Abstract

Scientific AI tools face a credibility problem. Predictions are generated at scale - but the reasoning behind individual outputs is rarely surfaced in a form that satisfies scientific scrutiny. This paper presents Atomis's full explainability architecture: a SHAP-based attribution layer built on top of our graph neural network prediction system, validated across 1,800 candidate predictions by independent research teams.

We document both what the system gets right and - critically - the categories of prediction where attribution confidence degrades. Failure cases are included in full.

Key findings:

SHAP attribution matched independent scientific reasoning in 91% of reviewed cases across drug, protein, and catalyst domains

• Explainability outputs reduced internal review cycles by an average of 3.2 weeks across 8 partner programs

• Three identified failure categories - high-novelty candidates, multi-target optimisation edge cases, and sparse training data domains - documented with recommended workarounds

• Full methodology, training data composition, and validation set available in supplementary materials