AI for science

Recent programs emphasize integrated scientific work: agent-assisted scientific software, AI workbenches, multi-step drug-discovery evaluation, and design engines that go beyond structure prediction. The historical through-line runs from expert systems, through machine discovery and inductive logic programming, to robot scientists that select and execute experiments.

The key boundary is epistemic: generating a molecule, proof sketch, hypothesis, or code change is not the same as establishing a scientific result. A credible workflow preserves provenance, compares baselines, represents uncertainty, records negative evidence, and submits the claim to a measurement that could be wrong.

Scientific learning can be framed as choosing an experiment e that is expected to reduce uncertainty about hypothesis H. Expected information gain is

EIG(e) = E_y[ KL(P(H|y,e) || P(H)) ],

where y is the possible result and KL measures how much the posterior differs from the prior.

Worked example: suppose two hypotheses have prior probabilities (0.5,0.5). An experiment has outcomes that are 90 percent likely under H1 and 10 percent under H2. If the positive outcome occurs, Bayes' rule gives posterior P(H1|+) = .9. Entropy falls from 1 bit to about 0.469 bits; the result provides about 0.531 bits of information. A more discriminating experiment has higher expected information gain, but cost and risk must also be considered.

The experiment-selection utility can be U(e)=EIG(e)-λ·cost(e)-ρ·risk(e). This makes the tradeoff explicit; the weights are decisions, not facts supplied by the model.

1. 1969 — DENDRAL. Buchanan, Feigenbaum, Lederberg, Djerassi, and collaborators used domain knowledge and heuristic search to infer molecular structures from mass spectra. CMU archive (opens in a new tab).
2. 1977 — AM. Douglas Lenat created a program that explored mathematical concepts using heuristics. Its apparent discoveries also raised questions about representation and designer-provided bias. IJCAI paper (opens in a new tab).
3. 1995 — Progol and inverse entailment. Stephen Muggleton advanced inductive logic programming for learning relational rules from examples and background knowledge. New Generation Computing (opens in a new tab).
4. 1996 — ILP for structure–activity relationships. King, Muggleton, Srinivasan, and Sternberg applied relational learning to biological activity. PNAS (opens in a new tab).
5. 2004 — Robot Scientist. Ross King and collaborators combined hypothesis formation, experiment selection, laboratory execution, and interpretation. Nature (opens in a new tab).
6. 2009 — Adam. The team reported an autonomous system generating and experimentally testing functional-genomics hypotheses. Science (opens in a new tab).

1. 2017: CheXNet. Pranav Rajpurkar, Jeremy Irvin, Kaylie Zhu, Matthew Lungren, Andrew Ng, and collaborators evaluated pneumonia detection on chest radiographs. The comparison is dataset- and metric-specific and does not establish prospective clinical benefit. Project and paper (opens in a new tab).
2. 2017 | ANI-1 neural potential. Justin Smith, Olexandr Isayev, and Adrian Roitberg trained a neural-network potential on density-functional calculations for organic molecules. The model inherited potential-energy surfaces and supervised approximation, but its accuracy depended on the chemical elements and configurations represented in costly reference data. Chemical Science (opens in a new tab).
3. 2019: CheXpert. Jeremy Irvin, Pranav Rajpurkar, Matthew Lungren, Andrew Ng, and collaborators released a chest-radiograph dataset with explicit uncertainty labels. Dataset quality and retrospective benchmark performance are not clinical validation. Primary source (opens in a new tab).
4. 2021: GFlowNets. Emmanuel Bengio, Moksh Jain, Maksym Korablyov, Doina Precup, and Yoshua Bengio proposed generative flow networks for sampling diverse high-reward candidates. Scientific usefulness still requires domain and experimental validation. Primary source (opens in a new tab).
5. 2021 | AlphaFold2. John Jumper and collaborators combined learned representations, evolutionary information, geometric reasoning, and end-to-end structure prediction. CASP14 performance approached experimental accuracy for many targets, but a predicted structure remained a model of a molecular state, not proof of dynamics, function, or every complex. Nature (opens in a new tab).
6. 2023 | GNoME. Amil Merchant and collaborators coupled graph networks, active learning, and density-functional validation to expand the set of predicted stable inorganic crystals. The work scaled a closed computational discovery loop, but thermodynamic stability alone does not establish synthesizability, kinetic stability, or useful function. Nature (opens in a new tab).
7. 2024 | AlphaFold3. Josh Abramson and collaborators used a diffusion-based architecture to model complexes containing proteins, nucleic acids, small molecules, ions, and modified residues. Broader interaction modeling built on AlphaFold2, while benchmark scope, molecular dynamics, confidence calibration, and experimental validation remained limits. Nature (opens in a new tab).
8. 2025 | AI co-scientist. Juraj Gottweis and collaborators reported a multi-agent system that generates, critiques, ranks, and evolves hypotheses, with selected biomedical validations. The work extends blackboard and scientific-discovery systems, but it was a vendor-affiliated preprint across three biomedical areas with selective rather than broad prospective validation. Primary paper (opens in a new tab).

Objective: reproduce hypothesis selection, experiment choice, observation, and update on a transparent synthetic system.

Environment: simulate bacterial growth from two candidate mechanisms, or use a simple physical law with hidden parameters. Expose a finite menu of experiments with different costs and noise. Keep a hidden ground truth for evaluation.

Procedure: define priors over hypotheses; compute predicted outcome distributions; select experiments using random choice, uncertainty sampling, and expected information gain; sample noisy observations; update by Bayes' rule; stop when posterior confidence or budget threshold is reached. Repeat over 500 simulated worlds.

Metrics: experiments required to identify the true hypothesis, total cost, false-confidence rate, calibration, and robustness when the true mechanism is absent from the hypothesis set.

Expected result: information-gain selection should reduce average experiment count when the model class is correct. All methods may become confidently wrong under model misspecification.

Success criterion: lower cost at matched identification accuracy, plus explicit detection or warning when no hypothesis fits.

Falsification: no advantage over random selection after computation and prior sensitivity are counted.

Safety note: keep the lab simulated. Real biological or chemical experiments require institutional procedures, domain supervision, and applicable safety review.

Claim: workflow-level scientific agents provide more value than isolated predictions because they improve reproducibility and experimental efficiency.

• Baseline: model generates one ranked prediction list.
• Intervention: workflow captures provenance, selects confirmatory experiments, records execution, and updates uncertainty.
• Primary metric: prospectively validated discoveries per unit cost.
• Secondary metrics: time, reproducibility by an independent operator, negative-result capture, and false-confidence rate.
• Uncertainty: prespecified sample size and interval on the difference in validation rate.
• Stopping rule: freeze protocols before prospective evaluation.
• Disconfirming result: no improvement in prospective validation or reproducibility after the workflow's extra cost is counted.

Google DeepMind (opens in a new tab), Isomorphic Labs (opens in a new tab), OpenAI Research (opens in a new tab), Anthropic Research (opens in a new tab), Microsoft Research (opens in a new tab), MIT (opens in a new tab), Allen Institute for AI (opens in a new tab), and relevant university and laboratory groups. Current source artifacts include OpenAI on scientific computing (opens in a new tab), Claude Science (opens in a new tab), DrugDiscoveryBench (opens in a new tab), and IsoDDE (opens in a new tab).

• Medium confidence, 12–36 months: more AI-for-science evaluations will be prospective and workflow-level, reporting failed experiments and human review rather than retrospective prediction alone. Disconfirming signal: retrospective scores remain strongly predictive of validated outcomes without those controls.
• Medium confidence, 18–48 months: provenance and machine-readable experimental records will become product differentiators for scientific agents. Disconfirming signal: independent reproduction rates show little relationship to provenance features.

---