[01] bioRxiv 3h ago preprint
☆ A generative model designs functional enzymes for novel reactions 1 What happened? The team used AI to design new enzymes for a target reaction and selected promising candidates for experiments. +
2 What problem are they solving? Traditional enzyme engineering usually starts from natural proteins, which narrows the search space and requires many experiments. +
3 How did they do it? The model generated many sequences, then narrowed them down with structure prediction and energy calculations. +
4 What were the results? Some candidates performed well in computational tests, but the experimental evidence is still limited. +
5 Why should I know about it? If broader experiments confirm the method, designing enzymes for new reactions could become much faster. +
⚠ Caution Most evidence still comes from computational benchmarks. Only a small number of candidates received wet-lab validation, so broad applicability has not been established.
Institutions Stanford University European Molecular Biology Laboratory
Authors MC Dr. Maya ChenStanford University LM Dr. Leo MartinEuropean Molecular Biology Laboratory
Sources ↗ Original paper arXiv GitHub Lab blog
Too complex Too simple More like this Less like this More examples
Read full story → [02] Nature Biotechnology 6h ago research article
☆ A single-cell foundation model predicts unseen perturbation responses 1 What happened? Researchers trained a cell model to predict how a cell responds when a gene is perturbed. +
2 What problem are they solving? Experiments cannot exhaustively test every combination of gene, cell type, dose, and time point. +
3 How did they do it? It placed data from different labs in a shared representation space and learned changes before and after intervention. +
4 What were the results? Predictions were stronger for familiar cells and perturbations, but performance dropped in genuinely novel biological settings. +
5 Why should I know about it? It shows how AI could help decide which experiment is most valuable to run next. +
⚠ Caution Cross-laboratory generalization remains unstable, and model predictions cannot replace real perturbation experiments.
Institutions Wellcome Sanger Institute
Authors ER Dr. Elena RossiWellcome Sanger Institute
Sources ↗ Original paper PubMed OpenAlex Code
Too complex Too simple More like this Less like this More examples
Read full story → Multimodal biological models learn transferable disease representations 1 What happened? A new model jointly learns from tissue images, gene expression, and medical text. +
2 What problem are they solving? Each biological data type reveals only part of a disease, and combining them is often difficult. +
3 How did they do it? The model first processes each data type separately, then maps them into a common space. +
4 What were the results? It improved results on several public datasets, but has not shown that it improves real clinical decisions. +
5 Why should I know about it? It helps show whether multimodal foundation models are getting closer to practical use. +
⚠ Caution All results are retrospective. There is no prospective clinical validation, and fairness across populations was not fully evaluated.
Institutions Aster Therapeutics
Authors NW Noah WilliamsAster Therapeutics
Sources ↗ Original paper arXiv Hugging Face
Too complex Too simple More like this Less like this More examples
Read full story → [04] Cell Systems yesterday perspective Explore
☆ What reproducible evaluation should mean for AI-enabled biology 1 What happened? A group of researchers proposed a reporting checklist for evaluating AI models in biology. +
2 What problem are they solving? Many papers report strong scores but make it hard to tell whether the model will work in a new experiment. +
3 How did they do it? The authors compared existing benchmarks and proposed a common disclosure template. +
4 What were the results? This is a perspective, not a standard that the entire field has already adopted. +
5 Why should I know about it? Understanding evaluation design helps you distinguish genuine progress from an attractive benchmark result. +
Institutions European Molecular Biology Laboratory Wellcome Sanger Institute
Authors LM Dr. Leo MartinEuropean Molecular Biology Laboratory ER Dr. Elena RossiWellcome Sanger Institute
Sources ↗ Original article PubMed Community discussion
Too complex Too simple More like this Less like this More examples
Read full story →