How AI Supports Scientific Discovery

A materials laboratory wants a coating that performs under a specified temperature condition. An AI system can propose many candidates, while the team has limited time to prepare samples and use its instruments. Which work should the system help change, and what would count as progress? This is Lesson twelve, How AI Supports Scientific Discovery. We will follow this fictional investigation from choosing a test through using its measurements, then examine the resources and decisions that turn suggestions into useful scientific evidence. Start with the research question.

The scientists define the property they want to investigate and the conditions under which it matters. The model ranks possible candidates using the information and method supplied to it. That ranking can change which samples receive scarce preparation and instrument time. Domain knowledge helps the team decide whether the suggested candidates and comparison are suitable for the question. A poor ranking could direct that effort away from something important. Follow one proposed coating to the next stage.

A technician must be able to prepare the sample, and the team needs a suitable measurement and comparison. Those steps can expose information missing from the model's proposal. Suppose a measurement conflicts with the predicted performance. The researchers can investigate the mismatch and revise what they test next. In this teaching case, the useful contribution is a more focused investigation that can reject an unsuitable proposal. This pathway gives different computational activities their place.

Prediction estimates a property or arrangement. Search helps explore possible candidates. Simulation, experiment design and analysis can support other parts of the investigation. Choose the activity around the scientific question and inspect how its result changes the next action. More generated candidates can increase the demand for screening if preparation and measurement remain the limiting steps. A documented prediction example comes from the 2024 AlphaFold Three research.

It reports predictions of biomolecular structures and interactions evaluated on specified benchmarks. A biomolecular structure describes how relevant molecules are arranged. A model that proposes an arrangement can give scientists something useful to investigate. The evaluation's reference cases and conditions determine what its reported performance supports; a new research application still needs evidence appropriate to its question. Now consider what happens when observations influence the proposed explanation.

A protein can adopt more than one arrangement, and a single predicted picture may leave measured behavior unresolved. Scientists can ask which possible arrangements are compatible with their measurements. The observations then help constrain the candidates, giving the prediction and experiment connected roles within the investigation. In a June 2026 Nature Biotechnology paper, Advaith Maddipatla and colleagues guided AlphaFold Three using experimental data to generate sets of possible protein structures. On their tests, the guided sets agreed more closely with measured constraints than unguided predictions,

and the authors identified alternative conformations in previously unmodeled electron density. In ordinary language, measurements helped identify several arrangements compatible with the evidence. The resulting explanation remains tied to the measurements and evaluation conditions examined in that work. For our coating team, the general question is how evidence changes the next proposal. It could compare measured performance with the ranking, inspect a failure and choose a revised candidate or research question.

This is a proposed investigation sequence, separate from the documented protein method. Scientists would select methods appropriate to coatings. Keep the model's version, candidate, test conditions, comparisons and results together so another researcher can inspect how the conclusion follows. Suppose the project record lists ten candidates, three prepared samples and one measured result. Those fictional counts describe different stages. They tell the team where work has reached and where screening remains.

The research conclusion depends on what the measurement establishes, how reliable the comparison is and which alternatives were examined. A later commercial product would require further evidence for its intended use. Count the stage that matters to the claim being made. The laboratory director now has a resource decision. Should it generate more candidates, expand preparation capacity or improve the measurements supporting the current question? Each option uses computing, staff and equipment differently.

Technicians and reviewers may receive more work as proposal generation becomes easier. The director should examine where useful evidence is constrained and how an investment would change that constraint, including the effort required to maintain reproducible records. Access to those resources also shapes participation. A smaller team may be able to generate a promising candidate but lack an instrument, relevant data or specialist support to investigate it. Shared facilities and usable datasets could widen the work it can attempt.

Restrictive access or unsupported tools could leave it unable to proceed. Funders and institutions choose which questions, people and infrastructure receive support. A model account is one part of that arrangement; participation depends on the connected research process. Ask what wider value has actually been demonstrated. A benchmark can establish performance on its specified evaluation. A measurement can support a conclusion about a particular sample and condition.

Replication, further experiments and domain judgment help determine what can be relied on more broadly. The record should make the current finding and remaining questions visible. This lets collaborators decide what to test or fund next without translating an early result into an unsupported product promise. Try the Learning Check. Why distinguish prediction from validation? Explain what a model proposes or estimates

and what evidence the scientific claim still requires. Pause and write your answer. A model proposes or estimates a result. The scientific claim needs appropriate evidence and testing under the field's methods. Identify the claim, evaluation conditions and remaining uncertainty. In the coating investigation, a ranking helps choose candidates; suitable preparation, measurements and comparisons establish what can be said about their performance. The validation must match the conclusion being drawn.

Return to the laboratory's limited testing time. It can examine whether the ranking improves candidate selection, whether measurements change the next investigation and where preparation or analysis needs support. The director can compare investments against the evidence the team is trying to obtain. You can now trace AI's contribution through the investigation and identify the people and resources needed to turn a suggestion into a supported scientific result. For Practice, use the different fictional enzyme-screening team in your worksheet.

Separate proposed candidates, prepared samples, measurements and the broader intended use. Write supported claims for the supplied stages and identify missing evidence and a possible screening bottleneck. Complete the separately identified scientific-workflow exercise to examine how observations can change a proposed explanation and what resources make participation possible. Make your attempts before using the separate key. The next lesson examines who can question, change or stop a system's actions.

Scientific work has shown how evidence and expertise shape a decision; we will now follow the authority, usable alternatives and accountability that make human choice practical in other settings.