Peter (@BiotechPedro) header
Peter (@BiotechPedro) profile picture

Biotechnologist & Biostatistician. I try to convert coffee into code and ideas for uncovering biological mysteries. PhD student @FraticelliLab @IRBBarcelona

Barcelona, Spain
Joined May 2017
Photo by Peter (@BiotechPedro)Photo by Peter (@BiotechPedro)Photo by Peter (@BiotechPedro)Photo by Peter (@BiotechPedro)Photo by Peter (@BiotechPedro)Photo by Peter (@BiotechPedro)Photo by Peter (@BiotechPedro)Photo by Peter (@BiotechPedro)Photo by Peter (@BiotechPedro)Photo by Peter (@BiotechPedro)
Deep learning has revolutionized the empirical sciences, but there have been cases where apparent progress turned out to be an artifact of missing baselines. We need to check this automatically to keep up with the deluge of new research articles. arxiv.org/abs/2610.40700
1
2
11
814
Hey, I heard about Bonsai, but now I heard there is another method called MILK that also visualizes scRNA-seq on trees. And it is much faster! Should I just use that then? A tree is a tree, right? TLDR: No. #methodsmatter On synthetic data MILK performs worse than UMAP. 1/n
6
13
1
118
13,071
#lipidtime: It's a lipid universe. Today in Nature we publish the Lipid Brain Atlas, the first map of membrane lipid composition across the entire mouse brain, a layer of organisation that cell-type and transcript maps had left out. #lipidomics #neuroscience #BrainAtlas
12
97
5
454
18,324
Mathematics has now fully morphed into biology: Project costs millions $ ✅ Put out paper that nobody has read ✅ Massive paper supplement nobody will read ✅ Advertise with pretty art ✅ Vicious authorship fight ✅ Say you're curing cancer ✅
45
655
48
8,043
315,531
As in biological evolution, one does not need to invoke consciousness, purposiveness, malice or ASI to see the destructive potential of powerful self-improving AI agents. [4/5]
1
1
1
289
Exactly... Because we are locked in an arms race between companies and between countries, we have been running with little control towards ever more powerful AI systems knowing that we do not have the systems in place to control the outcome... [1/5]
I’ve been reading a bunch of convoluted explanations and conspiracy theories about why AI companies are “suddenly” concerned about existential risk, which I’m finding genuinely bewildering. This isn’t that complicated. The labs are perfectly sincere in their concerns, and correct that AI is extremely dangerous. The people working in the field haven’t come to this view suddenly; many have been thinking and writing about this for years. Many researchers entered the field because they saw the existential risk posed by artificial minds bootstrapping themselves to superhuman capabilities, and decided the best way for them to reduce that risk was to work actively on building aligned AI systems. (Yes, this initially seems counterintuitive, but is a perfectly reasonable approach given the alternatives.) The risk isn’t overblown. If humans build computational systems capable of recursive self-improvement that aren’t exquisitely aligned with human interests, then the expected outcome really is that we all die. No-one really knows how far we are from that point, but the people who know best are exactly the same people currently sincerely expressing alarm. (People skeptical that misaligned superintelligence could kill us all are extremely confusing to me. Our species depends on a thin shell of habitable atmosphere around a single ball of rock. Even if a misaligned ASI wasn’t actually trying to kill us, almost anything it might want to achieve at scale would likely destroy us as an incidental side effect. Examining the catastrophic damage humans have managed to do to other species without intending to kill them off provides a valuable analogy.) The autonomous hacking incidents over the last couple of months are surprising in some of their details (especially the degree of inter-agent coordination), but not in their fundamental shape. Misaligned AI systems performing damaging actions in pursuit of their assigned goals is *exactly* what people thinking about this field have predicted for many years. Yes, there are ways that increasing regulation might benefit the big AI labs. No, that’s not the primary motivation for their proposals this week. We really are approaching the cusp of an existential threat: the technology for recursive self-improvement of AI systems is (probably) within sight, and market forces and geopolitics mean that without coordinated intervention our default path is a headlong scramble towards it. I know none of us want to be dealing with the impending threat of a sci fi apocalypse in 2026. It would be much simpler if this was all a confabulation or a self-interested conspiracy. But it really isn’t. I’m sorry.
1
2
14
3,759
Peter retweeted
WORLD'S TOP MATHEMATICIAN SAYS AI MAY BE MAKING US TOO EFFICIENT terence tao talks about how he used to go to the library for one paper and randomly find another one sitting next to it now u type exactly what u need into search or ai and get it in seconds his take: "you actually do need a certain level of distraction in your life." and i think this goes way beyond math we keep treating everything between the question and the answer like wasted time, but sometimes thats literally where the learning happened u looked for one thing, read something u didnt need, got confused, went somewhere else and came back knowing more than u originally asked for ai removing all that friction is insanely useful, i use it every day. i just dont think every second we remove from the process was useless sometimes getting the answer slower was part of what made the answer yours
123
264
72
2,139
296,779
Online today in @nature are the final correspondences on mtDNA lineage tracing, including our work with @Dana_peer, @livius_tacitus, @doctor_msc, @nawy_t, and @LeifLudwig, concluding a dialogue spanning 2+ years. A few final thoughts: 1/10 doi.org/10.1038/s41586-026-1…
6
33
1
208
34,280
There has been a slow but seemingly inexorable industrialization of academic science that, while apparently laudable for its “practicality”, risks stamping out that which makes academic science beautiful and precious to begin with.
15
32
4
270
23,040
FWIW the term “virtual cell” is neither high-level nor a concept. It’s marketing rubbish that is being used to hype vaporware in lieu of meaningful language that relates to solid science.
With all due respect, I'd like to offer a few points of clarification. First, I have no issue with "shortcut models." In fact, many of my own papers use relatively simple models to solve important real-world problems. If a simpler model ultimately proves capable of capturing complex cellular biology and helping cure disease, I'd be delighted. Science should reward what works, not what is most sophisticated. Second, terms like virtual cells, foundation models, and world models are high-level concepts that describe a class of models rather than a specific algorithm. Similar terminology has emerged naturally in computer vision and NLP as the field evolved. I think it's reasonable to adopt analogous concepts in biology as we explore whether they can unlock similar advances. Whether these ideas ultimately live up to their promise is, of course, an empirical question. Rigorous validation will decide. This is exactly what my original post is about. Healthy skepticism is essential, but so is giving ambitious new directions the opportunity to prove (or disprove) themselves. I don't think we should dismiss a promising research direction simply because the terminology sounds aspirational 🙏🙏
10
20
4
251
44,325
Peter retweeted
The day a blind man sees. The first thing he throws away is the stick that has helped him all his life
620
4,195
473
58,522
2,598,951
2/ Millions of papers a year, growing faster every year. Most aren't reproducible. Peer review is buckling. And every paper is a lossy compression of the work behind it — months of dead ends, judgment calls, and configuration tricks flattened into a clean story. The format was designed for a world where every reader was human. That world is ending.
1
12
1
100
8,641
I knew this was coming. We are so f*cked - fake and real now look the same. We desperately need tools to show journals and researchers that data is real, not AI-generated.
⚠️All of the below images were FABRICATED by ChatGPT Images 2.0, each with a single prompt❗️ ⚠️
65
238
18
992
127,623
Peter retweeted
I’m looking for an automated way to read others’s scientific data without giving credit or acknowledgement, and also claim full credit for insights from it. And I want it to have a fitting name OAI: say no more
Introducing GPT-Rosalind, our frontier reasoning model built to support research across biology, drug discovery, and translational medicine.
10
62
2
511
47,205
Yes we're all too busy publishing our own papers & hyping our own science but it really is our responsibility to also be part of the push and pull that is necessary to have a robust self correcting enterprise. 3/3
1
32
3,532
Replying to @dmitrypenzar
I'm actually not sure whether (1) EVO2 is the problem (2) training on ClinVAR is the problem (3) the predictor threshold is very poorly calibrated (this is one of the hardest things). Or all three.
1
5
22
2,514
Not cool at all. My colleagues and I checked several well-known pathogenic variants, and they were predicted as VUS. Certain variants are simply missing (e.g, there is no information for rs879254374). I like the idea, I don’t see any reason to use Evo2. Garbage in, garbage out.
very cool work. more dna language model pilled than i was an hour ago when i wrote my Evo 2 article a long time ago (owlposting.com/p/a-socratic-…) the question was always 'how do you interpret the outputs of these models if it just spits out log-likelihoods?' and increasingly the answer is 'use probes to annotate embeddings, and ask the LLMs to interpret those'. in feb 2025, when i wrote my article, i didnt really trust the models to do that well. now i do, and perhaps there is an immense amount of greenspace for really crazy things to be done but one of the big failure modes of LLM-for-science approaches is that the models try really hard to be consistent with the data instead of reality. theres no experimental validation here, so who knows whether that's happening here id be extremely curious for someone to cross-reference the annotated VUS's here with, e.g. the UK Biobank 500k~ exomes with linked phenotype data. maybe Claude one-shot human variant function annotation, or maybe the annotations are slop, or maybe its in between
6
16
3
111
27,226
7
43
5
191
20,174
AI for bio teams in academia & industry have been playing "offense" pumping out billion parameter models every other week that are worth less than a paper weight. How can anyone claim that the field needs "more offense" & less "rigor", when it requires literally the opposite?!?
This applies to academic labs too — maybe more than anywhere. Defense in academia looks like rigor: protect your niche, optimize your pipeline, wait for the field to settle. It isn't rigor. It's just slower irrelevance. The labs that matter in five years aren't asking "how do we use AI to do our old experiments faster." They're asking "what questions are now askable that weren't before?" That's the offensive posture. And academics have something founders don't — the freedom to chase questions with no obvious payoff. That's an enormous edge, if you use it.
3
15
159
16,561
Epigenetics is not a bystander in cancer. It links life’s exposures to disease risk. It’s time to bring DNA methylation into the clinic for risk prediction. nitter.gq/Aiims1742/status/20377…
Multiple recent papers (two just this week) highlighting the role of inflammation in sculpting “epigenetic memory” and promoting tumorigenesis across various solid organs. science.org/doi/10.1126/scie… nature.com/articles/s41586-0… science.org/doi/10.1126/scie… cell.com/developmental-cell/…
1
37
2
206
43,624