Back to home

AI Scientist research paradigm

Parallel hypotheses, delivered only after review

The agent reads the literature and makes a plan, then sends several sub-agents to test different hypotheses at once, sharing findings across branches. An independent review agent checks every conclusion.
Start researching

Agent algorithm accepted at ICML 2026 · 2.2k GitHub stars

Research process
3 branches · 1 running
  • Does free calmodulin already hold compact, binding-like conformations?SupportedPRE NMR shows about 5–10% compact states, some close to the peptide-bound form.
  • Does a more flexible central linker mean stronger target recognition?RefutedAffinity peaks with one inserted glycine and drops when the linker is made longer or shorter.
  • Is the linker always locked once a target binds?Running
Shared across branches The affinity data were measured at low salt and pH 5.5; other branches were told not to extrapolate directly.

How it works

Literature first, then a plan

It searches, screens and reads full texts, breaks the question into testable hypotheses and writes a plan before running anything, rerouting on its own when a source is blocked.

Research plan
Search three databases
Deduplicate and screen
Read full texts
Write the review
Reroute PubMed blocked → Europe PMC

Independent review makes sure it is right

Before delivery, an independent review agent checks the methods, code and conclusions. Problems go back for rework, and the next round starts only after review passes.

Review · Lysozyme molecular dynamics
Before delivery, a review agent checks the data and conclusions against the research goal.
The RMSD curve jumps unphysically: atoms of the reference structure cross the periodic boundary and the molecule is split.
Make molecules whole with the topology connectivity, center, re-align and recompute RMSD.
Passed: backbone RMSD averages 0.08 nm and agrees with the radius of gyration. Ready to deliver.
$ gmx trjconv -s em.tpr -f md.xtc -pbc mol -center
$ gmx rms -s em_reference.tpr -f md_whole.xtc

A report with citations

Every claim cites its source, and the code, data and figures stay in the project so you can rerun them anytime.

review.md

Calmodulin linker flexibility and target recognition

A single inserted glycine maximized both interlobe effective concentration and peptide affinity.[1]

[1] Anthis & Clore, JACS 2013

review.mdreferences.json

Start from a research goal

Explore next