How artificial intelligence, automation, data-first platforms, and new operating models will reshape drug discovery, drug development, and portfolio decision-making over the next decade.
Global R&D Pharma cost context
The arithmetic of modern discovery is brutal. €2.5 billion is the cost of bringing a single drug to market in modern drug development when one accounts for the failures upstream (Joseph di Masi). With only one in ten programs reaching the market after clinical trials, the story across pharma, materials, and specialty chemicals is the same: long cycles, high failure rates, and rising complexity.
While many diagnose the problem as decentralized ecosystems, talent gaps, patent cliffs, the real fix lives inside the experiments and scientific workflows. The bottleneck is the gap between running one experiment and deciding what to run next. That decision loop, often days or weeks of manual interpretation and discussion, is where time and money are lost. Closing this loop through connected data, building Self-Driving Labs® that augment scientists with AI and robotics to solve more
complex problems and accelerate R&D from years to weeks is not a 2030 ambition. It is the Moore’s Law of Discovery and brings exponential scientific discovery in action now.
While many diagnose the problem as decentralized ecosystems, talent gaps, patent cliffs, the real fix lives inside the experiments and scientific workflows. The bottleneck is the gap between running one experiment and deciding what to run next. That decision loop, often days or weeks of manual interpretation and discussion, is where time and money are lost. Closing this loop through connected data, building Self-Driving Labs® that augment scientists with AI and robotics to solve more
complex problems and accelerate R&D from years to weeks is not a 2030 ambition. It is the Moore’s Law of Discovery and brings exponential scientific discovery in action now.
Signs of stabilization in clinical productivity
Despite structural cost pressures, recent data suggests operational efficiency may be stabilizing. According to IQVIA’s Global Trends in R&D 2025, clinical productivity showed measurable improvement in 2024, with total program durations declining from a peak of 10.1 years in 2022 to 9.3 years. This response likely reflects the growing adoption of AI, innovative trial designs, and data-driven decision frameworks that Atinary is helping to standardize.
Trend 1 – Emerging Biopharma is leading early innovation
Early-stage innovation is no longer concentrated inside large pharmaceutical companies. In 2024, emerging biopharma companies were responsible for roughly 85% of new active substances reaching the market. Large pharma is shifting its strategy from internal discovery to becoming an “integrator,” partnering with these agile, EBP teams that are smaller, more agile teams, and often operating with focused budgets and fewer constraints.
This shift changes where competitive advantage lives. It is no longer defined by how many assets sit in a pipeline, but by how quickly and effectively a company can evaluate external opportunities, integrate them into existing workflows, and move them toward meaningful decision points. As the number of external programs grows, so does the operational complexity behind them.
In practice, this is where many organizations start to struggle. Without consistent data structures, shared experimental context, or scalable decision processes, integrating external innovation often slows teams down. What should expand optionality ends up creating friction. The organizations that benefit from this new model are not necessarily those sourcing the most ideas, but those able to learn from them faster.
In practice, this is where many organizations start to struggle. Without consistent data structures, shared experimental context, or scalable decision processes, integrating external innovation often slows teams down. What should expand optionality ends up creating friction. The organizations that benefit from this new model are not necessarily those sourcing the most ideas, but those able to learn from them faster.
Trend 2 – AI readiness is now measurable
Artificial intelligence in the pharmaceutical industry and R&D is no longer a side initiative. It is becoming measurable and, increasingly, comparable. According to CB Insights (2025), companies like Eli Lilly, Merck, and AstraZeneca are already being ranked by their execution capabilities.
This acceleration is not happening in a vacuum. The pressure is structural. Patent cliffs are approaching, pipelines must deliver faster, and projections suggest artificial intelligence could unlock hundreds of billions in annual value. In that context, AI readiness becomes a strategic lever: one that directly influences capital allocation, productivity, and long-term positioning.
Platforms like SDLabs™ reflect this shift. By embedding surrogate models, Bayesian optimization, and agentic workflows directly into the experiment loop, teams don’t need to wait for large transformations to see value. Results emerge from the first campaigns, the system improves with every run, and turning every run into a compounding asset from day one. This is also where Self-Driving Labs® move from concept to practice.
This acceleration is not happening in a vacuum. The pressure is structural. Patent cliffs are approaching, pipelines must deliver faster, and projections suggest artificial intelligence could unlock hundreds of billions in annual value. In that context, AI readiness becomes a strategic lever: one that directly influences capital allocation, productivity, and long-term positioning.
Platforms like SDLabs™ reflect this shift. By embedding surrogate models, Bayesian optimization, and agentic workflows directly into the experiment loop, teams don’t need to wait for large transformations to see value. Results emerge from the first campaigns, the system improves with every run, and turning every run into a compounding asset from day one. This is also where Self-Driving Labs® move from concept to practice.
Trend 3 – From isolated experiments to closed-loop learning systems
R&D value is moving from one-off projects to platforms that learn, or “Scientific Data Factories” like that of the Atinary Lab in Boston. Instead of starting from zero, scientists can feed experimental results, their domain expertise and insights, and data analytics outputs into a persistent infrastructure. That changes the unit of progress: every campaign starts with prior evidence rather than from zero. The result is a learning system where Design-Make-Test-Analyze and Learn (DMTA+L) cycles compound knowledge across discovery and development.
Many labs already have the hardware — liquid handlers, HTE plates, analytical instruments — but lack the surrogate modeling and acquisition-driven optimization layer that turns automation into learning. An optimizer that reads results, updates a surrogate model, and uses acquisition functions to propose the next experiment is what connects these pieces into a true system.
The payoff is both immediate and cumulative: faster iterations, less rework, and the ability to transfer models and protocols across projects so past learnings compound over time. This does not replace scientists—it amplifies them. Scientists still define objectives, review outcomes, and guide strategy, while the platform handles the iterative search and preserves knowledge so teams don’t repeat the same mistakes.
Atinary’s Self-Driving Labs® approach puts these 3 pillars integrated with science:
● Robotics & Lab Automation to perform experiments efficiently, precisely, and 24/7
● Computing power to manage experiments, collect data, and track results
● AI / ML to screen large spaces, learn from experiments, and make decisions.
Many labs already have the hardware — liquid handlers, HTE plates, analytical instruments — but lack the surrogate modeling and acquisition-driven optimization layer that turns automation into learning. An optimizer that reads results, updates a surrogate model, and uses acquisition functions to propose the next experiment is what connects these pieces into a true system.
The payoff is both immediate and cumulative: faster iterations, less rework, and the ability to transfer models and protocols across projects so past learnings compound over time. This does not replace scientists—it amplifies them. Scientists still define objectives, review outcomes, and guide strategy, while the platform handles the iterative search and preserves knowledge so teams don’t repeat the same mistakes.
Atinary’s Self-Driving Labs® approach puts these 3 pillars integrated with science:
● Robotics & Lab Automation to perform experiments efficiently, precisely, and 24/7
● Computing power to manage experiments, collect data, and track results
● AI / ML to screen large spaces, learn from experiments, and make decisions.
Trend 4 – Strategic value: AI, digital tools, and KPIs in Pharma R&D
AI and digital tools are fundamentally changing R&D performance assessment. Beyond “cost per approved drug,” leaders now measure:
● Iteration Economics: Cost per decision loop and cycle time across phases.
● Learning Velocity: Insights generated per experiment.
When metrics are visible, teams prefer investments that tighten decision loops over those that merely increase throughput. Embedding uncertainty-aware, Bayesian optimization inside the experiment loop reduces iteration counts and produces measurable productivity from the first pilot. Over time, datasets compound and raise the baseline for future projects. See Takeda’s use case for a concrete example.
● Iteration Economics: Cost per decision loop and cycle time across phases.
● Learning Velocity: Insights generated per experiment.
When metrics are visible, teams prefer investments that tighten decision loops over those that merely increase throughput. Embedding uncertainty-aware, Bayesian optimization inside the experiment loop reduces iteration counts and produces measurable productivity from the first pilot. Over time, datasets compound and raise the baseline for future projects. See Takeda’s use case for a concrete example.
Trend 5 – AI is unlocking chemical & biological design spaces
Modern discovery faces a search problem. The number of possible molecules for new materials and engineered proteins for vaccines far exceeds what any laboratory can make or test. Computational models shrink that gap by proposing high-quality candidates and ranking them for experimental follow-up. AlphaFold’s protein-structure breakthroughs made targets far more tractable for structure-based design, speeding the path from sequence to a testable hypothesis.
Generative AI models, like VAEs, GANs, and diffusion-style architectures, learn compact representations of molecules and protein sequences. When trained on property data and combined with physics-aware filters, those models produce novel scaffolds and candidates that sit outside existing libraries. Recent work demonstrates target-specific generative pipelines that yielded experimentally validated hits for kinase targets such as CDK2 and delivered promising scaffolds for difficult targets.
Generative AI models, like VAEs, GANs, and diffusion-style architectures, learn compact representations of molecules and protein sequences. When trained on property data and combined with physics-aware filters, those models produce novel scaffolds and candidates that sit outside existing libraries. Recent work demonstrates target-specific generative pipelines that yielded experimentally validated hits for kinase targets such as CDK2 and delivered promising scaffolds for difficult targets.
Use case saving costs with fine outcomes in R+D
In a collaboration with fragrance industry scientists, AI was applied to a constrained and costly problems in industrial chemistry: optimizing a hydroformylation reaction dependent on scarce rhodium catalysts. Instead of navigating the space through traditional trial-and-error, the team explored a 7-dimensional space of ~2.9 billion possible combinations in just 88 experiments, using Bayesian optimization to continuously guide each step.
● Impact: Rhodium usage dropped by up to 30x, and catalyst cost contribution fell by up to 97%.
● Result: Optimization became a learning system that reduced cost, time, and uncertainty simultaneously.
● Impact: Rhodium usage dropped by up to 30x, and catalyst cost contribution fell by up to 97%.
● Result: Optimization became a learning system that reduced cost, time, and uncertainty simultaneously.
Trend 6 – The scientist is still the core value
Nearly half of life sciences organizations report that AI and data skill shortages are their main barrier to transformation, but the issue runs deeper. Many AI tools still depend on specialists to operate, which keeps their impact confined to pilot projects. Bench scientists remain disconnected from the optimization process, while data scientists often lack the experimental context to act effectively.
By 2030, the benchmark will not be whether AI is present, but whether every scientist can independently run, interpret, and iterate on experiments without relying on intermediaries.
This is where the model shifts. The most effective platforms are not those that replace scientists, but those that augment them. Tools like SDLabs bring agentic workflows, Bayesian optimization, and active learning directly into the hands of bench scientists, via a code-free interface, allowing scientists to stay in control while AI handles the complexity of search and iteration. And more importantly, allowing scientists to focus on what they do best: creativity, judgment, and high-level strategy.
By 2030, the benchmark will not be whether AI is present, but whether every scientist can independently run, interpret, and iterate on experiments without relying on intermediaries.
This is where the model shifts. The most effective platforms are not those that replace scientists, but those that augment them. Tools like SDLabs bring agentic workflows, Bayesian optimization, and active learning directly into the hands of bench scientists, via a code-free interface, allowing scientists to stay in control while AI handles the complexity of search and iteration. And more importantly, allowing scientists to focus on what they do best: creativity, judgment, and high-level strategy.
An Integrated Response to Exponential Science
The shift toward Self-Driving Labs® fundamentally changes the economics of discovery. Traditionally, discovery was a linear expense; with Atinary, we are bringing a new paradigm for scientific discovery and scaling.
As our Co-Founder and CTO, Dr. Loïc Roch, describes:
“Today, we are still operating on a 1930s R&D mindset while trying to solve 21st-century problems. Advanced materials currently take up to 20 years to reach the market. That is a bottleneck that threatens our global competitiveness.”
By moving R&D from human-driven to an AI-driven R&D process that is continuously learning and compounding, we enable the pace of progress across chemistry and materials science to accelerate exponentially. By closing the loop between experiment design, physical execution, and learning, we collapse discovery cycles from decades into weeks and unlock the breakthroughs the 21st century demands.
As our Co-Founder and CTO, Dr. Loïc Roch, describes:
“Today, we are still operating on a 1930s R&D mindset while trying to solve 21st-century problems. Advanced materials currently take up to 20 years to reach the market. That is a bottleneck that threatens our global competitiveness.”
By moving R&D from human-driven to an AI-driven R&D process that is continuously learning and compounding, we enable the pace of progress across chemistry and materials science to accelerate exponentially. By closing the loop between experiment design, physical execution, and learning, we collapse discovery cycles from decades into weeks and unlock the breakthroughs the 21st century demands.