Helmut G. Katzgraber

Nature just published a paper on an AI model that learns how diseases unfold. The model was trained on ~400,000 people - or rather their medical histories - in the UK and tested on 1.9 million in Denmark. In then model estimates the chance and timing of 1,000+ conditions based on a person’s past diagnoses (plus basic factors like sex, BMI, smoking). It can also simulate plausible future health trajectories for planning purposes. What's interesting 💡: If you have one model that updates risks for many conditions as your history changes, then there is no need for dozens of separate calculators. Also, you can simulate trajectories than then let you simulate futures. This is quite neat, as it can inform prevention and health‑system planning (e.g., projecting disease burden in ageing populations). However, there are a number of important caveats. 🚨 For some outcomes (e.g., dementia and mortality) it performs on par with or better than well‑known clinical scores, but for others (like diabetes), a single lab test still wins. Also, many associations were already well known. And lastly, this was trained primarily on UK Biobank, which skews healthier and has gaps (e.g., limited data >80y). Patterns of what gets recorded (hospital vs primary care) will bias predictions. That said, I think this is an interesting proof-of-concept, particularly interesting for population‑level planning and targeted screening. Future versions can add labs, prescriptions, imaging, or wearables, and should be much more powerful. Which, coincidentally, is a good reason to start collecting such data already! Fun fact: Delphi’s architecture is based on GPT-2 🔥 Link OA paper: https://lnkd.in/eERWtS69

194

9 Kommentarer

Probably one of the most important articles in biotech AI this year...🧬👇 A new Nature Biomedical Engineering paper describes CRISPR-GPT, a large language model–driven, multi-agent system for automating CRISPR gene-editing workflows (link in the comments). ⚙️ The system coordinates multiple specialized AI agents to do these things: - Select appropriate CRISPR systems, design guide RNAs, and choose delivery methods. - Generate experimental protocols and assay plans. - Analyze results from wet-lab experiments and adapt subsequent steps. I can't stress enough the potential significance of this work, which IMO lies in the integration of computational reasoning with experimental execution, enabling “closed-loop” cycles where experiment design, execution, and analysis are connected and automated. The authors tested their agentic CRISPR-GPT system in real experiments, e.g., these: 🧬 Knockout of four genes in a human lung adenocarcinoma cell line. In the multigene knockout experiment targeting four genes (TGFβR1, SNAI1, BAX, BCL2L1) in A549 lung adenocarcinoma cells, the AI-generated protocol achieved consistently ~80% editing efficiency across all targets, as measured by NGS analysis 🧬 Activation of two genes in a human melanoma cell line. In the epigenetic activation experiments in a human melanoma cell line, the reported efficiencies were approximately 56.5% for NCR3LG1 and 90.2% for CEACAM1, based on flow cytometry comparing gRNA-edited groups versus negative controls. With this kind of agentic AI tools, it could become possible to explore larger experimental spaces more systematically and at greater speed. This really clicks with my vision of what the key role of AI systems is in drug discovery and biotech research: tying pieces together and allowing for holistic discovery vs classical reductionism (such as "target-ligand"-centric DD workflows). We outlined our ideas about this with Oleg Kucheriavyi earlier this year, in our industry report "Beyond Legacy Tools: Defining Modern AI Drug Discovery for 2025 and Beyond." The report is published with BioPharmaTrend.com, check it out via the link in the comments 📑 👇. Looking forward, systems like CRISPR-GPT could evolve into general-purpose lab intelligence, able to handle multi-modal data, integrate with robotics, and support continuous, iterative discovery. The space is worth watching. If you are following AI drug discovery and other deep tech trends, subscribe to Where Tech Meets Bio, a leading Substack newsletter in this niche (link in comments). Image from the article (citation in comments)

370

52 Kommentarer

Oh, Dear! Yet more magical thinking about reducing hallucinations of LLMs by using diffusion models. As Denis O. critiqued them, this won't happen. Diffusion models are at least as much subject to hallucinations as transformer based LLMs. Anyone who has used a GenAI image generator knows of the difficulty in getting it to faithfully follow the prompts. "It seems likely that diffusion LLMs could challenge auto-regressive tech, offering new capabilities like improved reasoning and controllability, but their full impact is still emerging." As always, the proponents of these new technologies keep talking about reasoning. It is just not going to happen and we know that image generators are highly uncontrollable. The idea of generating a whole block of text in parallel is wrong from a semantics and linguistics perspective. Language has sequential dependencies. An argument is developed sequentially. A critical evaluation is developed sequentially, not in parallel. "Andrej Karpathy and Andrew Ng, both renowned AI researchers, have enthusiastically welcomed the arrival of Inception Lab’s diffusion LLM." Surely, this is the kiss of death, not an endorsement. "This process begins with random noise which is then gradually refined and “denoised” into a coherent stream of tokens. This is analogous to how diffusion models generate images by starting with noise and iteratively removing it to reveal a clear image." It is not analogous, it is exactly the same, need to get the facts right. "Parallel Processing: Diffusion models can process and generate text in parallel, potentially leading to significant speed advantages." As with image generators, the final output of tokens will be fast, just transferring out of the output buffer to the user, but after a long pause (not even thinking) to do all the denoising. We also know that diffusion models eat TFLOPs, just as much as or more than transformers. https://lnkd.in/eq2HWhWZ

73

25 Kommentarer

🔮 Why AI4Science is hard: open the physics challenge? In the past few months at conferences and seminars at Stanford, Cambridge, and Yale I’ve been asked a lot: what are the real use cases for ML in experimental physical sciences - and why is it so difficult? Indeed, the headline successes over last several years - symbolic regression “rediscovering” Kepler laws, SINDY learning Newtonian mechanics or learning ODEs/PDEs on clean simulation and real world turbulence are inspiring, but transition to the physics labs seems to be elusive. In fact, my own journey into electron microscopy ~15 years ago started from realization that scanning probe microscopy data is very difficult to analyze by ML. At that time, we interpeted these difficulties as "calibration" and "uncontrolled factors". Now the word will be distribution shifts and hidden latent variables, but the central premise is the same. Conversely, in electron microscopy atom is an atom - what you see is what you analyze. Hence, we can think about “learning physics” from microscopy as e.g. fitting force fields from snapshots of atomic configurations. However, the reality is still more complicated - between observations and underpinnign physics there is a thick layer of physics abstractions. In scanning probe and electron microscopies, most phenomena are strongly entangled (ferroelectricity and electrochemistry for PFM, electronic and structural order for STM, and so on), and the physics is only partially known. After interetsing discussions with Georgia Channing, I have meditated on what would a good discovery benchmarks look like? Here’s a concrete one: STM on α-RuCl₃, a honeycomb magnet with Kitaev-type (bond-dependent) interactions. Briefly, Kitaev materials are magnets where the atoms sit on a honeycomb-like lattice and their spins interact in a very unusual, direction-dependent way: the type of spin coupling depends on which bond connects two atoms. This behavior comes from strong spin–orbit effects in certain crystals made of heavy elements. In the ideal case, these materials can host a “quantum spin liquid,” a state with unusual magnetic order and exotic fractional excitations. Now, this is the idealized physics. The task for AI: how much of this physics can be discovered or hypothesized from an STM image showing unusual non-periodic patterns of electronic density? How can prior physics context or simulation help understand these phenomena? Now, over my ~30 year career in scanning probe microscopy, I can probably identify 5-6 examples like that from my and other's work (chaos in ferroelectric switching, Berry phases from STM, few others) and several examples where the answer is still out. But may be this is a start! If we want ML that discovers rather than denoises, we need community benchmarks like this - where success means suggesting mechanisms and proposing the numerical model. You can read more about RuCl3 here: https://lnkd.in/eeRbFVDG

101

12 Kommentarer

We're pleased to announce that Fujitsu Quantum has released a new research paper: "Learning quantum many-body data locally: A provably scalable framework." This paper introduces a groundbreaking, provably scalable machine learning framework designed to efficiently learn quantum many-body data obtained from quantum experiments by leveraging data locality. We believe this work is a significant step towards fusing two innovative technologies, quantum computing and machine learning, to open new avenues for accelerating materials development, drug discovery, and the fundamental understanding of our universe. #Fujitsu #FujitsuResearch #FujitsuQuantum #Quantum #MachineLearning https://lnkd.in/gJ_a4Z_Y

167

2 Kommentarer

I moderated a thought-provoking panel on Generative AI's impact on patents at the PatentSemTech workshop, SIGIRConf 2025 in Padua. Panelists included Eleni Kamateri, Allan Hanbury, Linda Andersson, Sebastian Björkqvist, and Gabriella Pasi. Our discussion centered on "Human in the Loop" as crucial. Generative AI, particularly Large Language Models (LLMs), offers powerful capabilities for tasks like prior art search and drafting. Our panelists emphasized these tools are not infallible, likening LLMs to "teenagers" – brilliant but unpredictable. Human oversight is vital for accuracy in this precision-demanding field. We explored the "paradigm change" in patent search, from Boolean to relevance ranking, and now to vector search. Notably, while the shift from Boolean to relevance took a considerable time to be widely accepted in patent search, the current rapid adoption of vector search and Generative AI-driven search hints at a completely different pace. Yet, integrating LLMs presents hurdles: inherent biases, high costs, and limited AI democratization due to resource concentration in large companies. Beyond search, AI also holds promise to smoothen the overall patent process, and crucially, help scientists better understand the complex language of lawyers and patent examiners, bridging vital communication gaps. Panelists agreed "finetuning is not the solution" for patent accuracy, as it's often too slow and expensive to maintain. Instead, the future lies in "small models with domain knowledge." These models are less resource-intensive, making them more accessible and controllable for researchers. We also debated "Why do we need long patent text?" suggesting better, more symbolic information representations. The patent system, though complex, is here to stay. The challenge lies in enhancing prior art search quality amidst industry obfuscation and AI-generated "prior art." Ultimately, AI literacy is crucial. Professionals must understand these tools and combine them strategically with their indispensable domain knowledge. The path forward requires cautious, informed adoption, integrating diverse tools, and maintaining expert human judgment at the core of the patent process. It was an enlightening discussion on AI's exciting, complex journey in intellectual property. #SIGIRConf #PatentSemTech #AI #Patents

78

5 Kommentarer

Brand new. Flood gates have opened up for applied category theory. Category Theory for Supercomputing: The Tensor Product of Linear BSP Algorithms Thomas Koopman, Rob H. Bisseling, Sven-Bodo Scholz https://lnkd.in/gCvGRvPM We show that a particular class of parallel algorithm for linear functions can be straightforwardly generalized to a parallel algorithm of their tensor product. The central idea is to take a model of parallel algorithms -- Bulk Synchronous Parallel (BSP) -- that decomposes parallel algorithms into so-called supersteps that are one of two types: a computation superstep that only does local computations, or a communication superstep that only communicates between processors. We connect each type of supersteps to linear algebra with functors. Each superstep in isolation is simple enough to compute their tensor product in Vect with well-known techniques of linear algebra. We then individually translate the tensor product of supersteps back to the language of BSP algorithms. The functoriality of the tensor product allows us to compose the supersteps back into a BSP algorithm for the tensor product of the original function. We state the recipe for creating these new algorithms with only a minimal amount of algebra, so that it can be applied without understanding the category-theoretic details. We have previously used this to derive an efficient algorithm for the higher-dimensional Discrete Fourier Transform, which we use as an example throughout this paper. We also derive a parallel algorithm for the Discrete Cosine Transform to demonstrate the generality of our approach.

50

2 Kommentarer

What if the accelerating universe is not driven by an unknown force, but the informational residue of the universe's quantum fabric? In work published today, we show how the Quantum Memory Matrix (QMM) framework explains cosmic acceleration without invoking new exotic fields. In this picture, each Planck-scale cell of spacetime has a finite information capacity. Once saturated, it leaves behind a small, residual vacuum imprint energy, which is precisely of the cosmological-constant form and magnitude that we observe. We also find that if imprinting continues but is slowed by the Hubble expansion, it produces a slow-roll entropy field with an equation of state just slightly above -1. This yields concrete, testable predictions for Planck 2018, BAO, and Pantheon+ data and forecasts for next-generation surveys like Euclid, DESI, and the Roman Telescope. This result suggests that dark matter and dark energy are two sides of the same information field: one gradient-dominated, the other potential-dominated. Rather than being added ad hoc, the dark sector may simply be the informational footprint of spacetime itself. The work was just published in Astronomy under the title "Extending the Quantum Memory Matrix to Dark Energy: Residual Vacuum Imprint and Slow-Roll Entropy Fields": 🔗 https://lnkd.in/ggK6DG2E This paper continues a broader effort to explore information as the most fundamental layer of physics. Other recent directions include: 🧠 The Quantum Memory Matrix and black-hole unitarity 🌌 Cosmic cycles and primordial black holes 🔥 Informational dualities between entropy and complexity 🌍 Information-based interactions beyond gravity Each step strengthens the idea that information may not just describe reality. It might be reality itself. Thank you to my co-authors Eike Marx, Valerii Vinokur, Terra Quantum AG, and collaborators for making this journey possible. #Physics #Cosmology #DarkEnergy #QuantumInformation #FundamentalPhysics #SciPhi Terra Quantum AG | Leiden University

194

60 Kommentarer