Maurizio Morri – Science Blog

A major breakthrough in cancer care is making headlines. Scientists from the US, UK, and Switzerland have developed a cutting-edge test that predicts which men with aggressive, non-metastatic prostate cancer are most likely to benefit from the drug abiraterone. This is significant because while abiraterone can save lives, it also causes serious side effects such as high blood pressure, diabetes, and heart complications.

Until now, doctors had no reliable way to know who should receive the drug. This new test, powered by machine learning, changes that. It analyzes digital images of tumor biopsy samples and identifies a specific biomarker that indicates whether a patient is likely to respond to treatment.

In a study involving over one thousand men, the test found that twenty five percent of them had this biomarker. For these patients, abiraterone reduced the risk of death within five years from seventeen percent to nine percent. For the rest, the drug showed little to no effect, meaning they could avoid unnecessary treatment and side effects.

The test is designed to work with routine clinical data and can easily be added to existing hospital workflows. It promises to make prostate cancer treatment more precise, sparing patients who do not need aggressive therapies and ensuring those who do get the help they need.

Researchers also hope the findings will encourage broader approval of abiraterone for early-stage use, especially in the UK where its application has been limited. With this test, the decision becomes more scientific, more ethical, and far more personal.

Read more here:

https://www.theguardian.com/society/2025/may/30/new-ai-test-can-predict-which-men-will-benefit-from-prostate-cancer-drug

DeepMind and Isomorphic Labs have just open-sourced the code for AlphaFold 3, the latest and most powerful version of their AI system for predicting biomolecular structures. Unlike its predecessor, AlphaFold 3 can model not only proteins but also DNA, RNA, and small molecule interactions. This release is set to transform the fields of structural biology, drug discovery, and synthetic biology.

The newly released AlphaFold 3 pipeline allows scientists and engineers to predict complex molecular assemblies with unprecedented accuracy. This includes protein-ligand binding, nucleic acid interactions, and even multi-protein complexes — tasks that were previously the domain of expensive wet-lab experimentation.

What makes AlphaFold 3 particularly exciting for developers is the modular and Python-accessible architecture. Bioinformaticians, AI researchers, and biotech engineers can now directly integrate these models into their workflows, accelerating everything from target identification to molecule design. The fusion of machine learning and biology has never been more seamless.

As AI continues to unlock new frontiers in life sciences, AlphaFold 3 sets the stage for a future where protein design and therapeutic discovery can be driven by open, programmable, and intelligent systems.

https://github.com/deepmind/alphafold

https://www.deepmind.com/blog/alphafold-3-predicting-the-shape-and-interactions-of-everything-protein

NVIDIA has just announced native support for Python in its CUDA platform, marking a major shift in how developers can access GPU acceleration. For the first time, Python developers can write CUDA programs without needing to rely on C or C++ bindings. This native integration drastically lowers the barrier for using GPUs in scientific computing, AI, and data-heavy Python applications.

The new cuda-python package gives direct access to CUDA’s driver and runtime APIs, letting users launch kernels, manage memory, and control streams entirely from Python. It also includes support for just-in-time (JIT) compilation, which means you can write dynamic GPU code directly in Python, compile it on the fly, and run it immediately.

A key innovation is the new CuTile programming model, which brings a tile-based structure to CUDA operations. CuTile is designed to feel natural to Python users familiar with NumPy and CuPy, and it allows for efficient manipulation of large data arrays without needing to manage threads manually.

This move brings CUDA closer to Python’s ecosystem and could reshape how GPU computing is taught, deployed, and scaled across AI and HPC workloads.

Read the full announcement here:

https://thenewstack.io/nvidia-finally-adds-native-python-support-to-cuda/

Official documentation from NVIDIA:

https://developer.nvidia.com/cuda-python

A UK-based biotech company, Basecamp Research, is using artificial intelligence and environmental DNA to uncover a vast new world of biology. By sampling DNA from some of the planet’s most remote and untouched ecosystems, the company has identified over one million previously unknown species. These discoveries are not just academic. They are fueling a next-generation AI platform designed to radically accelerate drug discovery.

The company’s genomic database is already among the largest of its kind. It is being used to train AI models that predict protein structures, functions, and interactions at unprecedented accuracy. This includes boosting tools like AlphaFold, which helps researchers visualize the shape of proteins based on genetic code alone.

A major focus of the project is the identification of new large serine recombinases, a type of enzyme that can precisely insert large DNA sequences into the genome. These enzymes are considered highly promising for future gene therapies, including for cancer and rare genetic disorders.

Basecamp Research is also trying to set a new standard in ethical science. The team works directly with local researchers and governments in the regions where DNA samples are collected. In return, partner countries receive royalties and scientific credit, avoiding the extractive models of earlier biotech ventures.

This combination of field biology, advanced sequencing, and deep learning could reshape how we find drugs, understand evolution, and build genetic tools. It is a powerful example of how AI, when paired with real-world data, can help uncover the deepest layers of life on Earth.

Read the full story here:

https://www.ft.com/content/9765ab86-0156-4901-b6ec-fbee465ab819

Meta is moving to automate the majority of its internal risk assessments using artificial intelligence. Until now, most updates to Facebook, Instagram, and WhatsApp underwent manual reviews to ensure they would not harm users or violate laws. That is about to change.

According to a report published by TechCrunch, Meta plans to have AI handle up to 90 percent of these assessments. The company says this will speed up product development while still keeping humans in the loop for sensitive or unfamiliar issues.

This shift is part of a wider push inside Meta to cut costs and streamline decision making across its platforms. Engineers will soon be asked to input risk-related information into a system that scores each update and determines whether human review is needed. If not, the feature could go live automatically.

Critics argue that AI cannot yet fully understand complex social or political risks, especially those tied to misinformation, privacy, or cultural harm. Meta’s executives have acknowledged that risk automation will be a gradual process, and that human judgment will remain essential for edge cases.

Full story: https://techcrunch.com/2025/05/31/meta-plans-to-automate-many-of-its-product-risk-assessments/

Alternate source: https://economictimes.indiatimes.com/tech/artificial-intelligence/meta-to-handover-most-of-product-risk-assessments-to-ai/articleshow/121541117.cms

A new study from Google DeepMind is pushing the boundaries of how artificial intelligence can understand biology. Scientists have trained a model called AlphaFold3 that does not just predict protein structure like its predecessor, but can now also simulate how proteins interact with DNA, RNA, and small molecules. This brings researchers a step closer to modeling complex molecular machinery inside cells.

The innovation lies in how AlphaFold3 combines physical principles with deep learning, using a single unified architecture to simulate biomolecular complexes. Instead of relying on existing templates, the model generates predictions from raw sequence data. This means it can handle previously unseen combinations of biological components, giving scientists the ability to explore novel interactions that were hard or impossible to study before.

The model has already demonstrated major improvements in accuracy over existing tools. It predicts not just the shapes of individual proteins but also how they bind to other molecules, which is essential for understanding how cells work and how diseases begin. This is also critical for designing new drugs that target specific biological pathways.

While the full version of AlphaFold3 is not yet publicly released, DeepMind has launched a service that allows researchers to use the tool for non commercial purposes. This opens the door to new research in structural biology, drug discovery, and synthetic biology, where understanding how biological parts fit and move together is key.

By giving scientists access to predictions about protein interactions on a scale never before possible, this new AI model could accelerate discoveries across medicine and biotechnology.

A new deep learning model called MutSig AtlasAI is changing how we understand cancer. Built by researchers at the Broad Institute and MIT, it looks at tumor DNA from thousands of patients and learns to tell which mutations actually help the tumor grow, the so called drivers, and which ones are harmless.

Most mutations in cancer are just along for the ride. The hard part is figuring out which ones matter. This tool makes that easier. It uses not just how often a mutation appears, but also where it is in the gene, how it affects protein structure, and how evolution has treated that part of the genome. It also works across different cancer types, including rare and understudied ones.

In tests, it found new driver mutations that older methods missed, especially in cancers like bile duct and soft tissue sarcomas. The model is open source and ready for other scientists to build on, making it a potentially powerful step forward in precision cancer treatment.

https://www.broadinstitute.org/news/machine-learning-tool-pinpoints-cancer-driver-mutations

In a significant advancement for prostate cancer care, researchers from the US, UK, and Switzerland have developed an artificial intelligence (AI) tool capable of predicting which men with high-risk, non-metastatic prostate cancer will benefit from the drug abiraterone. This development, unveiled at the American Society of Clinical Oncology’s annual meeting, promises to tailor treatment plans more effectively, potentially improving outcomes and reducing unnecessary side effects.

Abiraterone, often described as a “gamechanger” for advanced prostate cancer, has been shown to significantly reduce the risk of death. However, its use in earlier stages of the disease has been limited due to potential side effects, including increased risks of heart issues and diabetes. The new AI tool analyzes biopsy images to identify a biomarker indicating likely benefit from abiraterone. In a study involving over 1,000 men, the AI identified that approximately 25% had biomarker-positive tumors. For these individuals, abiraterone halved the five-year mortality rate from 17% to 9%. Conversely, for those without the biomarker, the benefit was negligible.

This breakthrough could lead to more personalized treatment strategies, ensuring that patients receive therapies most likely to benefit them while avoiding unnecessary exposure to potential side effects. It also has implications for healthcare systems, potentially prompting revisions in treatment guidelines and drug approval policies.

Experts have lauded the development, emphasizing its potential to optimize personalized cancer care and improve patient outcomes.

For more details, read the full article here:

https://www.theguardian.com/society/2025/may/30/new-ai-test-can-predict-which-men-will-benefit-from-prostate-cancer-drug

In a major step forward for synthetic biology, scientists at the Center for Genomic Regulation in Barcelona have used artificial intelligence to design brand-new DNA sequences that can control gene activity in living mammalian cells. This study, published on May 8, 2025, shows that AI can create synthetic regulatory elements—small pieces of DNA that decide when and where a gene is turned on.

The AI model developed by the team doesn’t rely on copying natural DNA. Instead, it generates new sequences from scratch that are customized to work in specific cell types. For example, the model can design a sequence that activates a gene only in stem cells that are on their way to becoming red blood cells, while leaving other cells alone.

To test their work, the researchers built these sequences in the lab and inserted them into mouse blood cells. The results confirmed that the AI-made DNA successfully activated the right genes in the right cells. That kind of control is incredibly valuable for future treatments where you want to change gene activity only in a targeted group of cells.

This kind of precision could lead to safer and more effective gene therapies, where unwanted side effects are avoided because the treatment only works where it’s needed.

AI is now helping us go beyond what nature provides, letting scientists write DNA code that behaves exactly how we want it to. That changes the game for genetics, medicine, and our ability to design life at the molecular level.

https://www.sciencedaily.com/releases/2025/05/250508112324.htm