iiinova — where your digital self keeps evolving

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 11/09/2026

I left academia. I spent nearly the last decade doing diverse omics research at UCBerkeley and Vanderbilt. Both are great places for training. However, the world is changing and I could feel it was time to move on. Looking forward to exciting new adventures

4120

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 10/07/2026

PhyKIT v2.2.0 is out! PhyKIT has been rigorously profiled and optimized across the entire codebase, resulting in speedups ranging from ~1.0x to 16.6x, with a geometric mean of 2.0x Python 3.10 through 3.13 is also supported

143

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 08/06/2026

NEW results on the #geometry of "knowing" in #LLMs Where does an LLM's "I'm just guessing" live? Its "I'm fabricating"? Its "I'm deriving step by step"? The answer is in distinct areas of #activation space -- and this shared geometry of #epistemic states is observed across diverse #OS models.

141

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 26/05/2026

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

120

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 26/05/2026

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

sparkit.science

SPARKIT — A scientific research agent for any workflow

One API call. Hours of research, delivered in two minutes. Papers read, analyses run, claims cited.

110

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 23/05/2026

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!

030

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 22/05/2026

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...

000

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 20/05/2026

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...

011

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 20/05/2026

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:...

000

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 19/05/2026

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

030

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 19/05/2026

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

030

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 13/05/2026

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

000

Reposted by 🧬Jacob L Steenwyk

Em Stevens @emstevens93.bsky.social · 12/05/2026

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...

keele.ac.uk

FacultystudentshipinSchoolofLifeSciences

02024

Reposted by 🧬Jacob L Steenwyk

Maureen Barr @barrlab.bsky.social · 12/05/2026

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...

linkedin.com

INSPIRE is recruiting! Please share broadly with senior graduate students and beginning postdocs. INSPIRE, the NIH-funded Rutgers-IRACDA, offers postdoctoral training for individuals interested in… | ...

11533

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 12/05/2026

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...

010

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 12/05/2026

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

100

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 12/05/2026

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

120

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 12/05/2026

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

000

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 11/05/2026

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

120

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 14/04/2026

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. www.biorxiv.org/content/10.1101/202…

010

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 12/04/2026

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.

110

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 11/04/2026

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. www.biorxiv.org/content/10.1101/202…

021

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 10/04/2026

@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

000

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 10/04/2026

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...

020

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 10/04/2026

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. www.biorxiv.org/content/10.1101/202…

061

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 10/04/2026

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...

020

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 09/04/2026

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...

000

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 09/04/2026

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...

011

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 09/04/2026

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. www.biorxiv.org/content/10.1101/202…

1175

Reposted by 🧬Jacob L Steenwyk

Dr. Lauren E. Fuess @laurenfuess.bsky.social · 08/04/2026

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.

11017

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 08/04/2026

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

043

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 08/04/2026

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

000

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 08/04/2026

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)

131

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 08/04/2026

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?

162

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 08/04/2026

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. www.biorxiv.org/content/10.1101/202…

094

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 07/04/2026

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

030

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 07/04/2026

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. www.biorxiv.org/content/10.1101/202…

010

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 06/04/2026

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. www.biorxiv.org/content/10.1101/202…

052

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 05/04/2026

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

142

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 04/04/2026

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. www.biorxiv.org/content/10.1101/202…

182

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 31/03/2026

Plasmid classification has long been constrained by host range. www.biorxiv.org/content/10.1101/202…

100

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 30/03/2026

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

010

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 29/03/2026

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.

221

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 29/03/2026

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

5162

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 29/03/2026

2026 is gonna be a different kind of year

120

🧬Jacob L Steenwyk @jlsteenwyk.bsky.social · 28/03/2026

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. www.biorxiv.org/content/10.1101/202…

0102