AI has handed humanity the most powerful thinking tool it has ever had — one that speeds up every field it touches, and science most of all. The next generation of researchers won’t be the ones with the biggest lab or the longest tenure. They’ll be the ones who put the best tools of the day to work — and today the best tool is AI.
That tool is loud. AI now throws off hypotheses by the thousand — every model a fresh pile of molecules to make, structures to fold, targets to hit. And the money has followed the noise: companies are pouring capital into agent after agent to discover faster, test faster, screen faster — burning compute and time on data and hypotheses nobody has checked. The faster we generate, the more we build on ground no one has tested.
NewScience is the answer to that.
Introducing NewScience
NewScience is an AI research copilot for biology — it finds and reads the literature for you, benchmarks it by reproducibility instead of citation count, and, when you need certainty, verifies the science behind the number through systemized dry execution. Not for one domain alone: molecular docking, protein folding, molecular dynamics, genomics — NewScience hot-swaps the right simulation engine in behind each, so the full loop from paper to signed, verified result runs on one stack instead of scattered across a dozen. It operates over the world’s public literature, not data of its own — so the score it returns is neutral ground a scientist, a partner, and a regulator can all stand on.
Under the hood, it runs as an agent swarm. Users interact with a generalist coordinating agent — one wired to over 60 curated skills and connectors, pre-configured for genomics, single-cell, proteomics, structural biology, cheminformatics, and more. That agent spins up others on demand, and taps the specialist agents researchers build for themselves. And a reviewer agent stays in the loop the whole time — auditing citations and calculations, flagging errors and correcting them.
We’re opening this up. Beta testing for the platform — along with MCP and API access, so any lab can wire NewScience into its own pipelines — is launching soon, for every researcher and scientist who wants in. Register for the waitlist.
How it works
Think of NewScience as an AI workbench — built so a scientist can get to a trustworthy answer without reading fifty papers to find the three that matter. Here is what happens when you ask it something.
It gathers. The literature is scattered — PubMed, Europe PMC, preprint servers, structure and chemical databases, each with its own shape and its own quirks. NewScience reaches across all of them at once and pulls the work that speaks to your question into one place.
It reads down to the parameters. A number in a docking paper isn’t a fact — it’s the last line of a recipe. It depends on the receptor, the ligand, the software and its version, the search box, the random seed. NewScience takes those parameters out of each paper — sixteen of them for docking — and pins every one to the paper’s own words. If the sentence isn’t there, the parameter isn’t there. It reports what the paper says, down to the line, not what a model guessed it meant.
It ranks by what will hold. Some of those parameters carry the weight — without the search box or the ligand, there’s no way back to the number; others just force a guess. NewScience grades every paper on how completely it reports itself, from missing-the-basics to fully reproducible, and floats the ones that will actually hold up to the top. Not the most cited — the soundest.
It proves the number. When it matters, NewScience does the work behind the result — it pins every molecule to a fixed chemical identity, sends the job to the right simulation tool, and measures that tool’s own margin of error first, so a match means something. Out comes an output you can see for yourself, with a receipt that traces to a real run on real hardware. “The paper says the score is X” becomes “here is what the simulation gave back, and here is how close.”
An early use case of NewScience
We pointed the first version of this at one of the messiest corners of the recent literature — the flood of molecular docking studies aimed at the SARS-CoV-2 main protease, the enzyme half the world went hunting for a drug against.
We gathered 236 open-access docking papers — around 12,466 individual ligand results — and audited every one against a sixteen-field reporting standard we built for the job. The picture was bleak. Only 8% of the papers reported enough to run their docking from the methods alone. Nearly half were blocked outright by a single missing parameter. Not one of the 236 was fully reproducible. The search-box centre — the setting that decides where on the protein a molecule even docks — was missing from two-thirds of them. The random seed, the cheapest line a scientist can write down and the one thing that lets a result reproduce exactly, showed up in exactly one paper.
Then we checked whether our grading meant anything. We took 37 quality-controlled results across two protein structures, pinned every molecule to its exact chemical identity, and ran the docking ourselves with each paper’s own settings — after first measuring AutoDock Vina’s own run-to-run wobble, which turned out to be almost nothing. The reported numbers came back to within half a kcal/mol on average. And the grade predicted the gap: the cleanly-reported results landed right on the engine’s own floor, while the ones missing parameters drifted more than twice as far. How well a paper writes itself down tells you whether its number will hold.
One finding underneath the rest is the reason NewScience is built the way it is. When we let an AI read a paper and grade it holistically — by feel — it agreed with expert reviewers barely better than a coin flip. A plain set of rules reading the same pinned parameters, with a human signing the final call, matched two independent experts almost exactly. So NewScience runs on rules over evidence you can trace — never on a model’s gut.
Contribute to scientific progress
Biology is the field every one of us here has thrown in with — the one we’ve committed to accelerating. And we’re not doing it alone. We want to put the best tools going into the hands of the next generation of scientists and researchers, gather them into a community, and push the limit past where any single lab can reach.
The first version is launching soon. Stay tuned.