Trend watch: AI for scientific discovery in 2026 — signals, winners, and red flags

The convergence of foundation models, automated experimentation, and massive scientific datasets has transformed AI from a supplementary research tool into a primary engine of scientific discovery. In 2025, AI-driven systems accounted for the identification of over 2.2 million novel candidate materials in the Materials Project database, predicted protein structures for virtually every known organism through AlphaFold, and accelerated drug candidate screening timelines from years to weeks. As of early 2026, the AI for science market has reached an estimated $14.6 billion globally, with venture investment exceeding $5.8 billion in the trailing twelve months according to PitchBook. But the rapid pace of adoption has also introduced significant risks: reproducibility failures, compute cost escalation, and a widening gap between AI-generated hypotheses and experimentally validated discoveries. This trend watch identifies the signals that matter, the emerging winners, and the red flags that product teams and investors should monitor through the remainder of 2026.

Why It Matters

Scientific discovery underpins every major sustainability challenge, from developing next-generation battery chemistries and catalysts for green hydrogen production to identifying novel carbon capture sorbents and engineering climate-resilient crops. Traditional experimental science operates on timelines measured in years and decades. AI has the potential to compress these timelines by orders of magnitude, but only if the technology is deployed with appropriate rigor and realistic expectations.

The economic stakes are enormous. McKinsey estimates that AI-accelerated R&D could generate $200 to $300 billion annually in value across pharmaceuticals, materials science, and energy technology by 2030. In emerging markets, where research infrastructure is limited and scientific talent is scarce, AI for science platforms offer the possibility of leapfrogging traditional laboratory bottlenecks. India's BioE3 policy initiative, launched in 2024, explicitly prioritizes AI-driven biotechnology as a pillar of national economic strategy. China's National Natural Science Foundation has allocated over $2 billion specifically for AI-augmented scientific research programs through 2027.

For product and design teams building tools for researchers, the landscape presents both opportunity and complexity. Scientists increasingly demand AI capabilities integrated into their workflows, but they also require transparency, reproducibility, and interpretability that many current AI systems fail to deliver. The teams that successfully bridge this gap between powerful AI capabilities and rigorous scientific methodology will capture disproportionate market share in a rapidly expanding sector.

Key Concepts

Foundation Models for Science are large-scale AI models pretrained on broad scientific datasets that can be fine-tuned for specific research tasks. Unlike task-specific models, foundation models (such as Meta's ESM-2 for protein sequences or Google DeepMind's GNoME for crystal structures) encode general scientific knowledge that transfers across applications. The key advantage is reduced data requirements for downstream tasks: a foundation model pretrained on millions of protein sequences can predict the function of a novel protein from as few as 50 labeled examples, compared to thousands required for models trained from scratch.

Automated Scientific Workflows combine AI prediction with robotic experimentation in closed-loop systems. The AI component generates hypotheses or candidate designs, robotic platforms execute experiments, results feed back to update the AI model, and the cycle repeats. These "self-driving laboratories" have demonstrated 10 to 100x acceleration in materials discovery and optimization. The A-Lab at Lawrence Berkeley National Laboratory autonomously synthesized 41 of 58 AI-predicted novel inorganic materials in its first 17 days of operation, validating the concept at meaningful scale.

Inverse Design reverses the traditional scientific workflow. Instead of synthesizing materials and then characterizing their properties, inverse design specifies desired properties and uses AI to identify molecular or material structures that satisfy those requirements. This approach is particularly powerful for sustainability applications where target specifications (such as CO2 adsorption capacity, thermal conductivity, or catalytic activity) are well-defined but the chemical space of potential solutions is too vast for systematic experimental exploration.

Retrieval-Augmented Generation (RAG) for Scientific Literature applies large language models to extract, synthesize, and reason over the scientific literature. With over 3 million new scientific papers published annually, no human researcher can maintain comprehensive awareness of relevant work. RAG systems retrieve relevant papers, data, and experimental protocols to inform AI-generated hypotheses, reducing duplication of effort and surfacing connections across disciplines that individual researchers might miss.

Signals That Matter in 2026

Signal 1: Wet Lab Validation Rates Are the True Metric

The most important signal separating genuine AI-for-science progress from hype is the rate at which AI-generated predictions survive experimental validation. In materials science, the validation rate for AI-predicted stable compounds has improved from approximately 30% in 2023 to 52% in 2025, based on data from the Materials Project and independent replication studies published in Nature. In drug discovery, AI-nominated lead compounds entering Phase I clinical trials have increased from 15 candidates in 2023 to over 70 in 2025, but the critical question is whether these compounds outperform traditionally discovered candidates in Phase II success rates. Early data from Recursion Pharmaceuticals and Insilico Medicine suggest comparable or modestly superior Phase II success rates for AI-discovered candidates, but the dataset remains small and definitive conclusions are premature.

Signal 2: Compute Costs Are Reshaping Research Economics

Training and running large scientific AI models requires substantial computational resources. The compute cost for training a state-of-the-art protein structure prediction model increased from approximately $2 million in 2022 to $15 million in 2025 as model sizes grew. Inference costs for running predictions at scale remain significant: generating structure predictions for a million protein sequences costs approximately $50,000 to $100,000 using cloud GPU resources. For emerging market research institutions with limited budgets, these costs create a two-tier system where well-funded organizations can leverage cutting-edge AI while others cannot. The counter-trend is the emergence of efficient, smaller models and open-source alternatives that narrow (but do not close) this gap. NVIDIA's BioNeMo platform and open-weight models like OpenFold have reduced barriers to entry, but the compute intensity of frontier science AI continues to grow faster than efficiency gains.

Signal 3: Regulatory Frameworks Are Catching Up

Scientific discoveries generated by AI face increasing regulatory scrutiny, particularly in pharmaceuticals and food safety. The US FDA issued guidance in 2025 on the use of AI and machine learning in drug development, requiring documentation of model training data, validation procedures, and uncertainty quantification for AI-generated evidence submitted in regulatory filings. The European Medicines Agency has proposed similar frameworks. For materials science and chemistry, regulatory implications are less direct but growing: the EU's REACH regulation increasingly requires computational toxicology assessments, and AI-predicted safety profiles are gaining acceptance as supplementary evidence in chemical registration dossiers. Product teams building AI tools for regulated industries must design for auditability and explainability from the start.

Emerging Winners

Google DeepMind

DeepMind continues to set the pace in AI for scientific discovery. Following AlphaFold's transformation of structural biology, the GNoME (Graph Networks for Materials Exploration) system identified 2.2 million predicted stable crystal structures, of which 381,000 are close to the convex hull of stability and thus experimentally accessible. The A-Lab collaboration with Lawrence Berkeley National Laboratory demonstrated that autonomous robotic synthesis could validate GNoME predictions at scale. DeepMind's combination of model development, dataset creation, and experimental validation partnerships positions it as the leading platform for materials discovery.

Recursion Pharmaceuticals

Recursion has built the world's largest proprietary dataset of biological images (over 25 petabytes) combined with chemical and genetic perturbation data, creating a foundation for AI-driven drug discovery at unprecedented scale. The company's LOWE (Large-scale Omics-Wide Experiments) platform enables phenotypic drug discovery across multiple therapeutic areas simultaneously. In 2025, Recursion had 7 programs in clinical trials, with 4 AI-nominated compounds advancing to Phase II. The company's partnership with NVIDIA to build BioHive, one of the world's most powerful supercomputers for biological research, signals a long-term infrastructure commitment.

Emerging Market Platforms: India's BioAI Ecosystem

India has emerged as a significant player in AI for scientific discovery, driven by a combination of strong computational talent, government policy support, and lower labor costs for data annotation and model training. The Indian Institute of Science's Materials Genome Centre has deployed AI-driven materials discovery workflows for battery electrolytes and catalysts, with multiple validated discoveries published in peer-reviewed journals. Startups like Peptris (computational peptide design) and Niramai (AI for medical diagnostics) demonstrate the breadth of India's AI-for-science ecosystem. The BioE3 policy initiative provides $1 billion in funding through 2027 for AI-augmented biomanufacturing and bioscience research, creating a supportive environment for continued growth.

Red Flags to Watch

Red Flag 1: The Reproducibility Gap

AI-driven scientific claims are facing growing reproducibility challenges. A 2025 analysis in Science found that 28% of AI-generated materials discovery claims could not be reproduced when independent teams attempted experimental validation using the published computational methods. Common failure modes include: training data leakage (where test compounds inadvertently appear in training sets), overfitting to narrow composition spaces, and inadequate uncertainty quantification that presents low-confidence predictions as high-confidence discoveries. Product teams should build reproducibility checks into their platforms, including automated detection of data leakage and calibration of prediction confidence scores.

Red Flag 2: "AI-Discovered" Marketing Without Substance

As AI for science gains visibility, a growing number of companies are relabeling conventional computational chemistry or high-throughput screening as "AI-driven discovery." The distinction matters: genuine AI-for-science approaches use learned representations and generate novel hypotheses beyond the training distribution, while rebranded conventional approaches apply standard computational methods at scale. Investors and partners should probe whether a company's AI approach generates genuinely novel candidates that human researchers would not have identified through conventional methods, or simply accelerates known search strategies.

Red Flag 3: Emerging Market Data Quality Challenges

While emerging markets offer significant growth opportunities for AI-for-science platforms, data quality and infrastructure limitations present real risks. Laboratory data from institutions with less standardized experimental protocols may introduce systematic biases into AI training datasets. Instrument calibration differences, reagent quality variations, and documentation practices that differ from standards at well-resourced institutions can degrade model performance when deployed across diverse research environments. Platforms that succeed in emerging markets will need robust data quality assessment, standardization tools, and models that are explicitly designed to handle heterogeneous data quality.

Red Flag 4: Compute Concentration Risk

The concentration of AI-for-science compute capacity among a small number of cloud providers and hardware manufacturers creates systemic risk. NVIDIA's dominance in GPU hardware (estimated 80%+ market share for AI training) and the concentration of cloud AI capacity among AWS, Google Cloud, and Microsoft Azure means that pricing changes, supply constraints, or policy decisions by these companies disproportionately affect the entire AI-for-science ecosystem. Research institutions in emerging markets are particularly vulnerable to compute access disruptions. Diversification strategies, including investment in alternative hardware architectures and edge computing for inference, should be factored into long-term planning.

Key Players

Established Leaders

Google DeepMind leads in both protein science (AlphaFold) and materials discovery (GNoME), with the deepest integration between AI prediction and experimental validation.

Microsoft Research operates AI for Science initiatives across chemistry, materials, and biology, with the Azure Quantum Elements platform providing integrated simulation and AI capabilities.

IBM Research focuses on AI-driven molecular discovery through the MolGX platform, with particular strength in computational chemistry for materials and sustainability applications.

Emerging Startups

Recursion Pharmaceuticals combines the largest proprietary biological dataset with purpose-built supercomputing infrastructure for AI-driven drug discovery.

Orbital Materials applies AI to accelerate materials discovery for carbon capture sorbents and sustainable chemicals, with a focus on climate-relevant applications.

Atinary Technologies provides a Bayesian optimization platform for experimental design that integrates with laboratory automation systems, reducing the number of experiments needed to optimize materials and processes by 5 to 10x.

Key Investors and Funders

Lux Capital has been among the most active investors in AI for science, with portfolio companies spanning materials discovery, drug development, and computational biology.

ARCH Venture Partners has backed multiple AI-for-science companies, with a focus on platforms that combine AI prediction with proprietary data generation.

National Science Foundation (US) allocated $1.8 billion to AI-related research funding in 2025, with specific programs targeting AI for materials, chemistry, and biological science.

Action Checklist

• Evaluate whether AI-for-science tools in your pipeline provide transparent uncertainty quantification and reproducibility documentation
• Assess compute cost trajectories for your AI-driven research workflows and develop contingency plans for cost escalation scenarios
• Require experimental validation data (not just computational predictions) when evaluating AI-for-science vendor claims or partnership proposals
• Build data quality assessment capabilities into platforms targeting emerging market research institutions
• Monitor regulatory developments in your target markets for AI-generated evidence requirements in product registration and approval processes
• Invest in explainability and auditability features for AI predictions, as these will increasingly be required by regulators and scientific journals
• Track the validation rate (predictions confirmed by experiment) as the primary performance metric for AI-for-science platforms, rather than the volume of predictions generated
• Diversify compute infrastructure dependencies to reduce concentration risk from single hardware or cloud providers

FAQ

Q: What is the current state of AI for scientific discovery in emerging markets? A: Emerging markets are rapidly adopting AI for science, driven by government policy support and the availability of computational talent. India, China, and Brazil have launched national AI-for-science initiatives with combined funding exceeding $5 billion through 2027. However, adoption is concentrated in well-funded national laboratories and top-tier universities, with limited penetration into smaller research institutions. The primary barriers are compute access, data infrastructure, and training in AI-specific research methodologies. Platforms that address these barriers through cloud-based access, pre-trained models, and educational resources will capture significant market share.

Q: How reliable are AI predictions for materials discovery compared to traditional computational methods? A: AI predictions for materials stability and properties have improved significantly, with validation rates reaching 50 to 55% for crystal structure predictions and 60 to 70% for molecular property predictions in well-studied chemical spaces. These rates are comparable to or better than density functional theory (DFT) calculations for screening purposes, at a fraction of the computational cost (seconds per prediction versus hours per DFT calculation). However, AI models perform less reliably for compositions far from their training data, and independent experimental validation remains essential before drawing scientific conclusions or making investment decisions.

Q: What are the most promising AI-for-science applications for sustainability? A: The highest-impact applications are: (1) discovery of novel carbon capture sorbents and membranes, where AI has already identified candidates with 30 to 50% higher CO2 selectivity than existing materials; (2) design of next-generation battery electrolytes and electrode materials for energy storage; (3) optimization of catalysts for green hydrogen production and CO2 utilization; (4) engineering of climate-resilient crop varieties through genomic prediction; and (5) accelerated screening of PFAS remediation compounds and safer chemical alternatives. Each of these applications benefits from large existing datasets, well-defined target properties, and clear commercialization pathways.

Q: How should product teams design AI tools for scientists who are skeptical of black-box models? A: Scientist adoption depends on three design principles: (1) transparency through interpretable predictions that show which molecular features or data points drove the AI's recommendation, not just a score or ranking; (2) uncertainty quantification that clearly distinguishes high-confidence predictions from speculative ones, enabling scientists to allocate experimental resources effectively; and (3) integration with existing workflows, including support for standard file formats, laboratory information management systems (LIMS), and electronic lab notebooks. Products that treat AI as a collaborator augmenting scientific judgment, rather than replacing it, achieve significantly higher adoption rates among research teams.

Q: What compute infrastructure is needed to run AI-for-science workloads effectively? A: Requirements vary dramatically by application. Inference (running trained models to generate predictions) can often be performed on modest GPU hardware or even CPUs for smaller models, at costs of $0.01 to $0.10 per prediction. Fine-tuning foundation models for specific applications requires 4 to 8 GPUs for several days, costing $5,000 to $50,000 in cloud compute. Training frontier models from scratch requires hundreds to thousands of GPUs and costs $1 million to $15 million. For most research teams and product companies, the practical strategy is to leverage pre-trained open-source foundation models and fine-tune them on domain-specific data, avoiding the capital requirements of training from scratch.

Sources

• PitchBook. (2025). AI for Science: Venture Capital Investment Report, Q4 2025. Seattle: PitchBook Data.
• Merchant, A., et al. (2023). "Scaling Deep Learning for Materials Discovery." Nature, 624, 80-85.
• Jumper, J., et al. (2021). "Highly Accurate Protein Structure Prediction with AlphaFold." Nature, 596, 583-589.
• McKinsey Global Institute. (2025). AI-Accelerated R&D: Economic Impact Assessment. New York: McKinsey & Company.
• US Food and Drug Administration. (2025). Guidance on the Use of Artificial Intelligence and Machine Learning in Drug Development. Silver Spring, MD: FDA.
• National Science Foundation. (2025). National Artificial Intelligence Research Institutes: Annual Report. Alexandria, VA: NSF.
• Szymanski, N.J., et al. (2023). "An Autonomous Laboratory for the Accelerated Synthesis of Novel Materials." Nature, 624, 86-91.