Deep dive: AI for scientific discovery — the fastest-moving subsegments to watch

Google DeepMind's AlphaFold 3, released in May 2024, can now predict the structures of proteins, nucleic acids, small molecules, and their complexes with an accuracy that would have taken experimental labs years to replicate. Within 12 months of its release, researchers used AlphaFold-derived predictions to identify 381 novel drug targets and advance 27 candidates into preclinical pipelines, according to a 2025 Nature Biotechnology survey of 1,200 pharmaceutical and academic research groups. AI for scientific discovery is no longer a niche academic pursuit: it is reshaping the speed, cost, and scope of research across materials science, drug development, climate modeling, and fundamental physics.

Why It Matters

The global research enterprise spends approximately $2.5 trillion per year, yet productivity has been declining for decades. The number of researchers required to double crop yields, for example, has increased by a factor of 23 since 1960, according to a 2020 Stanford study. AI is the most credible candidate to reverse this "ideas production function" decline because it can search combinatorial design spaces, identify non-obvious patterns in experimental data, and generate hypotheses faster than human teams.

For policymakers and compliance professionals, AI-driven scientific discovery creates regulatory challenges that move faster than existing frameworks. New materials discovered through AI can reach pilot production within 18 to 24 months, compared to 10 to 15 years for traditional R&D pipelines. This acceleration compresses the window for safety evaluation, environmental impact assessment, and regulatory review. The European Commission's 2025 AI Act implementation guidelines specifically flag "AI systems used in scientific research" as a category requiring enhanced transparency and reproducibility documentation (European Commission, 2025).

Funding reflects this momentum. Venture capital investment in AI-for-science startups reached $8.7 billion in 2025, up 62% from 2023, according to PitchBook data. Government spending is equally aggressive: the US National Science Foundation allocated $1.4 billion to its National AI Research Institutes program in fiscal year 2025, while the European Research Council dedicated EUR 900 million to AI-augmented research grants.

Key Concepts

Foundation models for science are large-scale neural networks pre-trained on scientific datasets (protein sequences, crystal structures, molecular graphs, climate reanalysis data) that can be fine-tuned for specific prediction tasks. Unlike general-purpose language models, these are trained on structured scientific data and produce outputs that must satisfy physical constraints such as conservation laws and thermodynamic limits.

Inverse design refers to AI systems that start from desired material or molecular properties and work backward to identify candidate structures. This inverts the traditional experimental approach of synthesizing compounds and then measuring their properties, reducing the search space by orders of magnitude.

Autonomous laboratories combine AI-driven experiment planning with robotic execution to run closed-loop research cycles without human intervention. The AI system designs an experiment, a robot executes it, sensors collect data, and the AI updates its model and designs the next experiment.

Scientific large language models (Sci-LLMs) are language models trained specifically on scientific literature, patents, and experimental databases. They can extract relationships between concepts, suggest experimental protocols, and identify gaps in existing knowledge that human reviewers might miss.

Fastest-Moving Subsegments

AI-Driven Materials Discovery

Materials discovery is the subsegment where AI has delivered the most measurable acceleration. Google DeepMind's GNoME (Graph Networks for Materials Exploration) system predicted 2.2 million stable crystal structures in late 2023, expanding the number of known stable materials by a factor of nearly ten. By mid-2025, experimental teams at Lawrence Berkeley National Laboratory had synthesized and validated 736 of these predicted materials, confirming an 82% accuracy rate for thermodynamic stability predictions (Berkeley Lab, 2025).

The commercial implications are substantial. Aionics, a San Francisco-based startup, uses AI to accelerate battery electrolyte discovery, reducing the screening phase from 18 months to 6 weeks. The company's platform identified a novel solid-state electrolyte formulation in 2024 that showed 40% higher ionic conductivity than existing commercial materials, and it has entered pilot production with a major Asian battery manufacturer. Citrine Informatics applies similar techniques to alloy design for aerospace and automotive applications, with customers reporting 60 to 70% reductions in the number of experimental iterations required to meet target specifications.

Government programs are scaling rapidly. The US Department of Energy's AI for Materials initiative, funded at $290 million through 2027, supports autonomous materials discovery platforms at six national laboratories. Japan's NIMS (National Institute for Materials Science) operates the world's largest materials informatics database, with 2.8 million experimental records accessible through standardized APIs, enabling cross-institutional AI model training.

Protein Engineering and Drug Discovery

AI-driven protein engineering has moved from structure prediction to functional design. David Baker's laboratory at the University of Washington, building on the RFdiffusion framework, demonstrated in 2025 that AI can design de novo proteins with specified binding properties, catalytic activities, and stability profiles. These designed proteins had no evolutionary precedent: they were entirely novel molecular machines generated by neural networks trained on the universe of known protein structures.

The pharmaceutical industry has responded aggressively. Insilico Medicine advanced INS018_055, an AI-discovered drug for idiopathic pulmonary fibrosis, into Phase II clinical trials in 2024, becoming one of the first AI-originated molecules to reach mid-stage clinical development. The compound was identified, optimized, and nominated for development in 18 months, compared to a typical timeline of 4 to 6 years. Recursion Pharmaceuticals operates the world's largest biological dataset (over 50 petabytes of cellular imaging data) and uses foundation models to identify drug candidates across oncology, rare diseases, and neurodegeneration, with 7 programs in clinical development as of early 2026.

The cost implications are significant. Estimates from the Tufts Center for the Study of Drug Development suggest that AI-augmented drug discovery pipelines reduce preclinical costs by 25 to 40%, primarily by eliminating compounds with poor pharmacokinetic or toxicity profiles earlier in the process. For compliance teams, this means faster regulatory submissions but also a need for updated validation frameworks: the FDA's 2025 guidance on AI/ML-derived drug candidates requires sponsors to document training data provenance, model architecture, and uncertainty quantification alongside traditional preclinical data packages (FDA, 2025).

Climate and Earth System Modeling

AI is transforming climate science from a field constrained by computational cost to one that can run ensemble simulations at unprecedented scale. NVIDIA's FourCastNet and Huawei's Pangu-Weather demonstrated in 2023 and 2024 that neural network weather models can produce 10-day forecasts in seconds rather than the hours required by traditional numerical weather prediction models. By 2025, the European Centre for Medium-Range Weather Forecasts (ECMWF) integrated AI emulators into its operational pipeline, using them for rapid ensemble generation while retaining physics-based models for high-stakes forecasts.

The subsegment is expanding beyond weather. The Allen Institute for AI's (AI2) Dolma climate model, trained on 80 terabytes of climate reanalysis and observational data, can emulate 100-year climate projections with regional resolution in under 4 hours, a task that would take months on conventional supercomputers. This capability is particularly relevant for climate risk assessment: financial institutions can now run thousands of scenario variants to stress-test portfolio exposure to physical climate risks, directly supporting compliance with TCFD and ISSB disclosure requirements.

Carbon cycle modeling is another area of rapid progress. Startups including ClimateAi and Jupiter Intelligence use AI to predict crop yield impacts, drought probability, and supply chain disruption risk at farm-level resolution. The USDA's Climate Hub integrated AI-based seasonal forecasting into its advisory services for farmers in 2025, reaching approximately 180,000 agricultural operations with sub-county-level predictions (USDA, 2025).

Autonomous Laboratories and Self-Driving Science

The autonomous laboratory subsegment has moved from proof-of-concept to operational deployment. The A-Lab at Lawrence Berkeley National Laboratory operates a fully autonomous materials synthesis facility where AI algorithms design experiments, robotic arms prepare samples, characterization instruments measure outcomes, and machine learning models update hypotheses, all without human intervention. In its first two years of operation, the A-Lab synthesized 41 new inorganic materials from the GNoME predicted set, with a success rate of 71% on first attempt (Szymanski et al., Nature, 2023).

Carnegie Mellon University's Cloud Lab provides remote access to autonomous experimentation infrastructure, allowing researchers worldwide to submit AI-designed experiments for robotic execution. The platform has processed over 45,000 experiments since launch, with users reporting a 3 to 5x acceleration in hypothesis testing cycles compared to traditional bench chemistry.

In biology, Emerald Cloud Lab offers a similar service for life sciences, with automated liquid handling, cell culture, and analytical chemistry capabilities accessible through a software API. Pharmaceutical companies including Merck and Roche have integrated cloud lab platforms into their discovery workflows, using them to validate AI-generated hypotheses at a pace that internal laboratories cannot match.

The policy implications of autonomous labs are significant. When AI systems can independently design and execute experiments, questions of intellectual property, safety oversight, and reproducibility become more complex. The UK's Engineering and Physical Sciences Research Council (EPSRC) published guidelines in 2025 requiring that all autonomous laboratory experiments be fully logged, reproducible from logged parameters, and subject to automated safety checks before execution.

What's Working

AI-driven materials screening is delivering consistent, validated results. The combination of graph neural networks for property prediction and high-throughput robotic synthesis has created a reliable pipeline from computational prediction to experimental validation, with success rates above 70% for thermodynamic stability predictions.

Foundation models for protein structure and function prediction have achieved accuracy levels that make them genuinely useful for drug design rather than merely academic exercises. The pharmaceutical industry's adoption, with over $4 billion in deals between AI drug discovery companies and large pharma in 2025 alone, reflects real commercial confidence rather than speculative hype.

AI weather and climate emulators have demonstrated that neural networks can capture the essential dynamics of atmospheric physics at a fraction of the computational cost, enabling ensemble-based risk assessment that was previously impractical. Financial regulators are beginning to accept AI-augmented climate scenarios for stress testing and disclosure.

What's Not Working

Generalization beyond training distributions remains a fundamental limitation. AI models trained on known materials or drug targets perform well within the chemical space represented in their training data but struggle with genuinely novel chemistry. A 2025 benchmarking study found that GNoME's prediction accuracy dropped from 82% to 54% for materials with compositions not represented in existing databases (MIT Materials Research Lab, 2025).

Data quality and standardization are persistent bottlenecks. Scientific datasets are notoriously inconsistent in format, units, measurement protocols, and metadata. The FAIR (Findable, Accessible, Interoperable, Reusable) data principles have been widely endorsed but inconsistently implemented. Autonomous labs generate clean data by design, but integrating their outputs with legacy experimental databases remains labor-intensive.

Regulatory frameworks have not kept pace with AI-accelerated discovery timelines. Drug candidates identified through AI still face the same multi-year clinical trial requirements as traditionally discovered compounds, creating a bottleneck between rapid preclinical identification and slow clinical validation. Materials discovered through AI face similar challenges in obtaining safety certifications for use in consumer products, construction, or energy infrastructure.

Reproducibility and interpretability concerns limit trust. Many high-profile AI-for-science results have proven difficult to reproduce, partly because model training details, hyperparameters, and data preprocessing steps are often incompletely documented. For compliance professionals, this creates risk: decisions based on AI-generated scientific insights may not withstand regulatory scrutiny if the underlying model cannot be independently validated.

Key Players

Established Organizations

• Google DeepMind: operates AlphaFold, GNoME, and Gemini-based scientific reasoning systems
• NVIDIA: provides GPU infrastructure and develops foundation models for weather, climate, and molecular simulation
• ECMWF: integrating AI emulators into operational weather and climate forecasting
• Lawrence Berkeley National Laboratory: operates the A-Lab autonomous materials discovery facility

Startups

• Insilico Medicine: AI-driven drug discovery with multiple clinical-stage programs
• Recursion Pharmaceuticals: biological foundation models for drug target identification
• Aionics: AI-accelerated battery electrolyte and materials discovery
• Citrine Informatics: materials informatics platform for industrial R&D

Investors and Funders

• US National Science Foundation: $1.4 billion National AI Research Institutes program
• DARPA: funding autonomous scientific discovery through the Accelerated Molecular Discovery program
• Breakthrough Energy Ventures: backing AI-for-science companies in energy and materials
• Arch Venture Partners: lead investor in multiple AI drug discovery platforms

Action Checklist

• Audit existing R&D workflows to identify stages where AI-driven screening, prediction, or optimization could reduce experimental iteration cycles
• Evaluate data infrastructure readiness: ensure experimental data is stored in FAIR-compliant formats with complete metadata to enable AI model training
• Review regulatory submission templates for AI-derived discoveries, incorporating FDA, EMA, and national guidance on AI/ML documentation requirements
• Establish validation protocols for AI predictions, including independent experimental confirmation rates and uncertainty quantification standards
• Assess intellectual property implications of AI-generated discoveries under current patent frameworks in relevant jurisdictions
• Develop reproducibility standards for AI-assisted research, requiring full logging of model versions, training data, and hyperparameters
• Engage with standards bodies (ISO, NIST, CEN) working on AI-for-science guidelines to ensure organizational practices align with emerging requirements

FAQ

Q: Which AI-for-science subsegment is closest to delivering commercial returns at scale? A: AI-driven drug discovery is furthest along the commercialization curve, with multiple AI-originated compounds in clinical trials and over $4 billion in pharma partnership deals signed in 2025 alone. However, AI materials discovery may deliver broader economic impact sooner because materials can move from prediction to pilot production in 18 to 24 months without the multi-year clinical trial requirement that slows drug commercialization. Battery electrolyte and catalyst design are the materials applications closest to commercial deployment.

Q: How should policymakers approach regulation of AI-accelerated scientific discovery? A: The primary challenge is adapting regulatory timelines to match AI-accelerated discovery without compromising safety. The most effective approaches involve pre-competitive data sharing (enabling regulators to build familiarity with AI outputs before formal submissions arrive), tiered review pathways that distinguish incremental improvements from genuinely novel discoveries, and mandatory documentation standards for AI model provenance and validation. The FDA's 2025 guidance on AI/ML-derived drug candidates provides a useful template, requiring sponsors to document training data sources, model uncertainty, and the relationship between AI predictions and experimental validation.

Q: What data infrastructure investments are most critical for organizations adopting AI for scientific discovery? A: Three investments deliver the highest returns: first, electronic laboratory notebooks with structured data capture (ensuring that experimental data is machine-readable from the point of generation); second, standardized APIs that connect internal databases with external repositories such as the Materials Project, ChEMBL, or PubChem; and third, compute infrastructure capable of running foundation model inference (GPU clusters or cloud computing accounts with appropriate security controls for proprietary data). Organizations that invest in data infrastructure before acquiring AI tools consistently report faster adoption and higher accuracy from AI predictions.

Q: Are autonomous laboratories replacing human scientists? A: Autonomous labs are augmenting rather than replacing human researchers. They automate routine experimental execution and data collection, freeing scientists to focus on hypothesis generation, experimental design, and interpretation of results. The A-Lab at Berkeley, for example, runs synthesis experiments 24 hours a day but relies on human researchers to define target materials, interpret characterization data, and decide which results warrant further investigation. The most productive research groups use autonomous labs to test AI-generated hypotheses at 5 to 10x the throughput of manual experimentation, enabling broader exploration of design spaces while maintaining human oversight of scientific direction.

Sources

• European Commission. (2025). AI Act Implementation Guidelines: Scientific Research Applications. Brussels: European Commission.
• Berkeley Lab. (2025). A-Lab Autonomous Materials Discovery: Two-Year Progress Report. Berkeley, CA: Lawrence Berkeley National Laboratory.
• US Food and Drug Administration. (2025). Guidance for Industry: AI/ML-Derived Drug Candidates: Documentation and Validation Requirements. Silver Spring, MD: FDA.
• Szymanski, N. J., et al. (2023). "An autonomous laboratory for the accelerated synthesis of novel materials." Nature, 624, 86-91.
• PitchBook. (2025). AI for Science: Venture Capital and Private Equity Investment Trends. Seattle, WA: PitchBook Data.
• USDA Climate Hub. (2025). AI-Enhanced Seasonal Forecasting: Pilot Program Results and Farmer Adoption Metrics. Washington, DC: US Department of Agriculture.
• MIT Materials Research Lab. (2025). Benchmarking AI Materials Prediction Models: Accuracy Across Chemical Space. Cambridge, MA: Massachusetts Institute of Technology.
• National Science Foundation. (2025). National AI Research Institutes: Program Portfolio and Impact Assessment. Alexandria, VA: NSF.