←
Jake P. Taylor-King933 posts
Jake P. Taylor-King avatar

Jake P. Taylor-King

@wildtypehuman

Co-founder at Relation (@relationrx) - applying #singlecell (check @scTrends_update !) #genomics #crispr #machinelearning to #drugdiscovery and #healthcare.

◷ Joined October 2014▧ 49 Photos and videos
Jake P. Taylor-King avatar
Jake P. Taylor-King@wildtypehuman·11h

Why Iso will end up buying biology companies… At a ~$50bn valuation, biomolecular design deals (where pharma brings the target) won’t be enough. Their existing collaborations include upfronts, milestones and royalties reaching low double digits (probably with hurdle conditions associated). Very good deals, but a 5% royalty on a billion-dollar drug is $50m a year. You need *a lot* of successful drugs to make the valuation work. Add competition from Chai, Boltz etc (that operate on a fee-for-service basis), and you realise that Iso needs to own a much bigger share of the upside. It need its own future blockbusters, only partnering later for commercialisation. What probably won’t work… Designing a molecule against a target and knowing that manipulating that target will help patients are different problems. They could compete on established targets. But then they need to get there faster, deliver a better drug, or both. But, Iso is still preclinical. Buying existing drug assets could get them towards market faster. But it creates an awkward question for the platform story: If your pitch is that you can design drugs better than anyone else, why spend billions buying drugs someone else designed? Going after new biology is the most compelling option. First-in-class mechanisms give pharma a reason to buy into an entirely new franchise. Incremental molecular improvements put Iso into a crowded race against companies with every incentive to defend and extend existing franchises. The options… - University licensing? Useful, but buying a promising academic finding before it goes public isn’t necessarily a target package ready for drug discovery. - Buy datasets and build target discovery internally? Absolutely possible. But data still needs to become a biological hypothesis—and survive experimental testing. - Acquire companies with compelling targets, proprietary evidence and the teams who understand them? That last option is where I think they end up.

Thrilled to announce $5B in funding for @IsomorphicLabs to accelerate our mission to Solve All Disease. We continue to make rapid progress reimagining drug design from first principles with AI to radically cut the time & cost to find new medicines. Gradually, then all at once...

0:37
Rahul Satija avatar
⟳ Jake P. Taylor-King retweeted
Rahul Satija@satijalab·Nov 9

Excited to share Seurat v6 and scSLIDE! Single-cell studies usually compare cells, but scSLIDE lets us compare *samples*. Instead of forcing patients into case/control groups, we can model sample-level variation. Paper: cell.com/cell/abstract/S0092… Software: satijalab.org/seurat

Jake P. Taylor-King avatar
Jake P. Taylor-King@wildtypehuman·Nov 5

This is seriously impressive! In a world where CRO costs are steadily creeping upwards year on year, @LindusTx have been able to buck the trend!

Meri Beckwith @meribeckwith Nov 5 Article Cheaper clinical trials using AI - announcing a $1m phase 2a trial in Neurology The last two weeks have seen a frenzy of announcements in AI x bio; from a 3 person team at Citrini research creating a plausible scorpion antivenom (binder) in 3 days for $200k (1), to Anthropic

Hesam Hakim avatar
⟳ Jake P. Taylor-King retweeted
Hesam Hakim@hesamhakim·Nov 4

Most genetic variants linked to disease change how much of a gene is made, not what the protein is. Testing them one at a time is slow. A preprint from @JoachimDeJonghe, @JensBagerChris1, @wildtypehuman, and colleagues, not yet peer-reviewed, installs them by the thousand and reads the effect cell by cell. PEP-seq combines prime editing, which rewrites a short stretch of DNA in place without cutting both strands, with perturb-seq, which reads each edited cell's full set of active genes. They designed 2,164 guides covering common expression-altering variants (eQTLs), premature stop codons, and matched silent controls, and collected 749,110 single-cell transcriptomes over time. Two checks came out well. Stop codons produced the expected decay of the faulty messenger RNA. And for the eQTLs, the gene the variant was supposed to affect was often the most affected gene within a 1-megabase window around the edit, with effects tracking open chromatin near the site. The variant is edited in its own place in the genome, not tested on a reporter, and that is the strength of the design. As a preprint, these are claims until the full data are out. doi.org/10.64898/2026.10.28.…

Post attachment
Jake P. Taylor-King avatar
Jake P. Taylor-King@wildtypehuman·Nov 4

Do you want to feel good about politics and humanity? Just remember this: people *do not* want to go to war — despite what the politicians say. Same lesson in Europe and America, just different fronts (Russia / Iran).

Relation avatar
⟳ Jake P. Taylor-King retweeted
Relation@RelationRx·Oct 29

Relation not only builds AI models for drug discovery, but improves the substrate for the underlying data generation.

What kind of data do we need for DNA foundation models? At @RelationRx, we have developed a new protocol for prime-editing perturb-seq to help generate training data. biorxiv.org/content/10.64898… CC @JensBagerChris1 , @j_opzoomer (other authors not on twitter!) Thread below..

Post attachment
Dylan Webster avatar
⟳ Jake P. Taylor-King retweeted
Dylan Webster@DylanPWebster·Oct 30

Super cool. Seems like a lot of related recent work is pointing to a vision of amazing things to come unraveling eQTLs. These are the building blocks that make AI x bio promises tangible.

What kind of data do we need for DNA foundation models? At @RelationRx, we have developed a new protocol for prime-editing perturb-seq to help generate training data. biorxiv.org/content/10.64898… CC @JensBagerChris1 , @j_opzoomer (other authors not on twitter!) Thread below..

Post attachment
sam avatar
⟳ Jake P. Taylor-King retweeted
sam@discoplomacy·Oct 30

British politics would be improved somewhat if politicians replaced “Britain should do X” with “I want to do X”

Tessa Green avatar
⟳ Jake P. Taylor-King retweeted
Tessa Green@the_tessallator·Oct 30

Prime editing perturb-seq has been on my wishlist for ages 😍 thrilled to see it executed by @wildtypehuman and team

What kind of data do we need for DNA foundation models? At @RelationRx, we have developed a new protocol for prime-editing perturb-seq to help generate training data. biorxiv.org/content/10.64898… CC @JensBagerChris1 , @j_opzoomer (other authors not on twitter!) Thread below..

Post attachment
Mario Creatura avatar
⟳ Jake P. Taylor-King retweeted
Mario Creatura@MarioCreatura·Oct 29

Utterly heartbreaking. A must watch.

1:51
Jake P. Taylor-King avatar
Jake P. Taylor-King@wildtypehuman·Oct 29

Then, we wanted to see if we could edit in eQTLs. Yes we can, but there are nuances: you have to think carefully about what genes are edited within the cis- window. Often other genes change in gene expression, but the changes aren’t significant compared to the target gene. We also have multiple timepoints within our dataset… it seems the known eQTLs are constant across timepoints, but bystander genes are more likely to fluctuate in expression. Finally, the likelihood of gene expression changes improves if the target variant is within an open chromatin window. Takeaway, we have a scalable framework for interpreting non coding variants functionally, this opens new avenues for testing observed and synthetic variants genome wide within its genomic context, a new frontier for functional genomics :D More validation work is happening in the background to check the fidelity of our guides and edits! Big thank you to all of my coauthors and Joachim who led the charge! Follow to hear updates.

Post attachment
Jake P. Taylor-King avatar
Jake P. Taylor-King@wildtypehuman·Oct 29

So with that, we wanted to check we could measure PTC edits that induce nonsense mediated decay - we can! We also saw that knocking out essential genes reduced representation in the library. Some edits suggested that gene expression changes were uneven across the probe set, leading to compositional change.

Post attachment
Jake P. Taylor-King avatar
Jake P. Taylor-King@wildtypehuman·Oct 29

We set out to have a cheap, flexible method for profiling edits, considering the benefits of working with fixed cells, we bet that probes were the right approach -- but how to profile 1000's of edits in multiplex? We were inspired by @satijalab's VIPerturb-seq approach (biorxiv.org/content/10.64898… ) and created our own adaption for prime editing (see preprint for QC)

Post attachment
MOM avatar
⟳ Jake P. Taylor-King retweeted
MOM@mom_agency_·Oct 27

Someone turned the AI CEO drama into a Spanish soap opera!

1:25