Playbook: adopting AI for scientific discovery in 90 days

In 2024, private-sector investment in AI surged to $109 billion globally—12 times higher than China's total—while the rate of frontier AI capability improvement nearly doubled, accelerating from approximately 8 points per year to 15 points per year on standardized benchmarks according to Epoch AI research. The Nobel Prize in Chemistry 2024 was awarded to Demis Hassabis and John Jumper for their work on AlphaFold, marking the first time AI-driven scientific discovery received the field's highest honor. Meanwhile, 31 AI-discovered drugs entered human clinical trials by April 2024, with companies like Insilico Medicine demonstrating 10× cost reductions and compressing traditional 4–5 year drug discovery timelines to under 2 years. For research institutions, pharmaceutical companies, and sustainability-focused organizations across the EU, the question is no longer whether to adopt AI for scientific discovery, but how to implement it effectively within constrained timelines and regulatory frameworks. This playbook provides a structured 90-day rollout plan with concrete milestones, ownership models, and sector-specific KPIs to guide your transition.

Why It Matters

The intersection of AI and scientific discovery represents one of the most consequential technological shifts of the decade. According to the Stanford AI Index 2024, 90% of notable AI models now originate from industry rather than academia, a dramatic increase from 60% just one year prior. This concentration of capability in the private sector creates both opportunity and urgency for organizations seeking to remain competitive in research-intensive fields.

For sustainability applications specifically, AI-accelerated discovery holds transformative potential. DeepMind's GNoME (Graph Networks for Materials Exploration) identified 2.2 million new stable crystal structures in 2024, including 381,000 materials potentially relevant for energy storage and carbon capture technologies. CuspAI, a materials science startup, demonstrated the ability to screen millions of molecular combinations for direct air capture sorbents 10× faster than traditional methods. In climate modeling, NVIDIA's StormCast achieved 3-kilometer resolution hourly forecasts—an order of magnitude improvement over the standard 30-kilometer, 6-hour approach used by conventional systems.

The regulatory environment is also maturing. The FDA released draft guidance in January 2025 on AI use in drug development, signaling a pathway for clinical validation. The EU AI Act, now in implementation phase, requires organizations to classify AI systems by risk level and establish appropriate governance frameworks. Organizations that delay adoption risk not only competitive disadvantage but also the loss of institutional knowledge needed to navigate these emerging compliance requirements.

From an economic perspective, PwC projects AI will contribute €15 trillion to global GDP by 2030, with McKinsey estimating €2.3–3.9 trillion annually from generative AI applications alone. The scientific discovery segment—spanning drug development, materials science, climate modeling, and fundamental research—represents a significant share of this value creation opportunity.

Key Concepts

Foundation Models for Science

Foundation models trained on scientific literature, molecular structures, protein sequences, and experimental data form the backbone of modern AI-driven discovery. Unlike general-purpose language models, science-specific foundation models incorporate domain knowledge enabling them to predict molecular properties, suggest experimental designs, and identify promising research directions. AlphaFold 3, released in May 2024, exemplifies this approach by predicting structures of proteins, DNA, RNA, ligands, ions, and chemical modifications with 50%+ accuracy improvements over previous methods for protein-molecule interactions.

Autonomous AI Research Agents

Multi-agent AI systems represent the emerging paradigm for end-to-end scientific workflows. FutureHouse's platform, launched in May 2025, deploys specialized agents—Crow for literature synthesis, Owl for experimental design, Phoenix for hypothesis generation, and Finch for data analysis—that collaborate on complex research tasks. According to METR research, AI agent task completion length doubles every 7 months, suggesting autonomous research agents may match top human researchers in capability by 2027.

Generative Chemistry and Materials Design

Generative AI models trained on chemical space can propose novel molecules and materials not found in existing databases. Xaira Therapeutics raised $1 billion in 2024 specifically to advance generative chemistry approaches. These systems enable "zero-shot" design—generating candidates for specific properties without iterative experimental screening—potentially compressing discovery cycles by orders of magnitude.

Sector-Specific KPIs for AI-Driven Discovery

Sector	Key Performance Indicator	Baseline (Traditional)	AI-Enabled Target	Best-in-Class (2024)
Drug Discovery	Time to IND Filing	4–5 years	18–24 months	<18 months (Insilico)
Drug Discovery	Pre-clinical Candidate Cost	$10–15M	$1–3M	<$1.5M
Materials Science	Novel Material Screening Rate	100–1,000/month	100,000+/month	2.2M structures (GNoME)
Climate Modeling	Spatial Resolution	30 km	3–10 km	3 km (StormCast)
Protein Engineering	Design-Build-Test Cycles	6–12 months	4–8 weeks	<4 weeks (Cradle)
Literature Synthesis	Systematic Review Time	3–6 months	1–2 weeks	Days (FutureHouse)

What's Working

Closed-Loop Experimental Platforms

Organizations combining AI prediction with automated wet-lab execution demonstrate the highest ROI. Recursion Pharmaceuticals operates BioHive-2, the 35th most powerful supercomputer globally, integrated with robotic experimental platforms that generate proprietary training data. This closed-loop approach—where AI predictions are immediately validated and used to refine models—accelerates learning cycles dramatically. Recursion has 10 drug candidates in preclinical or clinical development as of 2024.

Strategic Pharma Partnerships

Emerging AI discovery companies that form deep partnerships with established pharmaceutical firms show stronger clinical translation rates. Isomorphic Labs partnered with Eli Lilly in 2024, combining AlphaFold-derived insights with Lilly's clinical development infrastructure. Insilico Medicine's $1.2 billion partnership with Sanofi exemplifies how AI-first biotechs can access the regulatory expertise and manufacturing capabilities needed to bring discoveries to patients.

Multi-Modal Data Integration

Success in AI-driven discovery correlates strongly with organizations that integrate diverse data types—genomics, imaging, clinical records, chemical libraries, and literature—into unified platforms. Owkin's $180 million partnership with Sanofi leverages multimodal AI combining histopathology, genomics, and clinical data for oncology drug development. This approach addresses the fundamental limitation that single data modalities rarely capture the complexity of biological systems.

Open Science Infrastructure

Collaborative initiatives that share pre-competitive data and models accelerate field-wide progress. The AlphaFold Database now contains 214 million protein structures accessible to 3+ million researchers across 190 countries. Meta's partnership with Georgia Tech and CuspAI produced OpenDAC, a dataset with 100+ million datapoints for direct air capture materials research. Organizations contributing to these ecosystems benefit from network effects while advancing collective scientific capability.

What's Not Working

Simulation-to-Synthesis Gap

A persistent challenge in AI materials discovery is translating computational predictions to real-world synthesis. MIT Technology Review noted in late 2024 that many AI-discovered materials remain theoretical, with synthesis validation lagging behind computational screening by years. Organizations that invest heavily in computational discovery without corresponding wet-lab validation capabilities frequently accumulate large candidate portfolios with uncertain practical value.

Data Quality and Provenance Issues

AI models are only as reliable as their training data. Many scientific datasets contain systematic biases, experimental errors, or incomplete metadata that compromise model performance. The failure of several high-profile AI drug candidates in clinical trials has been attributed to models trained on biased datasets that did not reflect patient diversity or real-world conditions. Establishing rigorous data governance and provenance tracking remains essential but underdeveloped in many organizations.

Regulatory Uncertainty

While the FDA's January 2025 guidance provides initial direction, significant ambiguity remains around AI model validation requirements, change control procedures, and liability frameworks. Organizations operating across jurisdictions face additional complexity as EU, US, and Asian regulators develop divergent approaches. This uncertainty slows deployment decisions and increases compliance costs.

Talent Scarcity

The intersection of AI expertise and deep domain knowledge in chemistry, biology, or physics remains exceptionally rare. Stanford AI Index 2024 notes that 78% of organizations now use AI, up from 55% the prior year, intensifying competition for qualified personnel. Salary inflation in AI roles has outpaced broader tech industry trends, creating particular challenges for academic institutions and smaller biotechs.

Over-Reliance on Benchmark Performance

Models that excel on standardized benchmarks sometimes fail to generalize to novel research questions. The gap between benchmark improvement (GPQA scores increased 48.9 percentage points from 2023 to 2024) and real-world scientific productivity suggests that current evaluation frameworks may not capture the factors most relevant to discovery success.

Key Players

Established Leaders

Google DeepMind (London, UK): The AlphaFold team won the 2024 Nobel Prize in Chemistry. AlphaFold 3, released May 2024, expanded capabilities to protein-ligand interactions. The AlphaFold Database provides free access to 214 million predicted structures. DeepMind's GNoME discovered 2.2 million new stable materials.

Recursion Pharmaceuticals (Salt Lake City, US): Operates one of the world's most powerful life science supercomputers (BioHive-2). Acquired Exscientia for $688 million in late 2024. Maintains partnerships with Roche/Genentech, Bayer, Merck, and Sanofi across oncology, immunology, and rare disease programs.

Insilico Medicine (Hong Kong/US): Demonstrated 10× cost reduction in drug discovery using generative AI. Nominated 18 pre-clinical candidates in 2 years versus the industry standard 4–5 year timeline. Multiple compounds in clinical trials, including a first-in-class fibrosis treatment.

NVIDIA (Santa Clara, US): Controls approximately 80% of the AI accelerator market. Provides BioNeMo platform for drug discovery. StormCast climate model achieved 3km resolution hourly forecasting. AI data center revenue reached $22.6 billion in Q1 2024, up 427% year-over-year.

Emerging Startups

Xaira Therapeutics (US): Raised $1 billion Series A in 2024 for generative chemistry and AI-powered drug discovery. Led by ARCH Venture Partners and Foresite Capital.

CuspAI (Cambridge, UK): Secured $100 million Series A at $520 million valuation in September 2025. Develops AI "search engine" for materials discovery, demonstrated 10× acceleration in screening for direct air capture materials. Partnered with Hyundai for battery and hydrogen fuel cell materials.

Periodic Labs (US): Raised record $300 million seed round in 2025 led by Andreessen Horowitz with participation from Jeff Bezos, Eric Schmidt, and Jeff Dean. Building autonomous AI laboratories for materials discovery. Founded by former OpenAI VP Liam Fedus and ex-DeepMind researcher Ekin Dogus Cubuk.

FutureHouse (US): Multi-agent AI platform for end-to-end scientific workflows. Released ether0 chemistry reasoning model in June 2025. Co-founded by MIT PhD Sam Rodriques.

Cradle (Netherlands): AI protein engineering platform with partnerships including Novo Nordisk, Johnson & Johnson, and Grifols. Raised $73 million Series B in November 2024, exceeding $100 million total funding.

Key Investors & Funders

ARCH Venture Partners: Led Xaira's $1 billion round. Historically one of the most successful life science venture firms with investments across AI-enabled drug discovery.

Andreessen Horowitz: Led Periodic Labs' record $300 million seed. Increasing allocation to AI-driven science applications across portfolio.

NVIDIA NVentures: Strategic investor in Terray Therapeutics, Radical AI, and CuspAI. Provides compute infrastructure advantages alongside capital.

Temasek: Singapore sovereign wealth fund co-led CuspAI's Series A. Active in AI science investments across Asia and Europe.

US Federal Government: Allocated $3.3 billion for non-defense AI R&D in FY2025. NSF, NIH, and DOE programs increasingly target AI-enabled scientific discovery.

EU Horizon Europe: Funding AI for science initiatives under Pillar II "Global Challenges and European Industrial Competitiveness." Emphasis on sustainable materials, health, and climate applications.

Examples

1. Insilico Medicine: From Algorithm to Clinic in Record Time

Insilico Medicine demonstrated what's possible when generative AI is applied across the full drug discovery pipeline. Using their proprietary AI platform, the company identified a novel target for idiopathic pulmonary fibrosis, designed a first-in-class small molecule inhibitor, and advanced the compound through pre-clinical development in under 18 months—compared to the industry average of 4–5 years. The lead candidate entered Phase 1 clinical trials in 2024. CEO Alex Zhavoronkov has stated the AI approach achieved approximately 10× cost reduction while also increasing probability of success by avoiding targets with unfavorable druggability profiles. The company's partnership with Sanofi, valued at up to $1.2 billion, validates pharma industry confidence in AI-driven approaches.

2. CuspAI: Accelerating Climate Materials Discovery

CuspAI applies AI to one of climate tech's most challenging problems: identifying new materials for direct air capture (DAC) of CO₂. Traditional materials screening for DAC sorbents requires years of experimental synthesis and testing. CuspAI's platform screens millions of candidate molecular structures computationally, predicting properties like CO₂ binding affinity, stability, and synthesis feasibility before any laboratory work begins. Their collaboration with Meta AI and Georgia Tech produced OpenDAC, a dataset containing over 100 million datapoints on DAC-relevant materials. In partnership with Kemira, CuspAI is also developing novel materials for PFAS water filtration. The company's $100 million Series A at a $520 million valuation reflects investor confidence in AI materials discovery for sustainability applications.

3. Recursion Pharmaceuticals: Building the Closed-Loop Discovery Engine

Recursion exemplifies the integrated approach to AI-driven drug discovery. The company combines BioHive-2—one of the world's most powerful supercomputers dedicated to biological research—with high-throughput robotic laboratories that generate proprietary cellular imaging data. This closed-loop system creates a continuous improvement cycle: AI models predict which experiments to run, robots execute them, results train better models. The December 2024 acquisition of Exscientia for $688 million consolidated two of the field's leading platforms. Recursion maintains active partnerships with Roche/Genentech (oncology), Bayer (fibrosis), and Sanofi (rare disease), with 10 programs in preclinical or clinical development. The company's $4 billion valuation reflects the strategic value of integrated AI-wet lab infrastructure.

Action Checklist

• Days 1–15: Assess organizational readiness. Inventory existing data assets, computational infrastructure, and domain expertise. Identify quick-win use cases where AI could augment current workflows (literature synthesis, experimental design prioritization, compound screening).

• Days 16–30: Establish governance framework. Define AI ethics principles, data governance policies, and model validation requirements aligned with EU AI Act risk categories. Designate responsible AI officer and establish review board.

• Days 31–45: Pilot foundation model deployment. Select 1–2 commercially available platforms (AlphaFold Server, NVIDIA BioNeMo, or domain-specific alternatives) for low-risk pilot applications. Train core team on prompt engineering and result interpretation.

• Days 46–60: Build data infrastructure. Implement data lake architecture connecting experimental records, literature, and external databases. Establish metadata standards and provenance tracking. Address GDPR and IP considerations for cross-border data flows.

• Days 61–75: Launch closed-loop experiment. Design experiment where AI predictions are validated experimentally. Instrument feedback loop to capture success/failure data for model refinement. Measure time-to-insight compared to baseline approaches.

• Days 76–90: Evaluate and scale. Assess pilot outcomes against KPIs. Document lessons learned. Develop 12-month roadmap for expanding successful applications. Identify partnership or vendor relationships needed for scale.

FAQ

Q: What is the minimum viable investment to begin adopting AI for scientific discovery?

A: Entry points vary significantly by use case. Organizations can access AlphaFold Server at no cost for academic research. Commercial cloud platforms like NVIDIA BioNeMo or AWS HealthOmics offer pay-per-use pricing starting at thousands of euros monthly for modest workloads. Meaningful enterprise deployments typically require €500,000–€2 million annually for platform licensing, cloud compute, and dedicated personnel. However, the 2024 Stanford AI Index notes that organizations seeing highest ROI invest early in data infrastructure and talent—areas where costs compound over time.

Q: How do we address AI validation requirements under EU regulatory frameworks?

A: The EU AI Act classifies most scientific discovery applications as "limited risk" unless they involve medical devices or clinical decision support, which may be "high risk." For high-risk applications, organizations must document training data provenance, conduct bias assessments, implement human oversight mechanisms, and maintain post-market surveillance. The FDA's January 2025 draft guidance provides useful reference points even for EU-focused organizations, as concepts like "predetermined change control plans" for adaptive AI systems are gaining traction with EMA as well.

Q: What talent profile is most critical for early-stage AI science adoption?

A: Organizations report the scarcest—and most valuable—talent combines deep domain expertise with practical AI engineering skills. A computational biologist who understands both protein biochemistry and transformer architectures, or a materials scientist comfortable with molecular dynamics and reinforcement learning, can bridge the gap between AI capabilities and scientific requirements. Stanford AI Index data shows 78% of organizations now use AI, creating intense competition for this hybrid talent. Consider apprenticeship models pairing domain experts with AI engineers as an alternative to hiring fully-formed hybrid candidates.

Q: How long before AI-discovered compounds typically reach clinical trials?

A: Based on publicly disclosed timelines from companies like Insilico Medicine and Exscientia, AI-discovered candidates have entered Phase 1 trials within 18–30 months of target identification—compared to 4–5 years for traditional approaches. However, clinical trial success rates for AI-discovered drugs remain similar to industry baselines (~10% from Phase 1 to approval), suggesting AI primarily accelerates pre-clinical phases rather than fundamentally changing clinical probability of success. The 31 AI-discovered drugs in human trials by April 2024 will provide more definitive data on clinical outcomes over the next 3–5 years.

Q: What are the biggest mistakes organizations make when adopting AI for scientific discovery?

A: Three failure patterns recur across early adopters. First, underinvesting in data infrastructure while over-indexing on model selection—even the most sophisticated AI cannot compensate for fragmented, poorly-annotated data. Second, treating AI adoption as a technology project rather than a scientific culture transformation requiring new workflows, incentives, and success metrics. Third, attempting to build proprietary platforms from scratch rather than leveraging existing infrastructure through partnerships or licensing, thereby consuming resources without generating competitive advantage.

Sources

• Stanford Institute for Human-Centered AI. "AI Index Report 2024." Stanford University. https://aiindex.stanford.edu/report/

• Epoch AI. "AI Capabilities Progress Has Sped Up." December 2024. https://epoch.ai/data-insights/ai-capabilities-progress-has-sped-up

• World Economic Forum. "Top 10 Emerging Technologies of 2024: AI for Scientific Discovery." https://www.weforum.org/publications/top-10-emerging-technologies-2024/

• Google DeepMind. "AlphaFold 3 Predicts the Structure and Interactions of All of Life's Molecules." May 2024. https://deepmind.google/science/alphafold/

• US Food and Drug Administration. "Draft Guidance: Artificial Intelligence in Drug and Biological Product Development." January 2025. https://www.fda.gov/regulatory-information/search-fda-guidance-documents

• Nature. "AI for Science 2025." Nature Collection. https://www.nature.com/collections/bfefgbacag

• McKinsey Global Institute. "The Economic Potential of Generative AI." 2024 Update.

• Recursion Pharmaceuticals. "BioHive-2 Supercomputer and Platform Overview." Corporate documentation, 2024.