待翻譯：Scaling Scientific R&D with AI Supercomputing Infrastructure

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

​Pharmaceutical and life sciences enterprises have proven AI can improve individual stages of discovery, development, and manufacturing — but the industry’s legacy IT infrastructure was never built for frontier-scale compute. This is a structural, industry-wide constraint. The economics make the stakes clear. According to the National Institutes of Health, a discovery can take almost 15 years to become an approved drug, with a failure rate exceeding 95 percent and pushing the effective cost per successful drug above $1 billion. A peer-reviewed study on NIH’s PubMed Central found that fewer than 10 percent of drugs entering clinical trials are ultimately approved, with attrition reaching 40–70 percent in Phase 2 alone. That volume of failed data is a latent asset — but exploiting it requires computational scale legacy systems cannot provide. Federal agencies confirm the gap is national, not proprietary. The National Science Foundation stated in August 2026 that access to AI infrastructure remains highly uneven even as AI transforms scientific discovery, and it committed $100 million to new regional AI infrastructure hubs to help close the gap. Meanwhile, the FDA reports its drug-evaluation center has already reviewed more than 500 regulatory submissions containing AI components since 2016, raising the performance and reproducibility bar infrastructure must meet. Where a purpose-built scientific compute has been deployed, returns are measurable. A 2025 NIH-hosted review found that digital twins in continuous manufacturing have improved active pharmaceutical ingredient consistency to 99.95 percent, while AI-linked protein structure models show potential to cut target validation time from months to days. These figures describe an industry where high attrition, unexploited historical data, rising regulatory scrutiny, and an acknowledged national compute shortfall converge on one conclusion: pharmaceutical IT must evolve from a transactional support function into purpose-built scientific compute capability. Emerj’s Matthew DeMello in conversation with Thomas Fuchs, Chief AI Officer at Eli Lilly, discusses how AI has become a scientific infrastructure problem in pharma; the industry should build purpose‑built compute to unlock meaningful discovery, model accuracy, and manufacturing impact. This article examines three core insights that matter most for pharmaceutical and life sciences leaders as AI becomes a central driver of scientific work: Unpublished failure data for higher-accuracy drug prediction: Capture and structure failed-experiment data into training pipelines, so prediction models learn from the full outcome space instead of only the fraction that succeeded. Purpose-built molecular models for real drug-design capability: Deploy molecular generative, physics-based, and diffusion models alongside LLMs, so biological complexity is modeled by architectures built for it rather than forced through language tools. Compute as core Scientific infrastructure, not IT overhead: Fund and govern supercomputing as a scientific instrument owned by R&D, so compute capacity scales with what discovery requires instead of what IT budgets allow. Listen to the full episode below: Episode: Scaling Scientific R&D with AI Supercomputing Infrastructure — with Thomas Fuchs of Eli Lilly ​ Guest: Thomas Fuchs, Chief AI Officer at Eli Lilly Expertise: Artificial Intelligence, Machine Learning, Computational Pathology, Medical Informatics Brief Recognition: Thomas Fuchs is Senior Vice President and Chief AI Officer at Eli Lilly and Company, where he leads AI initiatives across the organization. Previously, he was the inaugural Chair of the Department of Artificial Intelligence and Human Health and Dean of Artificial Intelligence and Human Health at the Icahn School of Medicine at Mount Sinai, where he also directed the Hasso Plattner Institute for Digital Health. He is the founder and former Chief Scientist of Paige, an AI-focused company in computational pathology, and serves on the board of GeneDx. Fuchs is also an Adjunct Professor at Mount Sinai and holds a Doctor of Sciences in Machine Learning from ETH Zürich, with postdoctoral research in Computer Vision at Caltech. Unpublished Failure Data for Higher‑Accuracy Drug Prediction Thomas Fuchs highlights that pharmaceutical R&D generates far more failed experiments than successful ones, and that these failures contain the strongest signal for improving molecular prediction. He argues that most AI systems in the industry are trained on an unrealistically narrow slice of outcomes because only successful results make it into the scientific literature. For leaders, the strategic shift is to elevate failed‑experiment data into a primary training resource rather than treating it as discardable noise. Fuchs explains why negative results are the strongest learning signal in discovery: “If you only train an AI on the positive outcomes — the results that get published — you’re giving it a tiny and misleading slice of reality. For every molecule that worked, we had millions that failed, and those failures are exactly where the real learning signal is. When you tap into decades of negative results, you can build models that understand what doesn’t work, and that’s what lets you design molecules you wouldn’t have thought of before.” – Thomas Fuchs, Chief AI Officer at Eli Lilly For organizations, the operational guidance is: Integrate failed assays, non‑binding attempts, and toxicity outcomes into model training, so systems learn from the full outcome space rather than the sliver that succeeded. Use negative‑trained models to prune non‑viable candidates earlier, reducing wet‑lab burden and lowering late‑stage attrition. Leverage proprietary failure data as a competitive advantage, since only large enterprises possess the depth required to outperform models trained solely on public literature. By elevating unpublished failures into a first‑class training resource, pharmaceutical leaders can build prediction systems that reflect the true complexity of discovery, and materially improve the reliability of AI‑driven molecule design. Purpose‑Built Molecular Models for Real Drug‑Design Capability Thomas Fuchs makes a sharp distinction that many pharmaceutical leaders still blur: large language models are useful inside discovery workflows, but they are not, and cannot become drug‑design engines. His point isn’t about hype management; it’s about biological complexity. A single cell operates at a level of physical, chemical, and temporal detail that language cannot encode. When organizations try to force drug‑design tasks through LLMs, they constrain themselves to the representational limits of text rather than the physics of biology. Fuchs describes how Lilly approaches the problem differently. Language models orchestrate work, handle documentation, and support regulatory Q&A, but the actual design and prediction engines are built on architectures meant for molecular behavior: generative models, diffusion models, physics‑based systems, and nucleotide/RNA models. These tools operate in spaces where binding affinity, toxicity, conformational change, and reaction dynamics can be modeled directly — not translated into sentences. He captures the limitation of language‑based approaches in a way that’s both technically honest and strategically useful for leaders: “The complexity of a single cell goes far beyond what human language can even describe. You would constrain yourself if you constrained yourself to language‑based models. The models that really drive design are different — molecular models, diffusion models, generative flow models — they can come up with new molecules.” – Thomas Fuchs, Chief AI Officer at Eli Lilly This distinction leads directly to a set of practical decisions for executives: Use LLMs where language is the substrate — orchestration, summarization, regulatory support, documentation, and workflow automation. Use molecular and physics‑based models where biology is the substrate — prediction, generative design, optimization, and exploration of chemical space. Architect discovery pipelines so each model class operates where it is strongest, rather than forcing a single model type to do everything. According to Thomas, this division of labor is not a technical nuance; it’s a strategic requirement. Leaders who treat LLMs as universal engines will hit hard ceilings in accuracy and scientific validity. Leaders who pair LLMs with purpose‑built molecular models can expand the frontier of what their discovery teams can actually design. Compute as Core Scientific Infrastructure, Not IT Overhead Thomas Fuchs frames compute as a scientific instrument, a shift that fundamentally changes how pharmaceutical organizations should govern and invest in AI. He points out that Lilly’s supercomputing strategy is not about faster experiments or generic performance gains; it’s about expanding the horizon of what scientists can even attempt. With the new system, researchers can train frontier‑scale models, run physics‑based simulations at meaningful resolution, and explore chemical and genetic spaces that were previously inaccessible. What stands out in his explanation is how compute changes the scope of scientific thinking. In the past, teams were constrained to small models and limited datasets because infrastructure couldn’t support anything larger. With a thousand B300 GPUs, that ceiling disappears. Fuchs describes the supercomputer as “a telescope” — a tool that lets scientists see further, ask deeper questions, and design models that were impossible under legacy constraints. This metaphor is not rhetorical; it reflects how compute directly shapes the ambition and creativity of R&D teams. He also emphasizes that compute delivers value far beyond discovery. In manufacturing, for example, Lilly used AI to optimize the drying process for APIs — a seemingly mundane step that resulted in millions of additional doses reaching patients faster. These kinds of gains don’t come from abstract innovation; they come from treating compute as a core capability owned by R&D, with clear metrics tied to patient impact, operational efficiency, and scientific rigor. Fuchs’s perspective gives leaders a different lens for evaluating infrastructure investments: Compute determines the ceiling of scientific ambition. Without frontier‑scale capacity, teams are forced to use smaller models, narrower datasets, and more limited exploration of biological and chemical complexity. Compute creates value far beyond discovery. Manufacturing, development, regulatory, and commercial functions can generate measurable returns when advanced AI systems have access to sufficient computational scale. Compute should be evaluated as a scientific capability, not solely as infrastructure. Like a telescope or microscope, greater computational power expands the questions researchers can ask and the models they can realistically build. This is the strategic shift Fuchs is pushing toward: compute is no longer a background technology. It is a scientific capability that determines how far researchers can push the boundaries of discovery, how effectively organizations can leverage their data, and how quickly AI innovations translate into measurable business and patient impact.