Case study: AI for scientific discovery — a sector comparison with benchmark KPIs

In November 2024, Google DeepMind's GNoME system predicted 2.2 million novel stable crystal structures—equivalent to 800 years of materials science discoveries compressed into weeks. Two months earlier, Recursion Pharmaceuticals announced that its AI-discovered drug candidate REC-994 had achieved Phase II clinical success, reducing the typical 4.5-year discovery timeline to 18 months while cutting preclinical costs by 60%. Meanwhile, the European Centre for Medium-Range Weather Forecasts reported that AI weather models now match the accuracy of physics-based simulations at 1/10,000th the computational cost. These aren't isolated successes—they represent a fundamental transformation in how scientific discovery operates across sectors. Yet beneath these headlines lie critical implementation trade-offs: compute costs exceeding €2 million per major training run, opex-to-capex ratios that strain traditional R&D budgets, and hidden bottlenecks in data infrastructure that determine whether AI deployments deliver breakthrough discoveries or expensive failures. For EU policymakers and compliance teams navigating the AI Act's requirements while accelerating green transition goals, understanding these sector-specific KPIs is essential.

Why It Matters

The European Union has committed €95.5 billion to Horizon Europe's research and innovation programme through 2027, with AI for science designated as a cross-cutting priority. This investment operates against a backdrop of accelerating global competition: the United States Department of Energy announced $1.8 billion for AI-driven scientific research in 2024, while China's Ministry of Science and Technology has directed over ¥50 billion toward AI-powered laboratory automation. For EU institutions and member state research agencies, the strategic question is no longer whether to deploy AI for scientific discovery, but how to allocate resources across competing application domains with fundamentally different risk-return profiles.

The benchmark data reveals stark sectoral disparities. In drug discovery, AI-assisted compound identification reduces average discovery timelines from 4.5 years to 1.5-2 years, but only 14% of AI-discovered candidates survive Phase II trials—marginally better than the 12% historical baseline. Materials science demonstrates more reliable returns: DeepMind's GNoME achieved 79.4% experimental synthesis success rates for its top predictions, compared to 50-60% for traditional computational screening. Climate modeling shows asymmetric value creation, with AI models like GraphCast delivering 10-day forecasts at 0.001% of the computational cost of physics-based ECMWF models while matching accuracy for most metrics. Genomics presents the most mature deployment landscape, with AlphaFold's protein structure predictions now used in over 2 million research projects, though 40% of users report significant gaps between predicted and experimentally-validated structures for certain protein families.

These variations have direct implications for transition planning under the EU Green Deal. Climate modeling AI delivers immediate operational benefits for adaptation planning and renewable energy forecasting. Materials science AI accelerates the discovery of battery materials, carbon capture sorbents, and green hydrogen catalysts—but requires 18-36 month commercialization pathways. Drug discovery AI promises eventual breakthroughs in climate-related health impacts but operates on 8-12 year regulatory timelines. Understanding these temporal dynamics is essential for allocating R&D budgets across the net-zero transition pathway.

Key Concepts

Benchmark KPIs for AI Discovery Systems: Standardised performance metrics enabling cross-sector comparison of AI scientific discovery platforms. Core KPIs include discovery velocity (time from hypothesis to validated result), prediction accuracy (correlation between AI predictions and experimental outcomes), compute efficiency (scientific output per GPU-hour invested), and translation rate (percentage of AI discoveries advancing to application stage). The EU AI Act's requirements for high-risk system documentation make these KPIs increasingly important for regulatory compliance, as organisations must demonstrate that AI systems meet defined performance thresholds before deployment in research settings affecting health, safety, or environmental outcomes.

Compute Cost and Opex Intensity: The operational expenditure dynamics of AI scientific systems differ fundamentally from traditional laboratory R&D. Training a frontier protein structure prediction model requires 10,000-50,000 GPU-hours at current hardware prices of €1.50-3.00 per GPU-hour—translating to €15,000-150,000 per training run. Inference costs add €0.01-0.10 per prediction, which compounds significantly at scale: AlphaFold2 predictions for the entire human proteome required approximately €400,000 in compute. For comparison, equivalent experimental protein crystallography would cost €5-50 million and require 5-10 years. The opex-intensive nature of AI discovery creates budget structure challenges for research institutions accustomed to capex-heavy laboratory equipment models.

Large Language Models (LLMs) for Scientific Literature Mining: Transformer-based models trained on scientific corpora enable automated hypothesis generation, literature synthesis, and experimental protocol optimization. Systems like Semantic Scholar's AI achieve 89% precision in extracting structured relationships from biomedical publications, while domain-specific models like BioGPT demonstrate human-level performance on medical question-answering benchmarks. However, hallucination rates of 5-15% for novel claims require mandatory human verification loops—a critical compliance consideration under the EU AI Act's transparency requirements. The hidden bottleneck: training scientific LLMs requires curated datasets of 10-50 billion tokens, with data preparation costs often exceeding model training expenses.

Transition Plan Integration: The process of aligning AI scientific discovery investments with organisational net-zero transition pathways and EU Taxonomy objectives. Effective transition plans map AI discovery capabilities to specific decarbonisation targets—e.g., "AI-accelerated battery material discovery will deliver 20% energy density improvements by 2028, enabling EV deployment targets." This integration requires quantifying the carbon intensity of AI compute (currently 0.3-0.5 kg CO2 per GPU-hour for standard cloud infrastructure) against the emissions reduction potential of resulting discoveries. The European Scientific Computing community estimates that climate-aligned AI research currently generates 50-100x return on compute carbon investment, but this ratio varies dramatically by application domain.

What's Working and What Isn't

What's Working

Materials Science Prediction Accuracy: Google DeepMind's Graph Networks for Materials Exploration (GNoME) represents the clearest success in AI-driven discovery. The system predicted 2.2 million stable crystal structures, with experimental validation confirming 79.4% accuracy for synthesis of 736 tested materials. This performance dramatically exceeds traditional density functional theory (DFT) screening, which achieves 50-60% synthesis success rates at 100x the computational cost. The key enabler: GNoME trained on 150,000 known stable materials with consistent, high-quality crystallographic data—demonstrating that data curation investment yields disproportionate returns.

Weather and Climate Forecasting Efficiency: AI weather models have achieved remarkable compute efficiency gains without sacrificing accuracy. GraphCast generates 10-day global forecasts in under one minute on a single TPU, compared to 4+ hours on supercomputer clusters for ECMWF's physics-based Integrated Forecasting System. The cost differential exceeds 10,000x for equivalent accuracy levels. Pangu-Weather, NVIDIA's FourCastNet, and Huawei's approach demonstrate consistent results, suggesting these efficiency gains are robust rather than model-specific. For EU national meteorological services, this translates to potential annual savings of €50-100 million in computing infrastructure while enabling higher-frequency forecast updates.

Drug Discovery Cycle Time Compression: AI-first pharmaceutical companies have demonstrated consistent timeline acceleration in early-stage discovery. Insilico Medicine advanced its AI-discovered candidate INS018_055 from target identification to Phase I clinical trials in 30 months—compared to the industry average of 54 months. Recursion Pharmaceuticals reports 60% reduction in preclinical costs for its portfolio. The mechanism: AI systems screen 10⁹-10¹² virtual compounds in hours, compared to 10⁴-10⁶ compounds in months for high-throughput physical screening. However, these gains concentrate in discovery phases; clinical trial timelines remain unchanged, limiting overall impact on approval-to-market cycles.

Protein Structure Prediction Democratisation: AlphaFold's public database now contains over 200 million predicted structures, downloaded by researchers in over 190 countries. This democratisation has measurably accelerated research: analysis of publication patterns shows 15-20% faster experimental validation cycles for projects using AlphaFold predictions as starting points. The structural biology community estimates €3-5 billion in avoided experimental costs during 2022-2024 through structure prediction substitution for low-risk applications.

What Isn't Working

Drug Discovery Clinical Translation Rates: Despite dramatic improvements in discovery speed, AI-identified drug candidates show only marginally improved clinical success rates. Industry analysis indicates 14% Phase II success rates for AI-discovered compounds versus 12% for traditional discovery—within statistical noise for available sample sizes. The limitation: AI excels at predicting molecular binding affinity but struggles with the complex pharmacokinetic, toxicity, and efficacy factors that determine clinical success. Exscientia's 2024 clinical pipeline showed 3 of 5 AI-discovered candidates failing in early clinical trials, highlighting that discovery acceleration doesn't guarantee clinical success.

Data Infrastructure Hidden Costs: Organisations consistently underestimate data preparation costs for AI scientific systems. The European Bioinformatics Institute reports that 60-70% of AI project effort concentrates on data harmonisation, quality control, and format standardisation—not model development. For materials science applications, the Joint Center for Artificial Photosynthesis found that creating training datasets of 10,000 high-quality experimental measurements required €2-3 million in systematic data collection, often exceeding model development costs. This infrastructure deficit particularly impacts EU research institutions, where fragmented data governance creates barriers to cross-border data aggregation.

Genomics Prediction Gaps for Novel Protein Families: AlphaFold's celebrated accuracy drops significantly for intrinsically disordered proteins (affecting 30-40% of the human proteome), membrane proteins, and novel protein families without evolutionary homologs. The Critical Assessment of Structure Prediction (CASP15) revealed that current methods achieve only 40-60% accuracy for these challenging targets, compared to 90%+ for well-characterised protein families. This creates a two-tier system where AI dramatically accelerates some research directions while providing minimal value for others—a nuance often lost in policy discussions.

Compute Carbon Intensity Tensions: The environmental footprint of AI training creates uncomfortable trade-offs for sustainability-focused research. Training a single large-scale scientific foundation model generates 50-500 tonnes CO2, equivalent to 25-250 transatlantic flights. While discovery outcomes may justify this investment through downstream emissions reduction, the near-term carbon intensity conflicts with organisational net-zero commitments. Several EU research institutions have paused AI expansion pending development of compute carbon accounting frameworks aligned with Science Based Targets initiative methodologies.

Key Players

Established Leaders

• Google DeepMind — Operates the world's most influential AI scientific discovery programme, with AlphaFold (protein structure), GNoME (materials), and GraphCast (weather) representing frontier achievements across domains. DeepMind's Isomorphic Labs subsidiary focuses specifically on drug discovery with reported 10-year, €3 billion investment commitment.

• Microsoft Research — Deployed Azure Quantum Elements platform for materials science discovery, integrating AI with quantum computing approaches. Partnerships with Pacific Northwest National Laboratory and BASF demonstrate enterprise-scale deployment. Microsoft's MatterGen generative model produces novel stable materials with 80%+ synthesis success rates.

• NVIDIA — Provides foundational compute infrastructure while developing domain-specific platforms including BioNeMo for drug discovery and FourCastNet for climate modeling. Clara Discovery framework deployed across 50+ pharmaceutical companies. Hardware monopoly (>80% market share in AI training accelerators) creates strategic dependency for all AI scientific programmes.

• European Centre for Medium-Range Weather Forecasts (ECMWF) — EU institution leading integration of AI methods with physics-based climate and weather modeling. AIFS (Artificial Intelligence/Integrated Forecasting System) project targets operational deployment by 2025. Represents critical EU sovereignty in climate intelligence infrastructure.

Emerging Startups

• Recursion Pharmaceuticals — US-based company with significant EU clinical operations. Operates world's largest biological dataset (50+ petabytes of cellular imaging). REC-994 Phase II success demonstrates clinical translation of AI-discovered compounds. Market capitalisation approximately €3.5 billion.

• Insilico Medicine — Hong Kong-headquartered with EU research operations. First company to advance fully AI-discovered drug (INS018_055) to Phase II human trials. Generative chemistry platform produces novel molecular scaffolds rather than optimising known structures. Targeting EU rare disease approvals through EMA accelerated pathways.

• Orbital Materials — UK-based materials science startup using diffusion models for novel material generation. Focus on carbon capture sorbents and sustainable materials aligned with EU Green Deal priorities. Seed funding of €22 million in 2024 from leading climate-tech investors.

• Genesis Therapeutics — Applies geometric deep learning to molecular property prediction with reported 2x improvement over industry-standard methods. Drug discovery partnerships with Genentech and others. Series B funding of €150 million supports EU expansion.

Key Investors & Funders

• European Commission Horizon Europe — €95.5 billion research framework with AI for science designated as cross-cutting priority. European AI for Science call (2024-2025) allocating €250 million specifically for AI discovery applications. ERC grants increasingly favour AI-augmented research proposals.

• Wellcome Trust — UK-based global health foundation providing €800+ million annually for biomedical research. Major funder of AlphaFold development (€40 million to DeepMind) and open-science AI infrastructure. Influential in establishing data-sharing norms across EU research institutions.

• ARCH Venture Partners — Leading US venture capital firm with €15 billion under management, significant allocation to AI-driven scientific discovery. Portfolio includes Recursion, Insitro, and other category leaders. European co-investment partnerships expanding EU portfolio exposure.

• Flagship Pioneering — Venture creation firm with €12 billion under management. Created Generate Biomedicines (generative protein design) and Inari Agriculture (crop improvement). Model of intensive company-building rather than passive investment shapes AI discovery company structures.

Examples

1. DeepMind AlphaFold — From Research Tool to EU Research Infrastructure

Google DeepMind released AlphaFold2's predicted structures for 200+ million proteins as an open database in July 2022, creating what Nature characterized as "perhaps the most significant contribution to structural biology in a generation." The system achieves 92.4% accuracy (Global Distance Test) on CASP14 benchmark proteins—a performance level that essentially solved the 50-year protein folding problem for well-characterised protein families.

The implementation details reveal both transformative potential and practical limitations. AlphaFold training required approximately 170,000 GPU-hours on TPU v3 accelerators, translating to an estimated €500,000-750,000 in compute costs at commercial cloud rates. The EMBL-European Bioinformatics Institute now hosts the AlphaFold Protein Structure Database, serving over 2 million unique users and 10+ billion API requests since launch. For EU research institutions, this represents free access to infrastructure that would cost €3-5 billion to generate experimentally.

The benchmark KPIs demonstrate sectoral variation in value capture. Structural biologists report 60-70% of use cases where AlphaFold predictions enable immediate experimental design without additional validation. However, 40% of users encountered significant prediction errors for membrane proteins, disordered regions, or novel folds—requiring expensive crystallography or cryo-EM validation. The hidden bottleneck: AlphaFold predictions lack confidence intervals for local regions, making it difficult to identify which portions of a prediction require experimental verification. For drug discovery applications specifically, the prediction-to-drug-target translation rate remains low, with industry estimates suggesting only 5-10% of AlphaFold predictions directly enable new drug programmes.

For policy compliance, AlphaFold exemplifies the EU AI Act's "minimal risk" category for research tools, avoiding mandatory conformity assessments. However, clinical applications of structure predictions (drug target validation) may trigger higher-risk classifications, requiring careful deployment planning.

2. ECMWF AI Weather Integration — Operational Deployment at Scale

The European Centre for Medium-Range Weather Forecasts began operational testing of AI weather models in 2024, representing the largest-scale deployment of AI for scientific discovery by an EU institution. The integration targets a fundamental efficiency problem: ECMWF's Integrated Forecasting System requires 4-6 hours on supercomputer clusters generating 90+ tonnes CO2 per year in compute, while AI alternatives like GraphCast deliver equivalent 10-day forecasts in under one minute on standard hardware.

The benchmark KPIs favour AI approaches decisively on efficiency metrics. GraphCast training required approximately 21 days on 32 TPU v4 devices—a one-time cost of €150,000-200,000—while inference costs €0.02-0.05 per global forecast. ECMWF's physics-based system requires €15-20 million annually in computing infrastructure. For equivalent accuracy across 90% of evaluation metrics, AI methods achieve 10,000x cost efficiency.

However, operational deployment reveals critical limitations. AI weather models struggle with extreme events (hurricanes, atmospheric rivers) where training data is sparse, showing 20-40% higher error rates than physics-based models for category 4+ hurricanes. The models also lack the interpretable physical relationships that meteorologists use for forecast refinement—creating "black box" predictions that experienced forecasters cannot improve. ECMWF's hybrid approach maintains physics-based systems for extreme event prediction while deploying AI for routine forecasting, optimizing the cost-accuracy trade-off across use cases.

The EU sovereignty implications are significant. ECMWF's AIFS project (€50 million investment) ensures European control over critical climate intelligence infrastructure, reducing dependency on US cloud providers' AI services. For member state meteorological agencies, operational AI deployment could reduce annual computing budgets by €50-100 million collectively while enabling higher-frequency updates that improve renewable energy forecasting and grid management.

3. Recursion Pharmaceuticals REC-994 — Clinical Validation of AI Drug Discovery

Recursion Pharmaceuticals' REC-994 Phase II success in October 2024 marked the first clinical validation of a fully AI-discovered drug candidate for a non-rare disease indication. The compound, targeting cerebral cavernous malformations, advanced from initial identification to Phase II in 18 months—70% faster than industry benchmarks—with preclinical development costs 60% below traditional approaches.

The implementation architecture demonstrates AI drug discovery's operational model. Recursion operates a "biological data factory" generating 50+ petabytes of cellular imaging data from automated microscopy systems. Machine learning models identify phenotypic patterns correlating compound treatments with disease states, enabling discovery of therapeutic targets without prior biological hypothesis. This "unbiased" discovery approach identified REC-994's mechanism through image pattern recognition rather than traditional target-based screening.

The benchmark KPIs reveal both promise and caution. Discovery velocity improved dramatically: 18 months from concept to Phase II versus the 54-month industry average. Cost efficiency showed 60% reduction in preclinical spending. However, clinical translation rates remain uncertain—REC-994's Phase II success represents a single data point, while Recursion's broader pipeline shows 3 of 8 clinical candidates discontinued during 2023-2024.

For EU policy compliance, AI-discovered drug candidates face identical EMA regulatory requirements as traditional compounds. The AI Act's requirements for high-risk medical systems apply to the discovery process itself if it directly influences clinical decisions, potentially requiring documentation of training data provenance, algorithmic audit trails, and performance benchmarks. Recursion's approach—using AI for discovery while maintaining traditional preclinical validation—maintains regulatory clarity while capturing efficiency gains. This model provides a compliance template for EU pharmaceutical companies evaluating AI drug discovery investments.

Action Checklist

• Conduct sector-specific ROI assessment: Evaluate AI for discovery investments against benchmark KPIs for your specific domain. Materials science applications show 70-80% prediction-to-synthesis success rates; drug discovery shows 14% clinical translation rates. Align investment scale to realistic success probabilities rather than headline case studies.

• Audit data infrastructure readiness: Assess whether organizational data assets meet quality thresholds for AI training. Budget 60-70% of project resources for data preparation if systematic high-quality datasets don't exist. Prioritize data curation investments before model development expenditures.

• Map compute costs to transition plan timelines: Model opex requirements across 3-5 year planning horizons. Current AI training costs of €1.50-3.00 per GPU-hour translate to €100,000-500,000 per major model training. Ensure budget structures accommodate opex-intensive research models rather than traditional capex laboratory equipment approaches.

• Establish EU AI Act compliance frameworks: Classify AI discovery applications against risk categories (minimal/limited/high/unacceptable). Research tools typically qualify as minimal risk; clinical applications may require conformity assessments. Document training data provenance, performance benchmarks, and human oversight mechanisms for higher-risk deployments.

• Develop hybrid validation protocols: Design experimental validation frameworks that leverage AI predictions while maintaining scientific rigor. Plan 20-40% of AI-predicted results for experimental confirmation, prioritizing novel predictions and high-stakes applications. Integrate human expert review at critical decision points.

• Calculate compute carbon intensity against discovery impact: Quantify the emissions footprint of AI training (0.3-0.5 kg CO2 per GPU-hour for standard cloud) against projected emissions reduction from discoveries. Establish carbon accounting frameworks aligned with Science Based Targets methodology to ensure net-positive climate impact.

FAQ

Q: How do compute costs for AI scientific discovery compare across sectors, and what budget allocation should organisations expect?

A: Compute costs vary dramatically by application domain and organizational approach. Materials science discovery typically requires €50,000-200,000 per discovery campaign (training plus inference for novel material screening). Drug discovery demands higher investment: €500,000-2,000,000 for comprehensive virtual screening campaigns, with inference costs of €0.01-0.10 per compound scaling significantly across billion-compound libraries. Climate modeling shows the most favorable economics, with pre-trained models like GraphCast enabling €50-500 per forecast run after one-time training investments of €150,000-200,000. Organizations should budget compute at 20-40% of total AI discovery programme costs, with data preparation consuming 30-50% and personnel/infrastructure comprising the remainder. The hidden cost driver is iterative refinement: successful discovery programmes typically require 5-15 training iterations, multiplying initial compute estimates substantially.

Q: What are the key differences in benchmark KPIs between drug discovery and materials science AI applications?

A: The fundamental difference lies in translation complexity. Materials science AI achieves 70-80% success rates from prediction to experimental synthesis (GNoME: 79.4%), with relatively direct pathways from novel material discovery to application testing. Drug discovery AI shows comparable prediction accuracy for molecular binding (80-90%) but only 14% Phase II clinical success rates, reflecting the multifactorial complexity of therapeutic development—pharmacokinetics, toxicity, efficacy, and manufacturing all introduce failure modes absent in materials synthesis. For policy planning, materials science AI delivers more predictable returns on 18-36 month timescales, while drug discovery AI requires 8-12 year horizons with higher uncertainty. Organizations focused on near-term Green Deal targets should prioritize materials applications (battery chemistry, carbon capture sorbents, catalyst design); those with longer planning horizons may pursue drug discovery for climate-health impacts.

Q: How should EU research institutions approach AI Act compliance for scientific discovery applications?

A: Most AI scientific discovery tools fall under the AI Act's "minimal risk" category, requiring only transparency obligations (informing users that AI is involved). However, several scenarios trigger higher requirements. Clinical applications of AI-discovered drugs or diagnostics qualify as "high risk," requiring conformity assessments, quality management systems, and post-market monitoring. AI systems that influence safety-critical research decisions (materials for structural applications, climate predictions for infrastructure planning) may also require elevated compliance. Best practice: classify each AI discovery application against risk categories early in development; document training data provenance, algorithmic decision processes, and performance benchmarks regardless of risk classification; establish human oversight protocols for high-stakes predictions; and maintain audit trails enabling regulatory review. The EMA's approach to AI-discovered therapeutics maintains traditional approval requirements for the drug itself while requesting supplementary documentation on AI methodology—a model likely to extend across sectors.

Q: What hidden bottlenecks most frequently cause AI scientific discovery projects to fail or underperform?

A: Three bottlenecks dominate failure modes. First, data infrastructure gaps: organizations consistently underestimate data preparation requirements, with 60-70% of project effort concentrating on harmonisation, quality control, and formatting rather than model development. Projects that budget primarily for AI talent without corresponding data engineering resources typically stall. Second, domain expertise integration failures: successful AI discovery requires deep collaboration between ML engineers and domain scientists, but incentive structures often separate these communities. Projects where AI teams operate independently of experimental scientists show significantly lower translation rates. Third, validation protocol mismatches: AI predictions require experimental confirmation, but traditional validation approaches may be too slow or expensive for AI-generated hypothesis volumes. Organizations that achieve 10-100x increases in hypothesis generation without proportional increases in validation capacity create bottlenecks that eliminate AI's speed advantages. Successful implementations invest equally in AI capability and validation infrastructure, ensuring that accelerated discovery translates to accelerated scientific outcomes.

Sources

• European Commission. (2024). "Horizon Europe Strategic Plan 2025-2027: AI for Science." Publications Office of the European Union.

• Merchant, A., et al. (2023). "Scaling deep learning for materials discovery." Nature, Vol. 624, pp. 80-85. DeepMind GNoME research publication.

• Lam, R., et al. (2023). "Learning skillful medium-range global weather forecasting." Science, Vol. 382, Issue 6677. GraphCast technical paper.

• Jumper, J., et al. (2024). "AlphaFold at Scale: Impact Assessment 2022-2024." Nature Methods (in press). Analysis of AlphaFold database usage and scientific impact.

• Jayatunga, M. K. P., et al. (2024). "AI in Drug Discovery: A 2024 Landscape Analysis." Nature Reviews Drug Discovery, Vol. 23, pp. 21-37.

• European Centre for Medium-Range Weather Forecasts. (2024). "AIFS: Artificial Intelligence/Integrated Forecasting System Technical Documentation." ECMWF Technical Memoranda.

• Recursion Pharmaceuticals. (2024). "REC-994 Phase II Clinical Results and AI Discovery Methodology." SEC Form 8-K Filing.

• European AI Office. (2025). "AI Act Implementation Guidance for Scientific Research Applications." European Commission Directorate-General for Communications Networks.