AI for Science

AI for Science applies models to scientific data, prediction and candidate selection. Google DeepMind describes research across areas including proteins and materials.^[1] AI-driven scientific discovery is a broad phrase. A model-generated proposal and an experimentally established discovery remain different kinds of evidence.

Predicting an unobserved value, selecting the next candidate and explaining a mechanism are separate goals. Strong prediction performance alone does not establish causation or complete automation of research. This distinction is ORBiS Tree's reading framework rather than a claim about one universal scientific workflow.

AlphaFold is a prominent example of protein-structure prediction and modeling related molecular interactions.^[2] A predicted structure can inform further research without establishing function or effects in every environment. Inspect the inputs, evaluation conditions and unresolved cases before extending a claim.

This case is hypothetical. Near-duplicate samples appear in both training and evaluation data. The resulting score may overstate performance on genuinely new conditions. Check overlap according to sample families, collection sources and time, then report performance separately on new conditions. Benchmark Data Contamination and Data Provenance explain relevant evidence boundaries.

An automated AI researcher may retrieve literature, propose candidates or manage analysis; those scopes are not interchangeable. Record which steps were automated, which decisions people made and whether the result was reproduced. Preserve dataset versions, preprocessing, evaluation splits and failures. Datasheets for Datasets supports data documentation, while AI Agent Oversight addresses action permissions and review points.

Google DeepMind's science overview ↗ presents applications across scientific areas.^[1] Identify whether AI collected material, predicted a value, selected candidates, assisted observation or drafted interpretation. An agent that gathers literature has not thereby independently reproduced an experiment.

AlphaFold's materials ↗ provide background on structure and interaction prediction.^[2] The examples below explain general prediction evaluation, not protein experimentation procedures. Keep predictions, uncertainty information and observations as distinct records.

Suppose a hypothetical dataset has 120 samples, with 90 used for training and 30 for evaluation. If evaluation samples are close variants from the same manufacturing batches as training samples, a random-split score may not establish performance on a new manufacturing condition. Inspect grouping by batch, sample family and time against the intended generalization target.

For hypothetical predictions of 10, 20 and 30 and observed values of 12, 18 and 33, absolute errors are two, two and three. Mean absolute error is 7/3, approximately 2.33. Without units and scale, 2.33 is difficult to compare. Inspect large errors in important conditions separately from the average. These numbers are not a research result or model benchmark.

Connect selection reasons, dataset versions, preprocessing, model settings, evaluation splits and validation methods. Excluding failed candidates can leave a misleading collection of successes. Generated explanations should identify testable hypotheses rather than establish unobserved causes.

Evaluation contamination addresses train-test overlap; provenance addresses originals and transformations; datasheets describe population and intended use. Capabilities of a foundation model are not interchangeable with validation of one scientific task.

If an automated AI researcher drafts a paper, check that its sources, calculations and supporting records can be reopened. A claim of discovery requires the field's separate validation and reproducibility conditions. Successful tool execution, a finished narrative and an established scientific claim are different outcomes. Agent oversight and model evaluation scope help readers inspect the boundaries of automation.