Search articles by title, topic or section.
Khabre desk
Recent reporting, analysis and explainers from the machine-learning desk.
health
A medRxiv preprint reports that muFM, a model pretrained on more than 1.7 million unlabeled micro-ultrasound images, improved performance on three prostate image-analysis tasks. Gains were largest when the model was fine-tuned with limited annotated data.
science
A bioRxiv preprint describes four rounds of computational screening that identified natural hydroxymethylfurfural oxidases with improved activity on a key FDCA-production reaction. Some candidates approached an engineered benchmark under the tested conditions.
science
A bioRxiv preprint describes a transformer model trained on 8.8 million measured SARS-CoV-2 RBD–ACE2 interactions. It predicts receptor-binding compatibility across animal species and pairs predictions with uncertainty estimates to help prioritise surveillance.
health
A medRxiv preprint reports that a RoBERTa model classified depression-related Twitter users with 0.822 accuracy on an 80/20 training–testing split. The study also used GPT-4o to identify themes in posts, including societal pressure and low self-esteem.
science
A bioRxiv preprint introduces PLABench, a benchmark comparing nine deep-learning methods for predicting protein–ligand binding affinity. Its authors report that strong overall accuracy did not translate into consistent performance across targets or protein families.
science
OpenAI says its most advanced AI model produced a mathematical result involving a singularity in the Navier–Stokes equations. The case highlights why continuum fluid models need molecular descriptions at very small scales.
science
A bioRxiv preprint describes a convolutional neural network that measured Douglas-fir pollen germination and elongation across temperatures from 5 to 40°C. The study found narrow, distinct thermal optima for the two reproductive processes, mostly between 19 and 23°C.
science
Apple researchers describe SimpleDesign, a Transformer-based model that jointly generates protein sequences and three-dimensional structures using a single-stage training approach. The system was trained on more than 2 million sequence–structure pairs, but the supplied evidence reports no experimental validation of its designs.
technology
Figure says its Index app is collecting real-world physical-task videos for training the Helix robotics system, but the company has not published independent evidence that the dataset improves robot performance.
technology
Figure says Creators have uploaded more than 16 million videos through its Index app, but the announcement provides no independent evidence that the data improves robot performance.