Genentech's Decima model explicitly learns cell type-specific expression by calculating its loss along the cell type axis.
nature.com/articles/s4159…
English
Kyle Farh MD PhD - Illumina AI Lab
41 posts
Kyle Farh MD PhD - Illumina AI Lab
@KyleFarh
Head of Illumina AI Lab. Deep Learning. Human Genetics. Perturb-seq.
San Francisco, CA Katılım Mayıs 2025
1.3K Takip Edilen264 Takipçiler
Most human 'recessive' disease genes may not actually be evolutionarily recessive - heterozygous LOFs in these genes have measurable effects on phenotypes and selection.
biorxiv.org/content/10.648…
English
One limitation of single-cell CRISPR screens is losing spatial context after tissue dissociation. This new paper introduces Spatial Perturb-seq to study gene knockouts in intact tissue. 🧠🧬
In mouse brain, it captures both cell-intrinsic effects and how perturbations alter ajacentd cell networks.
nature.com/articles/s4146…
English
Why do foundation models underperform in single-cell perturbation prediction?
This ICML’26 paper arxiv.org/html/2605.1934… suggests the issue isn’t scale, but representation: dominant cell-state variation overwhelms sparse intervention effects unless explicitly separated.
English
Kyle Farh MD PhD - Illumina AI Lab retweetledi
月刊David Baker Lab. 増刊号
nature.com/articles/s4158…
nature.com/articles/s4158…
日本語
Kyle Farh MD PhD - Illumina AI Lab retweetledi
Kyle Farh MD PhD - Illumina AI Lab retweetledi
Excited to share our collaborative work with the @TonyMustoe and @SuperSciJew labs, led by the fearless @RitwikaBose15.
In this study, we developed RAID-MAP, an approach to map RNA structure within biomolecular condensates in living cells.
biorxiv.org/content/10.648…
English
Kyle Farh MD PhD - Illumina AI Lab retweetledi
If you are interested in single-cell splicing QTLs, it will be worth trying our new tool ISSAC.
Sharing my talk at the Biology of Genomes 2026.
Preprint: medrxiv.org/content/10.648…
English
Kyle Farh MD PhD - Illumina AI Lab retweetledi
Regardless of arguments of efficiency or productivity, I highly recommend learning hands-on wetlab biology. There is something inexplicable when YOU, with your own hands, modify the genome of a living organism. Profoundly humbling experience, and nothing can emulate that feeling.
English
Kyle Farh MD PhD - Illumina AI Lab retweetledi
The writing is on the wall. There is no good reason why AI will stop improving at anywhere close to the human level of math capability. Especially since the rate of improvement from 0 to here was so fast, happening in about two years. Back in September 2024 I set to write a paper with my friend Kostas, with the theme "LLMs can't reason". And indeed, GPT-4o was terrible in solving deductive proofs that took me (and any trained logician) less than a minute to solve. While we were coding up, LLMs got to reason, and fast. Here is the paper we ended up writing: arxiv.org/abs/2605.12524 But models keep moving forward. Where will they be in two more years? Not where they are now, that's for sure.
English
Design of proteins with high specificity for binding arbitrary DNA sequences with RFDiffusion3 and AlphaFold3 helps identify specific binders for five therapeutically relevant targets; nearly a 100-fold boost over previous de novo methods. biorxiv.org/content/10.648…
English
Kyle Farh MD PhD - Illumina AI Lab retweetledi
Kyle Farh MD PhD - Illumina AI Lab retweetledi
David Liu’s lab used ProteinMPNN to tie up many loose ends in the evolved prime editors (PE6) by introducing as many as 163 amino acid residue substitutions, creating PE8 and showcasing the combined power of AI and directed evolution.
David R. Liu@davidrliu
Today in @NatureBiotech we report a new suit of PE8 prime editor proteins. PE8 variants were developed from laboratory-evolved PE6 proteins using AI-guided protein redesign. This approach combines recent advances in computational protein design and directed evolution to increase prime editing efficiency, especially in transient therapeutically relevant delivery settings such as mRNA+pegRNA electroporation into primary cells, eVLP delivery of prime editing RNPs, and LNP-mediated mRNA+pegRNA delivery in mice. drive.google.com/file/d/13VTcAs… 1/11
English
With AI it is all too easy to optimize the metrics by which we are measured by to the exclusion of others
nature.com/articles/s4158…
English
Recent work has shown that RoPE (rotary positional embeddings) can neither fully distinguish positions nor tokens in long sequences. Given the long DNA sequences used in to recent sequence-to-function models, it raises an interesting question: could better positional embedding schemes substantially improve these models’ capabilities?
arxiv.org/pdf/2605.15514
English
Kyle Farh MD PhD - Illumina AI Lab retweetledi
Our new paper just out in @FrontMicrobiol led by Emelia J Chamberlain.
From microbial #diversity to functional potential using dimensionality reduction
#MOSAiC #Arctic #Ocean #Microbiology
doi.org/10.3389/fmicb.…
English
Kyle Farh MD PhD - Illumina AI Lab retweetledi
People are realizing that AIs are nowhere near human intelligence and learning abilities.
Yet they have become very useful by compensating for their lack of common sense, lack of understanding of reality, and limited reasoning and planning abilities, by the accumulation of enormous amounts of declarative knowledge.
English
Kyle Farh MD PhD - Illumina AI Lab retweetledi
Here is another ground breaking research from Prof. @davidrliu - from Harvard & @broadinstitute and the co-founder of @BeamTx - $BEAM & @PrimeMedicine - $PRME, in which a new directed evolution of structured RNA motifs that enhance the efficiency of prime editing was introduced 🧵👇. According to this new research a use of these motifs resulted in substantial Prime Editing efficiency gains, including at sites of pathogenic mutation in primary human stem cells and in immortalized human bronchial epithelial cells using different delivery methods. The importance of these new findings is that this could make Prime Editing-based Therapies much more effective.
David R. Liu@davidrliu
Today in @NatBiotech, we report the directed evolution of structured RNA motifs that enhance the efficiency of prime editing. Iterated high-throughput pooled screens and mutagenesis of these small RNA elements improved transient pegRNA lifetime. drive.google.com/file/d/1Nyt_SS… 1/11
English
When evaluating single-cell perturbation prediction models, you need to be careful when using control cells to avoid overstating model performance. Basically, if control cells are shared, noise in the mean values for control cells can induce spurious correlations.
biorxiv.org/content/10.648…
English