Born in Italy, bioinformatician fascinated by non-coding vars. When not at computer, I like traveling, hiking and exploring new food.
Milan, Lombardy
Joined June 2009
- Tweets2.9K
- Following147
- Followers274
- Likes393
Edoardo Giacopuzzi retweeted
Why do genetically identical cells of the same type behave so differently? Can DNA shape how variable that behaviour is?
Our paper is out in @NatureComms. We built MEOTIVE to find genetic variants associated with cell-to-cell variability in expression. 🧵
nature.com/articles/s41467-0…
Edoardo Giacopuzzi retweeted
SAIGE-QTL is out in @NatureGenet! nature.com/articles/s41588-0…
SAIGE-QTL is a robust and efficient method for single-cell eQTL mapping, which 👇 (1/n)
Edoardo Giacopuzzi retweeted
1/ 🔥New study: Most perturbation models answer “if I knock out gene X, what happens?” We asked inverse: given a disease state, what drove them? 🔄
ExPert traces patient cells to their cause: hidden liver-disease biology, mitochondrial defects from blood biorxiv.org/content/10.64898…
Edoardo Giacopuzzi retweeted
Scooby-2: fine-tuning AlphaGenome for single-cell genomic track prediction biorxiv.org/content/10.64898…
Edoardo Giacopuzzi retweeted
Given substantial size relative to other widely available imputation panels, the 490k-based UK Biobank imputation panel demonstrates improved imputation across diverse ancestries nature.com/articles/s41588-0… @NatureGenet
Edoardo Giacopuzzi retweeted
CellWhisperer uses multimodal learning of transcriptomes and text to answer questions about single-cell RNA-sequencing data go.nature.com/3WPGJnW
rdcu.be/TamrmUPBYm6a
Edoardo Giacopuzzi retweeted
Our three TenK10K papers are out in @Nature today!
I am honoured to lead one of the studies addressing the "missing regulation" puzzle for human complex diseases, integrating gene expression and chromatin accessibility of 9M PBMCs from 1925 donors.
nature.com/articles/s41586-0…
Edoardo Giacopuzzi retweeted
Can someone find your genome using your electronic health record? 🔍 🧬
A new study in @NatureComms explores genome re-identification using polygenic scores.
Imagine someone has your medical records with your height, weight, BMI and blood test results. They also have a database of hundreds of thousands of anonymous genomes. Can they find yours?
The idea is simple. Calculate polygenic scores for each genome and compare predicted traits with real ones. The authors developed PGMatch, an algorithm that accounts for prediction accuracy and correlations between traits.
They tested it on UK Biobank data using 40 traits.
With equal numbers of matching and non-matching pairs, the algorithm achieved over 99% precision at 40% recall.
But with just one real match per 1,000 incorrect pairs, average precision dropped to 13.4%.
Interestingly, earlier research showed that GWAS participation could be detected with 98% sensitivity at a false positive rate of 0.1% using 50,000 SNPs in a study of 2,000 people. However, this required knowing the person's genotype.
The conclusion is that even with access to your medical records and a huge database of anonymous genomes, finding yours would be extremely difficult.
So for now, no one is likely to identify your genome and use it to poison you with whole milk after discovering that you are lactose intolerant.
nature.com/articles/s41467-0…
#PRS #GWAS #anonymization #UKBiobank
Edoardo Giacopuzzi retweeted
Upstream methodological decisions of RNAseq processing may influence the detection & estimation of eQTLs, thereby impacting coloc & TWAS results in previously unappreciated ways nature.com/articles/s41588-0… @NatureGenet
Edoardo Giacopuzzi retweeted
SCI member Michael Bassik & others mapped over 92 million enhancer-gene interactions across human cell types and tissues, improving our ability to link genetic variants to the genes they affect and understand the biology underlying disease. brnw.ch/21x6GdE
Edoardo Giacopuzzi retweeted
MACS 3.1.0 is released with a jaw-dropping 40-60x improvement in speed, thanks to @jmschreiber91's auto-optimize work! AI achieved in one weekend what I've been hoping for years! Now peak calling that once took minutes can be completed in seconds! github.com/macs3-project/MAC… 1/2
Edoardo Giacopuzzi retweeted
Our cells carry the same DNA, but how this DNA is interpreted differs wildly from cell to cell. In a new preprint, Science Fellow @zhou_jingtian and team introduce Bolero, a model conditioned on cell state that predicts chromatin accessibility and expression from DNA alone.
Edoardo Giacopuzzi retweeted
Introducing Omnii-1, our most powerful genome language model, and a major step toward general biological intelligence
In a world first, Omnii-1 designed the first AI-generated bacteriophages to be administered to a human patient with an antibiotic-resistant infection under compassionate use
Edoardo Giacopuzzi retweeted
Structural variant detection across sequencing technologies nature.com/articles/s41587-0…
Edoardo Giacopuzzi retweeted
Tahoe-100M: Mapping drug-induced molecular phenotypes at single-cell resolution cell.com/cell/fulltext/S0092…
Edoardo Giacopuzzi retweeted
Online Now! Single-cell CRISPR screen maps epigenetic regulators of lineage specification and malignancies in human HSPCs dlvr.it/TVvP2Z #stemcells
Edoardo Giacopuzzi retweeted
1/ Most genes and variants matter somewhere but look silent in any one cell model. Can we build conditions that reveal phenotypes? With @kanishkadey, we engineer cell states to expose hidden disease biology and trace how genetic risk reaches disease genes. biorxiv.org/content/10.64898…
ALT Visual abstract: activating TFs singly and in pairs engineers new cell states, and the Mosaic model learns how TFs combine. Findings: hidden AMD risk variants, state-dependent convergence, and chromatin priming.
Edoardo Giacopuzzi retweeted
SpliceAI told you a site might fire. SpliceAI2 asks the question your call sheet actually needs: which transcript got made, and how hard.
Illumina + Oxford / UCSF / NYGC put SpliceAI2 on bioRxiv (posted Oct 11). It takes ~200 kb of genomic sequence (196,608 bp in the architecture), predicts position-wise splice-site usage and pairwise junction usage, then scores full transcripts by walking a splice graph with dynamic programming — nodes = sites, edges = junctions, path weight = predicted isoform usage.
Training stack (paper numbers only):
• 314,745 short-read RNA-seq samples across human + 9 other mammals
• 330 long-read samples (human + mouse) for transcript usage
• Held-out genes: perfectly reconstructs the most abundant transcript for 82% (78% without the long-read signal)
Head-to-head vs SpliceAI / Pangolin / AlphaGenome on three independent benchmarks (essential splice variants excluded throughout):
• GTEx cryptic GT/AG: auPRC 0.77 vs 0.66 (SpliceAI); usage Spearman ρ 0.63 vs 0.47 (Pangolin). Bench: 4,545 cryptic + 573,967 null
• GTEx sQTL fine-map ranking: auROC 0.76 vs 0.73 (AlphaGenome); effect-size ρ 0.73 vs 0.68
• OpenSplice MPRA: Spearman ρ 0.59 vs 0.55 (SpliceAI)
Rare-disease math that sticks: in their cohort, SpliceAI2-prioritized cryptic splice variants account for 15% of excess genetic burden — on top of protein-truncating variants at 50%. About half of those cryptic hits sit >50 bp into introns, so exome panels were never going to see them. Blood RNA-seq from 5,435 GEL participants tracks the scores; at high thresholds, validation approaches essential-splice rates.
They also release pretrained weights plus precomputed effects for ~4 billion SNVs and 150 million population indels inside protein-coding genes (GENCODE v48 / GRCh38), code at github.com/Illumina/SpliceAI….
Bookmark if you triage WGS rare-disease cases, run sQTL / cryptic-splice QC, or still live on 2019 SpliceAI delta thresholds.
One concrete check: on your last unsolved WGS proband with a deep-intronic candidate, does SpliceAI2 push the score past 0.25 — and if you have blood RNA-seq, does the predicted junction show up?
biorxiv.org/content/10.64898…
