Drug development is a long, risky and wildly expensive game. Nine in ten drug candidates fail in clinical trials, often after years of preclinical work and hundreds of millions in investment. Now, a new generation of AI-native biotechs is trying to change the math — and they’re beginning to show results.
Clinical trials are the most failure-prone and costly part of the process. If AI can improve candidate selection, speed up patient enrollment or reduce trial errors, it could shave years off development timelines and help more drugs reach patients who need them.
Startups like Recursion, Inceptive, Cradle and Deep Genomics are embedding AI across the pipeline, from target discovery and molecule design to trial optimization.
“The TechBio approach to drug discovery begins with data, and we have a massive data advantage,” Recursion CTO Ben Mabey said. “We’re developing drugs at more than twice the speed to lead candidate, and producing molecules with 10x the efficiency of traditional pharma.”
Recursion’s platform, the Recursion Operating System (RecursionOS), integrates over 65 petabytes of proprietary biological data and uses foundation models like Phenom-2 and MolPhenix to simulate how compounds will behave before anything is synthesized in a lab.
In its REC-1245 program, now in Phase 1/2 trials, Recursion used an early version of RecursionOS to find a safer way to target a key cancer pathway — progressing from a novel biological insight to a viable preclinical candidate in just 18 months, less than half the industry average.
“This is a concrete example of how we are able to leverage AI throughout the discovery and development process,” Mabey said.
Deep Genomics, meanwhile, is focusing on RNA therapeutics and applying foundation models to a broader slice of biology. “Our platform brings together models like BigRNA, REPRESS, and DeepADAR that can predict outcomes directly from sequence data,” said founder and CIO Brendan Frey. “This lets us identify disease-driving mutations and design RNA therapeutics with unprecedented precision.”
Frey argued that foundation models represent a break from the past. Traditional AI approaches targeted narrow slices of the pipeline, like protein folding or small-molecule interactions, whereas Deep Genomics’ models are trained on trillions of datapoints across DNA, RNA, proteins and systems biology.
That means the platform can weigh multiple constraints simultaneously: identifying suitable targets, predicting mechanisms of action, ensuring safety and tolerability and even accounting for manufacturability. “With a single foundation model platform, we can intelligently discover new targets, mechanisms and molecules — including those we didn’t know to look for,” he said.
AI-driven drug development hinges on more than clever algorithms. You need scale in compute, data and lab throughput.
Recursion runs up to 2.2 million experiments per week in its automated labs and uses its in-house supercomputer, BioHive-2 (built with NVIDIA), to train models on terabytes of microscopy and transcriptomic data. This powers tools like Boltz-2, developed with MIT, which predicts both protein structure and binding affinity with accuracy approaching physics-based simulations — but 1,000 times faster.
Deep Genomics has built a similarly integrated setup, combining high-performance compute with pipelines that connect proprietary and public datasets to its models. “Our infrastructure allows seamless collaboration between machine learning, computational biology and experimental biology teams,” Frey said. “That enables rapid iteration from model development to in vitro validation and, ultimately, therapeutic design.”
AI isn’t just speeding up discovery — it’s also reshaping clinical development. Startups are using ML to select better trial populations, design adaptive protocols and model disease progression in silico. That can cut costs, improve statistical power and reduce unnecessary patient exposure.
To help drive adoption, Recursion recently joined the Council for Responsible Use of AI in Clinical Trials, alongside Sanofi and Velocity Clinical Research, to establish industry standards and define meaningful metrics for AI’s value across the trial lifecycle. Deep Genomics, for its part, works with pharmaceutical partners to slot its foundation models into early R&D workflows, tailoring predictions to specific disease areas. Frey said the company validates its AI-designed candidates through experimental assays and in vivo studies, and has made some tools, like REPRESS, available for academic use.
The ecosystem is maturing quickly. Cloud platforms can now handle AI workloads with biological-scale data. Open-source tools and public omics datasets are proliferating. And the transformer architecture behind large language models is being repurposed for protein and phenotypic modeling.
But AI isn’t a panacea. Regulatory frameworks still lag behind the pace of innovation, and models are only as good as the data they’re trained on. Biology is complex, noisy and context-specific — even the best simulations can miss real-world effects.
What success looks like:
Faster progression from concept to clinical trial
Fewer failed candidates through early digital triage
Smarter patient targeting and enrollment
Better understanding of disease mechanisms — and drug effects
AI is giving drug developers a new kind of leverage. With the right infrastructure, data and scientific grounding, it’s not just possible to speed up R&D — it’s starting to become the default.
“We’re seeing our AI-driven predictions and AI-designed drugs already having a real impact on patients in trials,” Mabey said. “That’s enormously motivating. Every version of our platform is smarter than the last — and the quest for a virtual cell, even a virtual patient, feels like a real possibility.”
Frey was equally bullish but noted the urgency: “What keeps us up at night is knowing how critical this work is. Patients can’t wait. That’s why we’re growing our team rapidly — to bring these AI-designed medicines to the people who need them most.”