AI-native Drug Discovery and Development

Drug discovery still runs on economics that resist improvement: development timelines that frequently exceed ten years, total costs that often surpass $1–2 billion per approved therapeutic, and attrition that stays stubbornly high in oncology, CNS, and rare disease. Sequential experimentation, fragmented data environments, and manual interpretation are a large part of why. This working paper from the SPARK AI Consortium at the San Diego Supercomputer Center examines what replaces that model — a continuous learning system in which prediction, experimentation, validation, and regulatory evidence connect in near real time — and why the organizations that rebuild drug discovery around it first may compound advantages competitors find difficult to replicate.

What is AI-native drug discovery?

AI-native drug discovery is a research operating model in which AI systems sit inside the scientific process rather than beside it — generating hypotheses, designing molecules, planning experiments, and interpreting clinical and regulatory evidence. Unlike AI-assisted workflows, the models retrain continuously as new experimental data arrives.

Highlights

• 10+ years and $1–2B per approved therapeutic — the economics AI-native R&D is designed to change
• 4 technology layers converging at once: foundation models, semantic infrastructure, autonomous labs, hyperscale compute
• 11-level semantic architecture, mapped in both an executive view and an operational view
• 5 constraints that could slow adoption — starting with the one most programs underestimate

What’s inside

• Beyond AI-assisted research — why decision-support tools bolted onto sequential workflows are breaking down, and what AI-native discovery architecture looks like when AI sits inside the scientific operating model rather than beside it
• The four-layer stack — foundation models as the reasoning layer, semantic infrastructure as the connectivity layer, autonomous labs as the experimental layer, and hyperscale compute as the infrastructure layer, with the vendor and partnership landscape forming around each
• Autonomous laboratories in practice — how closed-loop “self-driving labs” compress Design-Make-Test-Analyze cycles from weeks or months into continuous learning loops, and the compounding gains in throughput, predictive accuracy, and resource allocation
• The shadow asset: semantic reasoning — why semantic layering is not ontologies sprinkled over existing process flows, but the separation of intent, knowledge, and execution so that changes in scientific understanding propagate without redesigning validated systems
• Regulation moving with the technology — the shift toward evidence-centric, risk-based oversight, cloud-enabled continuous review, and growing regulatory support for in silico simulation, organ-on-chip, and synthetic control methodologies
• Five constraints that could slow adoption — data quality and interoperability, explainability, organizational readiness, interdisciplinary talent shortages, and the prospect that infrastructure concentration produces market concentration
• Seven strategic imperatives — the decisions that separate organizations layering AI onto existing workflows from those re-architecting discovery, development, and regulatory processes around it

Why the traditional drug discovery model is under pressure

For most of the industry’s history, the most valuable pharmaceutical asset was a molecule or a platform. The argument in this paper is that it is becoming the organizational learning system — the connected fabric of data, models, experiments, regulatory evidence, and knowledge that improves every time new data arrives. That reframes the investment question. It is no longer only which assets to fund, but whether your data environment, experimentation capacity, and semantic foundation can support a drug discovery system that gets better on its own. Model architectures and compute have advanced faster than data quality, semantic consistency, and interoperability have. In many organizations the data engineering challenge is now larger than the modeling challenge — which is precisely where the next round of advantage will be won or lost.

Download the Whitepaper

Foundation models. Semantic intelligence. Autonomous labs. One continuously learning R&D system underneath all of it.