NEW preprint: Sparse autoencoders were trained on ESM-2 (650M, seq-only) and ESM-3 (1.4B, multimodal) to ask a simple question -- when two protein language models look at the same protein, do they "see" the same biology? Answer: Yes, mostly t.co/FlSSMH4KAh

Nice to see the number of users for SPARKIT (sparkit.science) increasing There have been a wide variety of use cases, such as achieving a deeper understanding of the literature and brainstorming use cases for new methods

Had a truly wonderful project-based stint at the interface of ML and Biology with Sift Biosciences So impressed by the team, including Yue Clare Lou, Maddie Williams, and Usman Rashid, and definitely look forward to watching them grow!

Someone asked SPARKIT about a specific figure panel in a publication. SPARKIT found the paper and answered correctly Well harnessed agents can do so much, from understanding to more! bsky.app/profile/jlst...

SPARKIT is a research agent with SOTA performance at answering difficult scientific questions (HLE-gold; left) & multi-reasoning and search (GAIA; right) SPARKIT is useful for uncovering relevant literature, brainstorming new ideas & agentic automations 🔗: sparkit.science bsky.app/profile/jlst...

Founding memberships are currently 30% off with the code FOUNDING30. Currently working on a database of 118+ million documents (preprints, open source pubs, etc.) that the SPARKIT agent will be able to search across sparkit.science bsky.app/profile/did:...

Indexed 118,991,763 unique documents to better equip agents with search abilities More coming!

Building out some science-focused search functions to help enable agents to identify the relevant scientific context. More soon!

SPARKIT just works! It is awesome to have a research agent handy and help out with automating some of the most challenging knowledge work sparkit.science

Starting to build a general search agent. Will be sharing that one soon, too, if it continues to perform well on benchmarks!

Just closed an issue for someone who said they use ggpubfigs nearly every day Incredibly grateful to have created a tool with that kind of value for one person github.com/JLSteenwyk/g...

I'm currently advertising a PhD position in my lab, to study the transition from commensal to pathogen in Staphylococcus aureus. Details available here, please share! www.keele.ac.uk/research/pos...

POSTDOC OPPORTUNITY!!! The Rutgers IRACDA program is phenomenal: 70% of postdoctoral fellows land faculty positions. Martha Soto (LinkedIn below) is the amazing director and leader of this program. Please apply or reach out if you have questions. www.linkedin.com/posts/martha...

Happy to see some folks sign up! I'm very open to feedback. Similar to other tools I've built, I want to make sure that SPARKIT meets community needs so any type of feedback, big or small, is taken seriously :) bsky.app/profile/jlst...

30% off for Pro plans founding members A portion of the proceeds will be going toward global conservation efforts bsky.app/profile/jlst...

SPARKIT, a literature agent I built, has SOTA performance at answering difficult scientific questions (HLE-gold) & general multi-step reasoning & search (GAIA) sparkit.science SPARKIT often reads 20-40 papers in ~2 minutes & provides deep reasoning for your query

"Holy moly — major artifact detected." - Claude Opus 4.7

The literature agent, SPARKIT, is available here! sparkit.science bsky.app/profile/did:...

I thought this was a cool example of SPARKIT accurately searching through the internet to answer this tough question from GAIA sparkit.science/blog/gaia-ha... bsky.app/profile/jlst...

Introducing SPARKIT, a literature agent with SOTA performance SPARKIT has improved performance compared to frontier models for difficult scientific questions (HLE-gold; left panel) and general search and multi-step reasoning (GAIA; right panel) sparkit.science

I recently built a literature research agent that I will make publicly available. Using it will cost $; however, any profits that are made from users will be donated I'm really excited about this project - stay tuned!

Neuroplex combines miniscope Ca²⁺ imaging with confocal spectral unmixing to track nine projection-defined neuron types through one GRIN lens in freely moving mice—past the two-fluorophore wall that has bottlenecked circuit dissection. https://www.biorxiv.org/content/10.1101/2023.12.13.571122

This week in preprints: 6 papers on everything from meiotic drive in flies to the origins of eukaryotes. Some real classics alongside newer work this week.

From Pisani et al: CAT-GTR is one of the most flexible models in the phylogenomic arsenal. Phylogenomic mixture models outperform homogeneous and partitioned models academic.oup.com/mbe/advance-...

sOrganellomics uses organelle structure rather than gene expression to map cell states in tissue. In liver, hepatocytes form intermixed communities, not classic graded zones, a pattern that breaks down under nutritional stress. https://www.biorxiv.org/content/10.1101/2024.12.06.627285

@caseynewton.bsky.social Do you think that password management companies should also be allowed access to Mythos preview? Concerns about companies like these getting hacked seem very real

ClustKIT can now make figures of the results. Left: a histogram of cluster size. Right: cumulative sequence coverage, which depicts the percentage of all sequences that are in clusters of at least that size bsky.app/profile/jlst...

Coalescent analysis of T2T genomes finds a shared effective population size peak ~3-6 Mya in humans, chimps, and bonobos, but not gorillas or orangutans, suggesting complex, structured speciation at the Homo-Pan split. https://www.biorxiv.org/content/10.1101/2024.10.17.618932

These parameter sensitivity analyses informed the ClustKIT algorithm. Here, six different parameters are explored and the parameter with the best tradeoff (w/ respect to time running) is highlighted with a dashed line bsky.app/profile/jlst...

getting good feedback on things to watch out for and additional tests to run Many thanks to friends and others who provided feedback! bsky.app/profile/jlst...

Played around with adding fast, balanced, and accurate versions of ClustKIT. They perform pretty similarly (left panel). Going to be conducting a more comprehensive exploration of parameters like changing the sketch size, clustering method, lsh sensitivity, etc bsky.app/profile/jlst...

A foundation model trained on 36M cells across 8 species creates a universal embedding space for cell types—no fine-tuning needed for new data. It recovers developmental lineages and cell functions it was never explicitly trained on. https://www.biorxiv.org/content/10.1101/2023.11.28.568918

Update: I am definitely hiring 2 PhD students; preference to those who can start spring 2027, but reach out if you’re looking to start in Fall 27. Want to start earlier? Reach out and we can talk about technician positions starting as early as this summer.

Welp, didn't realize transparent backgrounds became black when I posted the logo 😅 bsky.app/profile/jlst...

More specifically, ClustKIT has a 41.8x speedup at 192 CPU threads and a 2.7x speedup with dual GPUs. bsky.app/profile/jlst...

NEW tool: introducing ClustKIT, an accurate #protein sequence clustering ClustKIT performs particularly well in the twilight zone of #sequence identity (L panel; low identity thresholds; Pfam benchmark w/ 22,343 sequences from 56 families) ClustKIT also scales well (R panel)

I built an agent (XXYYZZ) that excels at answering scientific questions. XXYYZZ was benchmarked against HLE-gold (149 human-verified questions from HLE). Totally at a loss for what to do with it. Any recommendations?

Eukaryotes didn't just arise from two partners. Phylogenomics of the LECA proteome reveals multiple waves of bacterial HGT, some predating mitochondrial endosymbiosis, plus a possible role for giant viruses. https://www.biorxiv.org/content/10.1101/2024.10.14.618062

Curious if Ant saw me (& probably others) make an ascii desktop buddy Fun feature to include in Claude Code! github.com/JLSteenwyk/a...

DeepTrio calls variants in parent-child trios using joint sequence data, learning inheritance patterns without explicit priors. At 20x coverage it matches DeepVariant at 30x. Models available for Illumina WGS, exome, and PacBio HiFi. https://www.biorxiv.org/content/10.1101/2021.04.05.438434

In Drosophila, a heterochromatin protein (HP2) must transcribe AAGAG satellite DNA before sperm packaging. Without it, Y-bearing spermatids die more than X-bearing ones, producing sex-ratio meiotic drive. https://www.biorxiv.org/content/10.1101/2025.01.10.632443

This week in preprints: 2 papers on how genes and genomes move around in ways we're still just beginning to map.

In Rhizaria, 8-20% of protein-coding genes arrived via lateral transfer. Eukaryote-to-eukaryote transfers outnumber prokaryote-derived ones, and they're concentrated in nuclear/informational processes rather than the cell surface. https://www.biorxiv.org/content/10.1101/2023.01.27.525846

Plasmid classification has long been constrained by host range. https://www.biorxiv.org/content/10.1101/2024.09.03.610885

Surprised and grateful to see the Genomely blog as #48 on the 'Rising in Science' category Subscription is free :) genomely.substack.com

This week in preprints: 4 papers on bats that refuse to age, a heart disease mechanism with a surprising fix, autism gene therapy in organoids, and ancient viral DNA hiding in human centromeres.

Funny how I don't really want to use em dashes anymore, even though I picked up their use pre-AI after reading Beth Shapiro's book, Life As We Made It, where she used them a lot lol I would guess that others are purposefully modifying their writing, or leaving in quarks, to present as human-made

2026 is gonna be a different kind of year

Human centromeres contain up to 140kb of entomopoxvirus-like sequence, totaling 2.42Mb genome-wide. These fragments cluster almost exclusively in hsat1A pericentromeric regions and are actively transcribed. https://www.biorxiv.org/content/10.1101/2024.11.07.622552

CRISPR activation targeting gene enhancers rescues haploinsufficiency phenotypes for CHD8 and SCN2A in autism organoid models, restoring expression and neuronal physiology without editing the mutant allele. https://www.biorxiv.org/content/10.1101/2024.03.13.584921