John Squires, the recently anointed director of the U.S. Patent and Trademark Office, has determined that a machine learning (ML) patent application met the standard for patent subject matter eligibility, an outcome that seems to bode well for ML-based patent applications going forward.

A team of U.S. and South Korean researchers have developed an AI model called MSI-SEER that can not only predict microsatellite instability-high (MSI-H) tumors based on tissue slides, but also flag “what it does not know.” “Have you ever asked ChatGPT anything, and the response was, ‘I don’t know?’” Cheong Jae-ho asked during an interview with BioWorld. “Probably not, and that is the problem with AI now.”

Pharma companies are collaborating to boost the power of artificial intelligence (AI) in drug discovery by allowing access to proprietary structural data to train a large language model. Each of the partners is contributing data from several thousand experimentally determined protein:ligand interactions, creating one of the most diverse datasets and the richest chemistry assembled to date for model training.

A new generative AI model trained on UK Biobank data can simultaneously forecast the risks and timing of developing over 1,000 different diseases a decade into the future. The developers applied similar algorithmic concepts to those used to develop large language models like ChatGPT and Gemini to build the model, using medical records from 402,799 participants in UK Biobank.

Researchers at the Massachusetts Institute of Technology have developed a generative AI model that was able to generate novel antibiotic structures from either chemical fragments or de novo, starting from ammonia, methane, water or no starting point at all. In a study that was published online in Cell, the team tested two dozen of more than 10 million structures that were proposed as potential antibiotics by the model.

Deep learning tools for protein design can also be used to create molecules that bind to them. Certain peptides, such as intrinsically disordered proteins (IDPs), are challenging to target due to their variable nature. However, scientists from the lab of Nobel laureate David Baker have developed a method to generate binders for IDPs by searching the world’s largest protein database with their AI-powered tool RFdiffusion.

Rakovina Therapeutics Inc. has reported progress in its AI-driven KT-5000AI program, advancing the development of precision ATR (ataxia telangiectasia and Rad3-related) inhibitors designed to disrupt the DNA damage response (DDR) pathway in cancer cells. Through its collaboration with Variational AI Inc., Rakovina has evaluated a vast chemical space of potential molecular structures using Variational’s AI Enki platform to identify novel compounds

Qanatpharma AG (QP), Zuse Institute Berlin (ZIB), Enamine Ltd. and Proteros biostructures GmbH have announced the launch of a research collaboration to accelerate the discovery of novel therapeutics targeting cerebral perfusion deficits associated with subarachnoid hemorrhage (SAH).

Chugai Pharmaceutical Co. Ltd. and Gero Pte Ltd. have entered into a joint research and license agreement to develop novel therapies for age-related diseases. Chugai will create novel antibody-drug candidates for new drug targets discovered by Gero using its AI target discovery platform.

Researchers at the Massachusetts Institute of Technology and Recursion Pharmaceuticals Inc. have released an open-source AI model that can predict the binding strength of small molecules as well as structures of proteins and biomolecular complexes. The model, which is called Boltz-2 and was released by the research team on the developer platform Github on June 6, addresses a major bottleneck in drug discovery with its improved ability to predict binding strengths.