AI for Science 2026: The Global Race for Scientific Discovery

5.  Institution Spotlights

5.1  Anthropic · Claude Science Workbench

Anthropic launched Claude Science (Beta) via an online event on June 30, 2026. It is positioned as an AI-native research environment built on existing Claude models, not a new model. Attendees included Anthropic board member and Novartis CEO Vas Narasimhan, Bristol Myers Squibb CEO Chris Boerner, and Genentech Head of Research Aviv Regev.

Key features:   60+ scientific databases; domain-specific toolkits for genomics, protein structure, and chemistry; multi-agent architecture (coordinator agent + specialist sub-agents); auditable research artifact generation.

Access:   Beta open to Pro/Max/Team/Enterprise subscribers. Project grant program: 50 slots, up to $30K compute credits each.

5.2  Singapore AI for Science Initiative

Singapore's National Research Foundation (NRF) leads the S$120M AI4S Initiative, supporting four flagship NUS projects in coordination with the Nobel Turing Challenge Initiative. Milestones: AI4Science International Conference (July 2024); AI4X 2025 conference at NUS; Horizon Europe Complementary Fund signed (December 2025). Website migrating to NRF official platform.

6.  RobotToday Analysis: Three Observations and Structural Analysis

Observation 1: Automation levels are rising, but "automated execution" and "autonomous discovery" remain distinct

Live demonstrations at WAIC 2026 from XtalPi, DeepModeling, and Tianwu carried specific technical details. However, "autonomous discovery" (independent hypothesis generation) and "automated execution" (workflow automation) are distinct concepts. Reporting should maintain this distinction.

Observation 2: US and China have developed largely independent AI4S ecosystem paths

The US (DeepMind, Anthropic, Meta) and China (Shanghai AI Lab, DeepModeling, XtalPi) have each built relatively self-contained AI4S ecosystems with low cross-border technical dependency. This will have medium-term implications for international data-sharing agreements and academic collaboration frameworks. The US pathway is primarily driven by private capital and big-tech monetization (Anthropic subscription workbench; Isomorphic Labs commercial licensing); China's pathway is built on a state-backed consortium model (Shanghai AI Lab as public infrastructure; DeepModeling, XtalPi commercializing on top). A key caveat: both pathways remain highly dependent on the same global pool of scientific literature for model pre-training—meaning a continued decline in code reproducibility rates would affect both ecosystems simultaneously.

Observation 3: Pharma and life sciences are the most active domain for commercial validation

Isomorphic Labs, Claude Science, BioMap, and XtalPi have all concentrated commercial deployment in the upstream pharmaceutical pipeline. The first disclosures of AI-tool-discovered drug candidates entering clinical stages may occur in 2026–2027.

Risk Flag 1: Reproducibility challenges may intensify under AI-driven publication cycles

Fewer than 30% of AI science papers provide accessible code. MIT's 2025 retraction of an AI productivity paper illustrates that data integrity challenges extend to AI4S research itself. The tension between fast-publication culture and academic validation is particularly acute in this field.

Risk Flag 2: Data governance gaps may affect the sustainability of global research collaboration

Scientific data from developing nations is being used to train commercial AI systems without adequate governance frameworks. The International Science Council has formally raised this concern. Unresolved, it may reduce participation in cross-regional scientific data sharing.

7.  Leading US Research Universities: AI4S Case Studies

Work from US research universities concentrates in three directions: engineering of protein and molecular generative models, construction of physical autonomous laboratories, and policy engagement. All cases draw from published papers or institutional public materials.

Massachusetts Institute of Technology (MIT) — Protein Binding Prediction and Drug Discovery Tools

Boltz-1 / Boltz-2  MIT Jameel Clinic + CSAIL × Recursion  |  2024–2025

MIT Jameel Clinic and CSAIL, in collaboration with Recursion, released the Boltz series. Boltz-1 (late 2024): open-source protein structure prediction benchmarked against AlphaFold 2. Boltz-2 (June 2025): adds binding affinity prediction to the same framework, outputting structure and affinity on a single GPU in approximately 20 seconds—approximately 1,000× faster than free-energy perturbation (FEP) methods (note: the speedup compares single-GPU inference time against full multi-CPU FEP simulation time; the two methods differ in computational scope and downstream validation requirements), with binding affinity performance approximately double that of prior methods on benchmarks. Both models are open-source.

BoltzGen  MIT Jameel Clinic  | November 2025

A generative protein design model built on the Boltz framework. Unlike structure prediction, BoltzGen generates novel protein sequences designed to bind specified biomolecular targets, producing candidates ready for downstream experimental validation in the drug discovery pipeline.

AI-Assisted Crohn's Disease Antibiotic Mechanism Analysis  MIT CSAIL  | 2024–2025

CSAIL researchers used AI models to analyze the selective mechanism of a narrow-spectrum antibiotic on the gut microbiome of Crohn's disease patients, identifying molecular pathways that target harmful bacteria without affecting beneficial flora. The findings provide a computational basis for personalized microbiome-targeted interventions.

Carnegie Mellon University (CMU) — Autonomous Laboratory Infrastructure

Coscientist  CMU Department of Chemistry + ML  | Gomes et al., Nature, December 2023 (doi:10.1038/s41586-023-06792-0)

Coscientist is the first peer-reviewed AI system demonstrated to autonomously plan, design, and execute chemical experiments. Developed by the team of Assistant Professor Gabe Gomes, it uses GPT-4 and Claude to process natural language experimental instructions and interfaces with robotic equipment for physical execution. Without pre-programming, the system completed a Suzuki coupling reaction—whose human inventors received the 2010 Nobel Prize in Chemistry—in minutes, with results consistent with manually executed reference experiments.

AI Science Foundry  Carnegie Mellon University  | Ongoing

The AI Science Foundry is a cluster of physical autonomous laboratories integrating AI, automation, robotics, and computational workflows across chemistry, biology, and hard materials, supported by large-scale storage and cloud-programmable laboratory facilities. Researchers submit experiment designs via remote interfaces; the system executes, acquires data, and returns analysis. CMU positions it as the largest US university co-located autonomous laboratory cluster, with the goal of reducing innovation-to-application timelines from years to months.

Stanford University — Genomic Foundation Models and Targeted Materials Discovery

Evo 2 (DNA Language Model) Arc Institute + Stanford + UC Berkeley |  February 2025

Evo 2 is a DNA language model trained on 9 trillion base pairs of genomic sequence covering nearly all catalogued life forms. 40 billion parameters. Stanford researcher Brian Hie was among the core contributors. The model supports mutation effect prediction and novel antibody sequence generation, released as open-source.

AI-Driven Targeted Materials Discovery  SLAC National Accelerator Laboratory + Stanford  |  2024

Researchers developed an AI method using active learning to reduce the number of experiments required in materials discovery, improving search efficiency in complex design spaces relevant to quantum computing, climate applications, and drug design. Integrated with self-driving experiment frameworks supporting automated parameter adjustment for robotic setups; currently in laboratory validation stage.

UC Berkeley — Scientific Foundation Models Research

BAIR + DL4SCI 2026  UC Berkeley + Lawrence Berkeley National Laboratory

Berkeley AI Research (BAIR) comprises 50+ faculty and 300+ doctoral students across ML, computer vision, NLP, and robotics. DL4SCI 2026 (Deep Learning for Science), co-organized with Lawrence Berkeley National Laboratory, is a five-day intensive program focused on training, adaptation, and evaluation of scientific foundation models, and on reasoning-centric research workflows. Open to doctoral students and researchers in scientific domains.

8.  US Academic Perspectives and Cautionary Views

Leading US research universities broadly acknowledge the efficiency value of AI4S, but hold explicit reservations on technical reliability, research integrity, and researcher training. The following is based on published papers, institutional statements, and public reporting.

8.1  Institutional Stances

MIT CSAIL — Active participation with policy conditions

MIT CSAIL submitted AI Action Plan recommendations to the US government in 2025, supporting AI for Science as a national priority. Specific recommendations: continued investment in basic research (especially new algorithms and ML architectures); maintaining a healthy mix of open-source and proprietary licensing; strengthening high-school mathematics and CS education. CSAIL also explicitly positioned interpretability, alignment, and safety as deployment prerequisites, not post-hoc review items.

Stanford HAI — Emphasis on human-AI collaboration

Stanford HAI, in its AI Index 2025 Report and public commentary, has explicitly positioned AI-accelerated scientific discovery as conditional on human-centeredness: AI provides computational support while researchers retain control over hypothesis generation and experimental interpretation. HAI frames AI4S as an upgrade to research tooling, not a replacement restructuring of the scientific paradigm.

CMU — System builder, attentive to scaling challenges

CMU's investments in Coscientist and the AI Science Foundry reflect a "build-first" approach. Its public statements describe Coscientist's demonstrated automation capabilities as a structural change in research execution, while noting that the next challenge is extending such capabilities to more complex scientific problems rather than reactions well-covered by existing literature.

8.2  Technical Reservations

"LLMs Are Not Scientists Yet" (2026):   Multiple academic teams, including scholars affiliated with MIT and Stanford, published reviews summarizing lessons from four autonomous AI research attempts. Core conclusion: current LLMs lack the capacity to generate genuinely novel scientific hypotheses; their strength lies in optimization and retrieval within known knowledge frameworks, not in exploring the space of unknown unknowns.

Autonomous pipeline execution failures:   Systematic evaluations of "AI scientist" pipelines have documented instances where systems generate synthetic placeholder data upon execution failure and continue running rather than halting. Hallucinated citations and inflated performance claims in unreviewed outputs have also been recorded; experiment failure rates as high as 42% were observed in some evaluations.

MIT data integrity incident (May 2025):   MIT disavowed a doctoral student paper on AI's productivity benefits in scientific research, citing "no confidence in the provenance, reliability, or validity of the data." The paper had been widely cited. The incident highlights that AI4S-related research is itself subject to data validation challenges, with elevated risk in high-expectations environments.

8.3  Structural Risk: Researcher Deskilling

Several US academics have raised the concern that heavy reliance on AI for automated experiment execution and analysis generation may reduce training opportunities for early-career researchers in core competencies: experimental design, data judgment, and anomaly identification. Referred to as "researcher deskilling," systematic data remain limited, but the concern has been incorporated into multiple academic policy recommendation documents as a risk warranting active monitoring.

Research data sourced from public information as of July 19, 2026: WAIC official releases, Anthropic.com, EU Commission policy, Singapore NRF, Qbitai, 36Kr, TMTPost, and other major media.