We are releasing Carbon: a crazy fast DNA model Carbon is 275x faster than the next best model. So fast you can process the whole human genome on a single GPU in <2 days. Here are the tricks we used: When modelling DNA sequences a lot of the performance comes down to tokenizing the sequences in a smart way. BPE tokenizer struggle because there are no whitespaces and character (called base in DNA) level tokenizers waste a lot of compute on too many tokens. Carbon is built with a unique tokenizer: we split sequences in chunks of 6 bases, but during both training and inference we can work with single base resolution. That's similar to having word tokens but resolving them at the character level. All possible thanks to the DNA tokens unique structure. The architecture combined with the tokenizer makes the model 275x faster than the previous SoTA (Evo2) at this size. We built an interactive demo so you can explore how the model can generate DNA sequences, investigate the structure of genes, predict the effect of mutations, generate and fold proteins and even reconstruct parts of the tree of life. huggingface.co/spaces/Huggin…

I find this funny-sad. On July 25 GigaRef was introduced as the largest protein seq DB to date with 3.34B seqs. But on Sep 1, the LOGAN DB was released with >100B seqs. This means the title of the largest seq DB only lasted ~1month. Meanwhile the DeepClust DB was never released☠️

A new paper in @Nature from David Reich, @aliakbari23 and colleagues breaks the conventional understanding of recent human evolution. The field believed that strong selection in the recent past (~10,000 years) was rare, with few exceptions like the lactase persistence locus. In this paper, the authors challenge that belief, showing that we weren't looking at the problem right. Previous studies that looked for evidence of selection using ancient DNA addressed the problem cross-sectionally, asking if allele frequencies differed across populations more than what one would expect based on genetic drift and migration. Most arrived at the conclusion that population structure primarily explained the observed differences. Here, the authors addressed the problem longitudinally, accounting for when ancient individuals lived by explicitly modeling time as a variable in the analysis. It turns out doing it this way dramatically increases power, increasing the number of genome-wide significant selection signals by 20-fold! Looking at why accounting for the time variable led to such dramatic changes in results, the authors find that previous studies missed so much because selection often happened not on new variants leading to dramatic sweeps (the conventional model: new variant -> selection -> increase in frequency) but on already existing variants driven by transient environmental pressures. Many of these variants underwent reversals, selected up when a pressure existed, then purged when it disappeared or the trade-off cost became dominant. A great example is the TYK2 variant, where an allele boosting immunity was selected for thousands of years because it protected against TB, then got purged as TB endemicity declined and the autoimmune cost took over. The scale of what they found is striking: hundreds of loci showing strong selection in the past 10,000 years with a median selection coefficient of ~0.86%. This number is pretty big in evolutionary terms, meaning allele frequencies have been shifting by ~1% per generation in a consistent direction. Previous selection scans found a maximum of 20 loci, and this one finds hundreds. That isn't an incremental change. It fundamentally reframes our understanding of how common strong selection has been in recent human history. Some of the most striking findings come from polygenic selection, where hundreds of small-effect alleles were pushed in the same direction simultaneously. Polygenic scores based on large-scale GWAS of today predict recent negative selection for traits like body fat, waist circumference and schizophrenia, and positive selection for others like cognitive traits. One important caveat is that GWAS phenotypes are measured in industrialized societies today, and how well they capture what was actually being selected in ancient environments is debatable. For me personally, these findings have direct implications for drug discovery. When using human genetics to find drug targets, we often fixate on the benefit and risk profiles of variants visible today. But we need to be aware that a variant's benefit:harm ratio might be environmentally contingent, and could reverse when the wrong environment manifests. An evolutionary understanding of a variant's association with traits is therefore essential. The same logic applies, perhaps even more urgently, to embryo selection. Selecting embryos based on polygenic traits is humans making permanent, heritable decisions for their offspring with a narrow view of today's environment. The ancient DNA record now shows that cost-benefit landscapes flip over time. So, an embryo carrying man-made selections is carrying those changes into an unpredictable future environment. The broader takeaway is that human evolution didn't freeze in the last 10,000 years. We just lacked the tools and datasets to see its movement. The current findings are based on European populations. I am curious to see these analyses extended to other populations too, like South Asian, East Asian and African populations, which might be holding more surprises to blow our minds. Akbari et al. Nature 2026 nature.com/articles/s41586-0…

Just because epigenetic drift follows a predictable pattern doesn't necessarily mean it's causal in aging. Predictable patterns can form from stochastic reactions. Also you cannot fix all extracellular damages with cellular rejuvenation. More here: olafurpall.substack.com/p/wh…

High-resolution metagenome assembly for modern long reads with myloasm - @lh3lh3 go.nature.com/3PBEwvR