Excited to share our latest work in Nature. Applying single-molecule and single-cell DNA sequencing methods, we uncover an extraordinary landscape of somatic mutations in immune checkpoint genes in autoimmune B cells, suggesting that somatic mutations may be key to autoimmunity [1/n] rdcu.be/fdqbr
Results from an @sfiscience.bsky.social postdoc "research jam" now published at @prxlife.bsky.social
NEE Focus issue on "Evolution in medicine", including a comment article from @alisonfeder.bsky.social and me.
Can we simulate realistic evolutionary trajectories and “replay the tape of life”? In this work, we propose a flexible, generalizable deep learning framework for modeling how the entire protein sequence evolves over time while capturing complex interactions across sites. 1/n
doi.org/10.64898/202...
How much better is an ancestral recombination graph (ARG) than a site frequency spectrum (SFS)? For recovering mutation rate history, we can answer fairly precisely because both ARG and SFS are linear transforms of mutation rate history. This blog post uses spectral analysis to clarify the picture.
This isn't in international media yet but Lithuania is in crisis. The ruling coalition is trying to seize control of our public broadcaster LRT using expedited parliamentary procedures. European & Venice Commissions are aware & we've been protesting for the last two weeks. Things are looking grim.
Over the past 5+ years I've had the honor of working with @wsdewitt.github.io @victora.bsky.social and many others on a project to "replay" affinity maturation evolution from a fixed starting point.
matsen.group/general/2025...
The last five months with Claude Code have completely changed how we work.
matsen.group/agentic.html details:
• How agents work (& why it matters)
• Git Flow with agents
• Using agents for science
• The human-agent interface
Questions? What has your experience been?
Check out our latest preprint on the effects of antibody-mediated feedback on ongoing germinal center reactions, led by Alex Barbulescu and @janabilanovic.bsky.social
www.biorxiv.org/content/10.1...
excited that this paper is finally out in @pnas.org :
www.pnas.org/doi/10.1073/...
Led by Gian Marco Visani (effort initiated by Michael Pun), fantastic collaboration with @pgtimmune.bsky.social @asya-minervina.bsky.social and Phil Bradley.
In another display of incredible resilience, Ukrainian mathematicians in 2022(❗) opened a new International Centre for Mathematics in Ukraine (ICMU): icmu.ua/en
It was pleasure to give an online mini-course on Bayesian Statistics to Ukrainian students and scientists: icmu.ua/en/events/in...
We are excited to share GPN-Star, a cost-effective, biologically grounded genomic language modeling framework that achieves state-of-the-art performance across a wide range of variant effect prediction tasks relevant to human genetics.
www.biorxiv.org/content/10.1...
(1/n)
My MSc student Indrė Blagnytė has a #preprint up on @biorxivpreprint.bsky.social on influenza D virus: www.biorxiv.org/content/10.1.... Flu D was discovered back in 2011, mostly circulating in cattle. Despite lots of research, a comprehensive analysis of its phylogeography had been missing. 1/6
A short blog post about a recent preprint (work hatched as a very fun all-postdoc collaboration at @sfiscience.bsky.social)
Why does selection feel so weak relative to mutation in affinity maturation? A new blog post giving three perspectives, including our new transformer-based model of natural selection on antibodies: matsen.group/general/202...
And @vmminin.bsky.social! Somehow I couldn't find you before.
Ultimately, direct comparison between the approaches is hard since their three models are so different, but that's what makes them so complementary. With @matsen.bsky.social @yun-s-song.bsky.social @wsdewitt.github.io, and others I can't find on here, but may have missed!
effective vs intrinsic birth rates over time for several simulated GCs using final, fitted data parameter values
they infer what we call an "effective" birth rate (left column) that is more biologically interpretable than the "intrinsic" rate inferred by deep learning (center column), but which also varies with time and across GCs.
It also turns out there's some subtleties in comparing to the more analytic traveling wave and branching process approaches:
Data results: affinity fitness response curves
The results consist of a curve, or rather, two versions of (hopefully) the same curve: one infers the parameters of a sigmoid shape, the other infers independent bin values.
diagram of inference procedure
Finally, we applied the model to real data, inferring the affinity-fitness response curves for many potential parameter values, choosing the best combination based on summary statistic matching.
diagram of simulation training workflow
We then trained a deep learning model on simulation samples with a wide variety of parameter values.
diagram of simulator workflow
Here I'll focus on the deep learning approach. We first built a birth-death-mutation simulator and carefully matched it to data to ensure we understood the processes underlying the experimental results.
whereas a branching process model arxiv.org/abs/2508.09519 and deep learning (this thread) zoom in to use detailed lineage trees to inform inference.
A traveling wave model takes a very high level view, essentially modeling histograms of affinity (Fig 6 in doi.org/10.1101/2025...)
Higher affinity antibodies, on average, have more offspring. But what exactly does this relationship look like? We used three complementary approaches to measure it:
In the weeks after we're exposed to a pathogen, our antibodies evolve toward higher affinity. The cellular mechanisms here are fairly well understood, but a recent experiment from @victora.bsky.social gave us the opportunity to also learn about the mathematical dynamics of this selection.
