will provide. You can’t begin with a well formulated question about a single topic.
Students don’t know how to formulate questions well at first.
8 Can we live on renewables? The red stack in figure 18.1 adds up to 195 kWh per day per person. The green stack adds up to about 180 kWh/d/p. A close race! But please remember: in calculating our production stack we threw all economic, social, and environmental constraints to the wind. Also, some of our green contributors are probably incompatible with each other: our photovoltaic panels and hot-water panels would clash with each other on roofs; and our solar photovoltaic farms using 5% of the country might compete with the energy crops with which we covered 75% of the country. If we were to lose just one of our bigger green contributors – for example, if we decided that deep offshore wind is not an option, or that panelling 5% of the country with photovoltaics at a cost of £200 000 per person is not on – then the production stack would no longer match the consumption stack. Furthermore, even if our red consumption stack were lower than our green production stack, it would not necessarily mean our energy sums are adding up. You can’t power a TV with cat food, nor can you feed a cat from a wind turbine. Energy exists in different forms – chemical, electrical, kinetic, and heat, for example. For a sustainable energy plan to add up, we need both the forms and amounts of energy consumption and production to match up. Converting energy from one form to another – from chemical to electrical, as at a fossil-fuel power station, or from electrical to chemical, as in a factory making hydrogen from water – usually involves substantial losses of useful energy. We will come back to this important detail in Chapter 27, which will describe some energy plans that do add up. Here we’ll reflect on our estimates of consumption and production, compare them with official averages and with other people’s estimates, and discuss how much power renewables could plausibly deliver in a country like Britain. The questions we’ll address in this chapter are: Is the size of the red st…
@drkatemarvel.bsky.social Are there any good current equivalents of the red stack from David MacKay 2009? It lives rent-free in my head but would like to know if it still reflects current knowledge. www.withouthotair.com/c18/page_103...
This may be the most important paper ever published about NIH funded research.
www.science.org/doi/10.1126/...
The science is, in fact, clear. Good studies [1] consistently show a weak association between tylenol & autism. Those studies are very careful to say they can't establish whether this is causal. A huge Swedish study used a clever sibling design to address this, and showed zero causal effect [2].
Unexpectedly, @jurgjn.bsky.social found that running Alphafold3 predictions for protein interactions can yield ipTM scores that are more predictive of true interactions when run in pools of proteins instead of pairwise predictions. Presumably, this reflects some sort of "competition effect".
The war on science in the US is already having an effect on private sector research like AlphaFold. Bears repeating but the private sector builds on top of things created by academic research for the public good. This hurts everyone.
Examples of alternate conformations with environmental explanations. (A) a protein in two states (open and closed) mediated by allosteric effects; (B) a population of two discrete rotamer states of a sidechain; (C) bound and unbound states due to ligand interaction; (D) different oxidation states of a cysteine pair. Furthermore, some regions in many protein structures exhibit such a degree of flexibility, that no accurate modelling is possible (dashed lines, E). The backbone conformation cannot be determined with crystallographic methods in such cases.
Failure of ensemble prediction algorithms in altloc segments. Although structure and ensemble prediction techniques based on amino acid sequences are capable of accurately reconstructing the overall backbone structure (left column), they consistently fail to produce viable samples corresponding to one of the two altloc conformations (as shown in the magnified regions in the second through fourth columns).
Some crystal structures show evidence of multiple stable conformations, such as loops or side chains with two distinct regions of density. These can be correctly modeled by MD, but not modern protein ensemble prediction tools like AlphaFlow or DiG (BioEmu not tested)
www.biorxiv.org/content/10.1...
I miss when the biggest controversy about mRNA was whether it correlated with protein or not.
LLMs can't take responsibility for their mistakes. When a human journalist puts their name on AI-written text, they take on that responsibility.
Increasingly I see inaccurate and badly written news stories authored by AI, many of which have actual humans listed as authors or editors.
The best first sentence of a grant application I've read was (paraphrasing), "Tool X is widely used to do task Y; we will make it accessible to people living with condition Z (13% of the population) so that it is more equitable and more widely used". Let's unpack why, because there's a lesson. 🧵
We’ve now explored this on all GPU models available on our cluster and some non-A100/H100 do not have the issue; seems like something nuanced that can hopefully be fixed:
github.com/google-deepm...
We’ve now explored this on all GPU models available on our cluster and some non-A100/H100 do not have the issue; seems like something nuanced that can hopefully be fixed:
github.com/google-deepm...
Jurgen in our group run AF3 locally after fiddling. It worked well on a A100 and produced some bad models on a lower end GPU. We don't have access to many A100s so unless we get it working on lower-end GPUs we won't be able to use this much.
The former finished without errors, but the output was noise (something similar to 100% spaghetti with AlphaFold2). The latter finished with the structure posted by @pedrobeltrao.bsky.social (2/2)
We tried running the example from README.md (“2PV7”) on a lower-end GPU with --flash_attention_implementation=xla (described in performance.md), and on an A100 GPU without that option. (1/2)
1. I'm conflicted about moving old tweets to Bluesky. I like the idea in general of using @blueark.app to preserve that content.
I have a few threads posted during COVID—for example, the original analysis of the IHME model right after it came out—that seem a valuable part of the scientific record.
