Browse Articles | Nature Methods

Benchmark pitfalls expose need for expert-guided spatial clustering

Benchmarking computational tools for cutting-edge technologies such as spatially resolved transcriptomics is challenging owing to the lack of reproducibility, of data and of evaluation strategies that are faithful to multiscale spatial biology. However, we found that complementing computational consensus strategies with feedback from histology and cell biology experts circumvents these limitations to accelerate biological discovery.

The physics of sculpting a living creature

Cell biologists have long known that mechanical forces, such as tension and compression, influence tissue differentiation. Now, advanced methods are letting them observe and manipulate these forces in real time, in living systems.

• Caroline Seydel

Amplifying multiplex spatial profiling with tyramide oligonucleotides

Integrating different modalities and flexibly amplifying signals presents a major challenge for imaging-based spatial biology. We present protein and nucleic acid serial tyramide amplification (PASTA), which combines tyramide oligonucleotide conjugates with diverse assays for horseradish peroxidase readout to enable the seamless, flexible integration and amplification of multiple signals.

Disruption of the research software landscape through AI software generation

Specialized research software has historically been costly and time-consuming to create, but large language models (LLMs) have become capable enough at code generation to fundamentally change this. We describe how LLM-assisted programming disrupts the landscape by allowing researchers to build tools without support from software engineers, illustrate this with an example built rapidly by a single LLM-assisted developer, and discuss opportunities and risks.

• Nelson D. Medina
• Joergen M. R. Kornfeld

Alignment with experimental data improves protein generative modeling

Direct preference optimization (DPO) aligns a pretrained protein language model with experimental stability data, yielding ProteinDPO, a model that scores and generates thermostable protein sequences. Applied to H5N1 influenza hemagglutinin, it achieved large improvements in thermal stability of hemagglutinin while preserving antibody recognition.

Aligning protein-generative models to experimental fitness with ProteinDPO

This Article demonstrates that direct preference optimization (DPO) can be used to effectively align an unsupervised structure-conditioned language model with biophysical information. The aligned model, ProteinDPO, achieves stability prediction competitive with that of task-specific models and consistently outperforms unsupervised and fine-tuned versions of the model.

• Talal Widatalla
• Ashir A. Borah
• Brian L. Hie

Code to plot

Biologists can reuse, remix and vibe code scientific figures.

• Alexander Lex
• Helena Klara Jambor