iiinova — where your digital self keeps evolving

Eric Talevich @etalevich.bsky.social · 22/04/2026

Bsky is, believe it or not, still the place to be. The mood has shifted, sure. Scientific X is gone -- remember its algo sinks posts with links, so effectively you can't share articles there.

000

Reposted by Eric Talevich

DNAnexus, Inc. @dnanexus.bsky.social · 20/04/2026

This week: accelerate your drug discovery with AI-driven workflows for protein structure prediction: hubs.ly/Q04cxxTV0 #AI #DrugDiscovery #Pharma #ProteinPrediction #PrecisionMedicine #AlphaFold #Nextflow

Live Webinar: April 24, 9 AM PST | 12 PM EST. Accelerating Protein Structure Prediction and Visualization: Scalable Workflows with Nextflow. Speakers: Darren Ames, Head of Solution Science DNAnexus, and Eric Talevich, VP of Bioinformatics DataXight

131

Eric Talevich @etalevich.bsky.social · 09/04/2026

What explains the difference in conclusions between Scale AI / Secure Bio and Active Site? Different lab flows being assessed, or a hint of incentives and selection bias?

010

Eric Talevich @etalevich.bsky.social · 09/04/2026

I'll give a webinar talk April 24th, covering: - nf-core/proteinfold (AlphaFold2, ESMFold, etc.) - Molstar/Mol* - what comes before and after them This kicks off a series on molecular modeling using open-source tools on DNAnexus.

000

Eric Talevich @etalevich.bsky.social · 09/04/2026

I wrote about how I write code for clinical use, as of 2026: etal.github.io/2026/03/30/v... Briefly: - Claude Code in one terminal, zsh in another - Papers and docs in a browser - Plugins: feature-dev, serena, python-lsp, explanatory-output-style - Lots of deterministic tools.

etal.github.io

Are you really vibe-coding CNVkit?

Clinical considerations for bioinformatics development with AI coding agents.

001

Reposted by Eric Talevich

Hunter Shain @shainlab.bsky.social · 10/06/2025

I have been listening to the Hubermann-JB podcast. It is 4 hours, so it will take some time to finish, but here are some thoughts...

111

Eric Talevich @etalevich.bsky.social · 21/05/2025

Regeneron picks up the 23andMe dataset -- one of the best possible outcomes, I think. REGN has a stellar track record of handling patient data responsibly and effectively, e.g. UK Biobank, Geisinger Health. www.biopharmatrend.com/post/1252-re...

biopharmatrend.com

Regeneron Acquires 23andMe Assets for $256M to Expand Consumer Genomics and Data-Driven Drug Discovery

021

Reposted by Eric Talevich

Andrey Kislyuk @kislyuk.bsky.social · 10/05/2025

www.savensf.com/take-action

savensf.com

Take Action | Save NSF

011

Eric Talevich @etalevich.bsky.social · 10/05/2025

Gut feeling: AAV is probably out aside from what is already at the clinical stage in pipelines; LNP with various payloads is probably the future, and there might be a lull in between while scientific gaps get figured out.

000

Reposted by Eric Talevich

Rohit Singh @rohitsingh8080.bsky.social · 05/04/2025

Bio foundation models are great design and engg tools. But can they help decode the fundamental principles of life? We harnessed a single-cell FM for decoding the long-debated relationship between genome arch. and gene coregulation. 1/ Preprint here: www.biorxiv.org/content/10.1...

biorxiv.org

Tracing the Shared Foundations of Gene Expression and Chromatin Structure

The three-dimensional organization of chromatin into topologically associating domains (TADs) may impact gene regulation by bringing distant genes into contact. However, many questions about TADs' fun...

23111

Eric Talevich @etalevich.bsky.social · 01/04/2025

To be fair, there was some chicanery and sloppy work happening under LDT's lighter scrutiny, per @annaleighclark.bsky.social

000

Eric Talevich @etalevich.bsky.social · 31/03/2025

Whereas, the LDT route allows easier commercial ramp-up from research use only (RUO) to CAP/CLIA-certified lab-developed tests (LDT), requiring technical/analytical validation of the new test on banked samples but not patient recruitment. Faster, less overhead, lower reimbursement. Stepstone to IVD.

000

Eric Talevich @etalevich.bsky.social · 31/03/2025

The rule was seen as unfair by both academic medical centers and biotech startups because it would have essentially required a clinical trial to bring any new test to market. Only the biggest existing players can afford to do that, and it takes a long time to build up to that point.

100

Eric Talevich @etalevich.bsky.social · 31/03/2025

A district court has struck down the FDA's Final LDT rule, which would have required clinical test developers to navigate full FDA IVD approval to bring new products to market -- a much tougher bar to clear. Tl;dr: Diagnostics biz isn't dead. Source: www.courtlistener.com/docket/68802...

courtlistener.com

American Clinical Laboratory Association v. U.S. Food and Drug Administration, 4:24-cv-00479 - CourtListener.com

Docket for American Clinical Laboratory Association v. U.S. Food and Drug Administration, 4:24-cv-00479 — Brought to you by Free Law Project, a non-profit dedicated to creating high quality open legal...

220

Eric Talevich @etalevich.bsky.social · 20/03/2025

Element Bio says this on-instrument hybrid capture approach can bring total library prep time down from 12-24 hours to as little as 5 hours (before sequencing). That faster TAT should be great for NGS diagnostics.

030

Reposted by Eric Talevich

Galaxy Project @galaxyproject.bsky.social · 17/03/2025

The GBCC2025 Scholarships are here! Thanks to the JXTX Foundation and Galaxy Project, 4 genomics and data science graduate students will receive scholarships to attend GBCC2025 in person. 📅 Apply by April 1, 2025 👉 jxtxfoundation.org/news/2025-2-...

jxtxfoundation.org

JXTX + GBCC 2025 Scholarships

013

Eric Talevich @etalevich.bsky.social · 17/03/2025

Deprioritizing posts with links is probably a deal-breaker for scientists on X. (On top of all the other stuff.)

000

Eric Talevich @etalevich.bsky.social · 17/03/2025

It's critical to 1. choose the right market, and then 2. bring a working product to that market, in the right way, considering pricing and reimbursement. For GT in particular, patients sometimes hold out because you usually only get one shot at a cure.

000

Eric Talevich @etalevich.bsky.social · 17/03/2025

Scientifically, there's some bitter irony there because CRISPR gene therapies / gene editing therapies "ought" to be able to anything a cell therapy can do, but more directly. But they're not ready yet.

100

Eric Talevich @etalevich.bsky.social · 17/03/2025

Cell therapies, on the other hand, might actually be doing OK. For example: www.biopharmadive.com/news/astraze... Maybe the clinical endpoints are more compelling in cancer and autoimmune diseases. Maybe the tech just works better.

biopharmadive.com

AstraZeneca adds ‘in vivo’ cell therapy capabilities with EsoBiotech deal

The pharma is paying up to $1 billion to buy the Belgian startup, whose technology it sees as useful in treating cancer and autoimmune conditions.

100

Eric Talevich @etalevich.bsky.social · 17/03/2025

Roche deems Spark Therapeutics, a flagship gene therapy company, has fizzled with a US$2.4bn write-off. Luxturna didn't sell well, surprisingly. I hadn't thought gene therapies were in a bubble but it seems like a broader correction is happening. www.fiercepharma.com/pharma/roche...

fiercepharma.com

Roche overhauls Spark gene therapy unit, recording $2.4B in full impairment

Roche has recently launched a “fundamental reorganization” of Spark Therapeutics, the gene therapy unit the Swiss pharma bought for $4.3 billion in 2019. | Roche has recently launched a “fundamental r...

110

Reposted by Eric Talevich

Daniel MacArthur @dgmacarthur.bsky.social · 15/03/2025

New preprint! We worked with @msftresearch.bsky.social and @broadinstitute.org to see whether large language models (LLMs) can be useful to variant scientists in deciding whether genetic variants seen in a patient are responsible for their disease. tl;dr yes they can: www.biorxiv.org/content/10.1...

biorxiv.org

Evidence Aggregator: AI reasoning applied to rare disease diagnostics

Retrieving, reviewing, and synthesizing technical information can be time-consuming and challenging, particularly when requiring specialized expertise, as is the case of variant assessment for rare di...

114247

Eric Talevich @etalevich.bsky.social · 13/03/2025

What a great resource! Python developer tooling has matured a lot over the past decade.

011

Eric Talevich @etalevich.bsky.social · 12/03/2025

E.g. say you have an in-house notebook that supplements the NGS analysis you get from laava. You decide during development that you'd like to copy in one of laava's plots and maybe tweak it. With Quarto that could be a more seamless copy-paste, versus Jupyter <> Rmarkdown.

000

Eric Talevich @etalevich.bsky.social · 12/03/2025

The reporting mechanism would also benefit from using Quarto. I have various notebooks to create plots using both Python and R from the same LAAVA output data, and while Jupyter and Rmarkdown can be multilingual, for the purpose of generating and remixing visual reports, Quarto handles it better.

110

Eric Talevich @etalevich.bsky.social · 12/03/2025

Ideally I'd translate the Bash scripts to Python, convert it all into a proper Python package, and refactor. But modern Python packaging handles CLI scripts in a quirky way that I worry might harm the experience for developers who wear other hats most of the time. Still worth doing, but not today.

100

Eric Talevich @etalevich.bsky.social · 12/03/2025

The motivation for making it modular and easy to modify is that some parameters deserve studies to tune properly, particularly the read mapping parameters. Even the choice of minimap2, bowtie2, or maybe megablast is worth evaluating.

100

Eric Talevich @etalevich.bsky.social · 12/03/2025

I focused on the developer experience in this release. The codebase now supports multiple entry points fairly smoothly: portable Nextflow, Form Bio GUI, docker image, a Bash script mirroring each process in the workflow, and a Python quasi-package (with room for improvement).

110

Eric Talevich @etalevich.bsky.social · 12/03/2025

LAAVA 4.0 is out! Get the latest on GitHub: github.com/formbio/laav... Long-read AAV Analysis (LAAVA) is an open-source bioinformatics pipeline that reports on AAV vector genome integrity using a standardized, published nomenclature and read classification scheme. 🧵

github.com

Release v4.0.0 · formbio/laava

What's Changed Functional changes: Handle mITR/wtITR directionality for scAAV classification via new input "mitr_label" (#70) Output aggregate summary tables for reporting (#67) Add version number...

210

Reposted by Eric Talevich

Richard Sever @richardsever.bsky.social · 11/03/2025

Big news: we are setting up a new non-profit organization to run bioRxiv and medRxiv. It's called openRxiv [no it's not a new preprint server; it's dedicated organization to oversee the servers] openrxiv.org 1/n

openrxiv.org

Homepage - openRxiv

openRxiv is an independent non-profit, the new organizational home for bioRxiv and medRxiv, enabling researchers to instantly share groundbreaking findings with the global scientific community.

552558846

Eric Talevich @etalevich.bsky.social · 06/03/2025

Clearly CNVkit is in a league of its own.

011

Eric Talevich @etalevich.bsky.social · 06/03/2025

In the best of all possible worlds, someone will have done an objective third-party benchmarking study on multiple competing methods. I love those papers. E.g. journals.plos.org/ploscompbiol... You get not only the winners, but also a deep dive into design decisions, tradeoffs and relevant factors.

journals.plos.org

Comprehensively benchmarking applications for detecting copy number variation

Author summary As an important type of genomic structural variation, CNVs are associated with complex phenotypes because they change the number of copies of genes in cells, affecting coding sequences ...

100

Eric Talevich @etalevich.bsky.social · 06/03/2025

How I pick a bioinformatics tool to start: - Check nf-core -- is there a pipeline I can use as-is? If not, what tools are used inside other relevant pipelines? - Check the methods of studies similar to what you're trying to do -- not searching methods papers themselves initially, but who used them

121

Reposted by Eric Talevich

Catharine.Aquino @catharineaquino.bsky.social · 15/02/2025

Because a lot of people had problems registering for the @RocheSequencing seminar about their new SBX sequencer , they made it easier to join . Here is the link : streamingmedia.roche.com/id/1_arrcrfhm

streamingmedia.roche.com

Introducing Sequencing By Expansion (SBX) - 20 Feb 2025 - Webcast&Video Portal

132

Eric Talevich @etalevich.bsky.social · 09/02/2025

There is some ancient wisdom about never coding at the limit of your own cleverness, because modifying that code will require you to be twice as clever. AI doesn't care. You can go from idea to usable app to unmaintainable legacy codebase in minutes/hours, rather than months/years.

010

Eric Talevich @etalevich.bsky.social · 09/02/2025

For scientific code, we're hosed because the whole point is to do something that's never been done before. Libraries are good enough that the new code is close to declarative. That means we're stuck with the older approach (thinking & looking up error messages).

000

Eric Talevich @etalevich.bsky.social · 09/02/2025

Use the most common and stable frameworks, basically just JS/Node, Python, and Java, because that's most of the training data. Major API changes between versions confuse the AI. (This means scientific computing is inevitably in the danger zone. Nextflow, yikes.)

100

Eric Talevich @etalevich.bsky.social · 09/02/2025

Agentic AI coding is leading to some new software development best practices: Keep projects small and modular to fit within the context window. Start with a good PRD. If the initial architecture is wrong, just start over. Use distinct, unambiguous names or the AI will get confused.

210

Eric Talevich @etalevich.bsky.social · 09/02/2025

Cline.bot and Replit are not really like a senior dev or a team of junior devs. They're amazing and a good fit for the needs of one-off web app development. They struggle with legacy and maintenance, as we all do, and they have their own unique quirks.

cline.bot

Cline - Autonomous Coding Agent for VSCode

Cline is an AI-powered coding assistant for Visual Studio Code.

100

Eric Talevich @etalevich.bsky.social · 09/02/2025

Is this chart specific to biotech? Seems like biotech is still in a cool season because AI has been soaking up new investment since 2022.

100

Eric Talevich @etalevich.bsky.social · 27/01/2025

Eleven years after Roche acquired Genia's nanopore sequencing tech, we see what they're up to: short-read sequencing competitive with Illumina and Ultima on price, throughput, and quality.

021

Eric Talevich @etalevich.bsky.social · 23/01/2025

For compbio in particular, so much of it is done for free in the public interest -- e.g. NCBI/EBI/DDBJ provide platforms, no? We commoditize eagerly. And I see AWS et al. expanding into this area partly because the barriers are so low.

010

Eric Talevich @etalevich.bsky.social · 23/01/2025

Extraordinary. You answered most of my questions in footnote 4. If we follow the money: Clinical development, sales, and marketing are where pharmas spend real money, right? Clinical research orgs (the other CRO) e.g. IQVIA, Paraxel, Icon, are not platform companies, but they toy with the idea.

010

Eric Talevich @etalevich.bsky.social · 16/01/2025

The fast, proprietary reimplementations are popular for "production at scale" use cases e.g. NGS diagnostics. You can prototype with OSS tools, then swap in Dragen/Sentieon/Parabricks equivalents to save time and compute costs, even after licensing.

130