Most AI tools give you a number. A confidence score, a ranking, a probability. What they rarely give you is a reason.

That's a problem in scientific research. When a prediction shapes a million-dollar experimental decision, "the model said so" isn't good enough for your PI, your institution, or your own scientific standards.

We built our explainability layer- based on SHAP (SHapley Additive exPlanations) attribution - to answer one question: why did Atomis surface this candidate?

How it works

Every prediction Atomis generates includes a feature-level attribution report. It shows which input properties drove the prediction, how much weight each contributed, and where the model's confidence drops off. The output isn't a black box summary — it's a traceable chain of reasoning your team can interrogate, challenge, and cite.

"Our scientists needed to defend every AI recommendation to an internal review board. The explainability report became the document we submitted. That wasn't something we expected to be possible."

- Jennifer taylor, Head of Computational Chemistry, European pharma partner

Why we're publishing the methodology openly

We want every research team evaluating AI-assisted discovery to be able to verify our claims. The full SHAP attribution methodology, validation dataset, and benchmark results are published in Nature Machine Intelligence (2023) and available to download below — no sign-up, no paywall.

If you find a flaw, tell us. That's how science works.