Predicting unknown mammalian viral hosts with Dynamic Positive–Unlabeled learning
nature.com/articles/s41467-0…
1
53
Stefano Giagu retweeted
When missing labels are not negatives: graph learning separates biology from sampling bias in virus-host prediction
In many scientific datasets, “not observed” quietly becomes “does not exist.” That can be a very bad assumption.
For mammal-virus associations, a missing link may mean that the virus truly does not infect that host. But it may also mean that nobody has looked carefully enough yet. Sampling effort is highly uneven across species, viral families and geographic regions.
Pignalberi and coauthors build that distinction directly into the machine learning. Their Dynamic Positive-Unlabeled framework learns two different things at once: a graph neural network estimates whether a mammal-virus association is biologically plausible, while a separate propensity model estimates how likely that association would have been observed if it really existed.
That means an unobserved association in a heavily studied species is treated differently from an unobserved association in a poorly sampled one.
The test I find most convincing comes from associations that were discovered only after the training data were assembled. Those newly confirmed links received a median predicted probability of 0.82, compared with less than 0.01 for associations that remained unlabeled.
The broader lesson is important: When labels are produced by an imperfect observation process, model that process too. If “missing” can mean either “false” or simply “not measured yet,” treating every missing label as a negative teaches the model your sampling bias as if it were science.
Pignalberi et al., Nature Communications (2026), CC BY-NC-ND 4.0. doi.org/10.1038/s41467-026-7…
Stefano Giagu retweeted
Nature has just published a pretty incredible paper.
The authors introduce a framework that turns research papers into interactive AI agents.
Rather than engaging with a paper only as static text, researchers can question its agent, ask it to reproduce an analysis, or apply the paper’s methods to new data.
This could make research methods much easier to understand, reproduce, and use.
It also opens up a strange possibility: research papers that can, in a sense, talk to one another. The authors demonstrate this by connecting agents derived from different papers so that they can combine their methods and data.
Used carefully (and with human scientific judgment still firmly in the loop) this could become a powerful new form of scientific communication.
*
Full paper in the first reply
Stefano Giagu retweeted
Join the EPIGRAPHY colloquium by James Clegg, Principal Optical Scientist at Microsoft Research Lab (Cambridge, UK) on April 16, 2026 at 3:00 PM (CET).
Title: Analog Optical Computing at Microsoft Research UK: Improving Efficiency in the AI Era
indico.cern.ch/event/1667483
Stefano Giagu retweeted
EPIGRAPHY had its kickoff event in Lisbon last week, from Feb 12-15, 2026! Lots of interesting talks on #deeplearning, #EdgeAI, and #particlephysics. We also connected with other communities like #EuCAIF, #NextGenTriggers, and @MLQC4FC ! See you soon at the next EPIGRAPHY event!
Stefano Giagu retweeted
We are extremely delighted to announce the kickoff school for our COST Action EPIGRAPHY in Lisbon, from Jan 12-15. While in-person invitations have gone out to EPIGRAPHY members, a public Zoom link will be provided to follow the talks & tutorials online: indico.cern.ch/event/1608897…
Stefano Giagu retweeted
New Anthropic research: Persona vectors.
Language models sometimes go haywire and slip into weird and unsettling personas. Why? In a new paper, we find “persona vectors"—neural activity patterns controlling traits like evil, sycophancy, or hallucination.

ALT Our automated pipeline takes as input a personality trait (e.g. “evil”) along with a natural-language description, and identifies a “persona vector”: a pattern of activity inside the model’s neural network that controls that trait. Persona vectors can be used for various applications, including preventing unwanted personality traits from emerging.
Stefano Giagu retweeted
🚨 Our paper "Mixture of Expert Graph Transformer for Particle Collision Detection" is now published in Scientific Reports!
🔍 We propose a novel graph-based approach to improve particle collision detection. link.springer.com/article/10…
@s_scardapane, @devoto_alessio, @StefanoG
Stefano Giagu retweeted
🚀 And we're off! On 5 May 2025, the ATLAS Experiment recorded its first 13.6 TeV collisions of the year.
These event displays capture the first moments of a promising new data-taking season at @CERN's flagship particle accelerator. Learn more: cern.ch/run3-2025
Stefano Giagu retweeted
📣 We have beam in the tunnel 💫
Today, the Large Hadron Collider received the first particle beams of 2025. #BeamTime
Protons were injected into the #LHC for the first time since November 2024, when the annual year-end technical stop (YETS) started. This is the fourth year of #LHCRun3, and the new data-taking season is about to begin.
Stay tuned for more #ParticlePhysics news.
Quantum Computing and Quantum Neural Network workshop at IJCNN2025 web page is now available at sites.google.com/view/nesya/…
1
1
115
call for paper is also open, deadline March 30th: drive.google.com/file/d/1Tkr…
1
73
Stefano Giagu retweeted
*MoE Graph Transformers for Interpretable Particle Collision Detection*
by @d_genovese @devoto_alessio @StefanoG
We propose a MoE graph transformer for particle collision analysis, with many nice interpretability insights (e.g., expert specialization).
arxiv.org/abs/2501.03432
Stefano Giagu retweeted
L’Italia si conferma protagonista nello sviluppo dell’ #intelligenzaartificiale in Europa.Si chiamerà infatti #IT4LIAAIFactory la piattaforma strategica, una delle prime in Europa, di intelligenza artificiale. Qui tutte le info
👉🏻 mur.gov.it/it/news/martedi-1…
Stefano Giagu retweeted
#intelligenza_artificiale #fisica #NobelPrize2024 Lo stretto legame tra fisica e intelligenza artificiale tinyurl.com/mpaw5cu8























