What is AI for Science?

Most clinical AI news is a benchmark. Before you trust a model near patients or trial design, look at how it was tested and on what.

Also in the toolbox

• Bedside Bench

  Doximity's clinical AI benchmark: 500 questions written by physicians, across 10 parts of clinical care. Source

• AutoMedBench

  Tests medical AI agents at every step (planning, setup, validation) as well as the final answer. 6,300+ runs, 7 tracks. Source

• MONAI Physio

  An open-source toolkit that turns 3D and 4D medical images into heart and lung digital twins. Source

What happened lately

• ICON builds Claude into clinical trials

  A large CRO plans to use it for site selection, enrollment forecasting and protocol design. Per a third-party post. Source

This is where most of the action is this year: design tools that hand designs to a wet lab and get data back, and models built for biology.

Also in the toolbox

• Codex with the Adaptyv plugin

  Design peptides in Codex, then send them to Adaptyv's lab for testing. Source

• Rosalind Workbench

  A workbench for biologists, built into Codex. Source

• Anthropic Life Sciences Verification Program

  Verified labs and companies can apply for access to Anthropic's most capable models for biology and drug development work. Source

What happened lately

• Claude designs protein binders with up to 35% hit rates

  With Adaptyv Bio and Twist Bioscience, binders for 14 of 15 targets. A human expert wrote the 30k-token prompt. Source

• Antibodies designed by an AI harness, then tested

  Will Hua's opendde-harness designs antibodies from a plain request. The designs went to the lab. Source

• Claude finds an unknown enzyme system in phage DNA

  It sits next to a CRISPR-like repeat array. Anthropic says what it does is not yet known. Source

Reading, checking, coding and reproducing. These are the tasks every lab has, and where a new tool can save the most hours this week.

Also in the toolbox

• Paper Scorecard

  Finds the citing papers that actually engage with a paper and shows their critiques. Source

• Lazarus

  An agent that gets dead research code running again, starting from a GitHub link. Source

• PaperBanana (Google)

  Turns methodology text into figures for papers. Source

What happened lately

• A model finds one experiment run 1,400 times

  Pointed at 40 years of ribosome methods sections, it flagged results only one lab had ever reported. Per a third-party post. Source

• Claude makes 30 open bio models 4x faster

  Anthropic open sourced the optimization code. Cheaper runs on the models your team already uses. Source

• Ginkgo builds four autonomous labs at universities

  At MIT, Caltech, Northwestern and Maryland. Source

A short chapter for now. Most of what we saved this year is biology. Weather and chemistry are where the useful tools are.

Also in the toolbox

• MagNET on Rowan

  An ML model that predicts 1H and 13C NMR shieldings, a drop-in for DFT in NMR workflows. Built with groups at Merck, MIT, Harvard and CSU. Source

Shared data and open infrastructure decide what your models can learn from. Some of it is growing fast, and some of it is going away.

Also in the toolbox

• The full arXiv corpus

  All 3,148,796 arXiv papers, every version, 16 TB, as a Hugging Face dataset. Source

• Neuroscience dataset browser

  Search 10,000+ open neuroscience datasets across EEG, MRI, MEG, NIRS, PET and more, from about 1,000 institutions. Source

What happened lately

• A $1.8B push for open, AI-ready biology data

  Biohub, DOE, NIH, Meta, Google DeepMind and Isomorphic Labs. The data layer your models will train on. Source

• OSF Projects to go read-only in February 2027

  Preprints and preregistration stay. If your lab keeps materials on OSF, plan the move. Source