Nextflow Frontier Develops Discovery System to Solve Science's Most Complex Problems with Breakthrough Results

 The new closed-loop discovery system outperformed AI systems from OpenAI and Anthropic on a public biology benchmark, advancing scientific discovery through autonomous experimentation at scale.

NEW YORK, Nov. 12, 2026 /PRNewswire/ -- Unveiled at the recent Nextflow Summit, Nextflow Frontier is a recursive, closed-loop discovery system for the unsolved problems in life sciences. In a head-to-head evaluation, Nextflow Frontier outperformed systems from OpenAI and Anthropic on BioML-bench, a public benchmark built from published laboratory experiments spanning drug discovery, biomedical imaging, single-cell omics and protein engineering. Across 22 of its 24 tasks, Nextflow Frontier surpassed an average of 96% of leaderboard entries, compared with 75% for OpenAI's Rosalind and 66% for Anthropic's Claude Science.

Additionally, against previously published results from other AI agents on BioML-bench, Nextflow Frontier leads across all 24 tasks, performing better than 92% of leaderboard entries on average, compared with 75% for the best-performing published agent. The results demonstrate the potential of AI to go beyond generating answers to independently exploring, testing and refining new scientific approaches, opening new possibilities for drug discovery and medical research.

Given a problem and a measure of success, Nextflow Frontier forms ideas, turns each into a real Nextflow pipeline and runs thousands in parallel, scoring and validating the results in silico, then learning and running again. Unlike other systems that simply generate answers or recommendations, it delivers novel, working scientific models and the reproducible methods behind them. Scientists guide the discovery process, creating new scientific capabilities they can validate, deploy and build upon.

Built on Nextflow's decade-long foundation as the open-source home of computational biology, Nextflow Frontier runs on infrastructure already in place across the industry. Fifteen of the top twenty pharmaceutical companies run the platform, and both Anthropic and OpenAI have built on Nextflow, with Anthropic integrating Nextflow skills into Claude for Life Sciences and OpenAI building its Rosalind Workbench on nf-core pipelines.

The breakthrough lies in Nextflow Frontier's ability to sustain scientific experimentation over long durations of continual optimization. Across the 22 comparable tasks, Nextflow Frontier searched for a median of 62 hours per task, compared with nine minutes for Rosalind and six minutes for Claude Science. This extended discovery process enables Nextflow Frontier to explore hundreds of potential approaches, learning from each result and pursuing promising scientific methods that might otherwise go unexplored.

Designed to work across different scientific problems and AI models, Nextflow Frontier adapts its approach to each research challenge, evaluating its solutions against previously unseen data.

Results include:

• Drug discovery: Ranking small molecules by their likelihood of hitting EGFR, a cancer target, Nextflow Frontier conducted 820 scored attempts over 80 hours, independently developing and combining simple chemistry models that train and run on a laptop in seconds. The resulting model achieved a higher score than the leaderboard's billion-parameter neural network and both OpenAI's Rosalind and Anthropic's Claude Science, demonstrating how sustained experimentation can uncover simpler, more effective scientific approaches.
• Drug metabolism: In a challenge to predict how quickly the human liver breaks down a drug, Nextflow Frontier independently identified and incorporated additional data from rat liver measurements in the same public study, going beyond the benchmark's supplied data to develop its own scientific approach. The result demonstrates Nextflow Frontier's ability to identify new paths to solving complex problems rather than simply executing predefined methods.
• Biomedical imaging: Forecasting lung-function decline in pulmonary fibrosis from one baseline CT scan, a first breathing test and basic clinical details, Nextflow Frontier outscored a retrained copy of the competition-winning method on the same held-out patients by measuring patterns of lung density instead of training its own image-reading neural networks.
• Protein engineering: Predicting how inserting or deleting one amino acid changes a protein's stability, it fine-tuned a protein language model 80 times smaller than the one behind the leaderboard's leading method and kept pace with it.
• Single-cell omics: Nextflow Frontier identified different types of kidney cells from a previously unseen donor more accurately than the leading published method. It also improved predictions of proteins found on cell surfaces by accounting for differences in how data was collected across laboratories, outperforming the leading published approach when tested on the same cells.

"We are entering an entirely new era of scientific discovery. For more than a decade, Nextflow has been the trusted foundation powering the world's leading computational biology research. Now, we're building on that foundation to move beyond executing scientific workflows to a system that can independently reason, experiment and adapt its approach to complex scientific problems," said Evan Floden, CEO and Co-Founder of Nextflow. "Where existing AI methods follow a predefined approach, Nextflow Frontier can explore thousands of potential approaches and pursue novel ones that no one told it to try, uncovering methods that weren't obvious to experts, and opening new routes to progress on problems science has yet to solve. This isn't simply faster science. It's an entirely new frontier for what science and medicine can achieve."

Beyond public benchmarks, Nextflow Frontier is being applied to proprietary scientific challenges. ALLOX, a drug-discovery company specializing in deep mutational scanning of proteins, is using Nextflow Frontier to develop models that predict mutation effects on protein-protein interactions and guide future experiments.

"We gave Nextflow Frontier one of the hardest problems in our field: predicting how mutations change protein binding affinity. On our own held-out data, the model it built outperformed every published method we tested, including recent deep learning approaches," said André Faure, Co-Founder and CTO of ALLOX. "The scale of experimentation it enables changes what's possible for a company like ours."

Nextflow Frontier is designed for close scientific partnerships with life sciences companies, embedding forward-deployed scientists alongside customer research teams to define problems, guide discovery and accelerate results. It runs securely within each customer's existing infrastructure, reading data in place and supporting their preferred models, with zero data egress and no training on customer data.

Learn more at nextflow.com/frontier.

About Nextflow
Nextflow is the most widely adopted workflow system in bioinformatics, cited in more scientific publications than any other. The open-source software lets scientists turn code written in any language, from Python and R to Bash, into scalable, portable and reproducible pipelines that run on any infrastructure. It powers research at academic institutes, biotech companies and pharmaceutical organizations worldwide.

SOURCE Nextflow Frontier