Enveda Is Launching the Next AlphaFold Contest. This Time, to Solve the Structures of 90% of Biological Molecules That No One Can Identify.

CASMI 2026 invites teams to identify molecules from unpublished mass spectra and invites machine learning researchers who have never worked in chemistry to solve one of the biggest gaps in our understanding of how life works.

BOULDER, Colo.--(BUSINESS WIRE)--Enveda launched CASMI 2026 this week, a 90-day machine learning and AI competition hosted on Kaggle in which teams identify a molecule's chemical structure from its mass spectrum. The contest revives the Critical Assessment of Small Molecule Identification, a blind challenge founded in 2012, inspired by CASP, the protein-structure prediction competition that produced AlphaFold. CASMI 2026 is the first challenge since 2022 and the first designed to reach the broader machine learning community.

Mass spectrometry can detect thousands of chemical signals in a drop of blood or in a leaf extract. Yet in a typical untargeted study, only about 10% of those signals can be matched to a known molecule.¹ The remaining 90% are measured, recorded, and never identified, leaving vast amounts of biology and potentially new medicines unexplored.

“We can detect far more of nature’s chemistry than we can identify,” said Viswa Colluru, Ph.D., Founder and CEO of Enveda. “That leaves a lot of the life sciences out of reach. With CASMI, we’re giving researchers completely new data and a rigorous test of their ideas. Better tools for identifying these molecules could expand our understanding of biology and our ability to discover new medicines.”

Why a blind challenge?

Machine learning and frontier AI have rapidly advanced scientific discovery, from predicting protein structures to proving long-standing mathematical problems. In this environment, what advances the field the fastest are well-constructed experimental datasets and benchmarks to measure progress against. Blind challenges, in which the answers are withheld until the contest ends, have been the most reliable source of truth.

In the field of mass spectrometry, blind challenges are rare because they require spectra that no one has seen before, which means acquiring new data on hundreds of molecules, including compounds that are difficult to obtain or have never been isolated or characterized. Enveda generated the CASMI 2026 test set on the isolation and profiling platforms it built for drug discovery: about 400 molecules represented by roughly 2,500 mass spectra, none of which have been published. The molecules are real or potential natural molecules and their analogs. Winning methods will be open-sourced, and the test will be a persistent benchmark for the field.

Participants will be able to train their models using approximately 2.5 million spectra from 275,000 molecules, combining existing open libraries with Enveda-180, an open dataset recently released by Enveda containing approximately 1.16 million spectra from 184,330 synthetic small molecules (link).

An invitation to solve one of the major outstanding challenges in chemistry

CASMI 2026 helps to grow PRISM, Enveda’s foundation model developed with Microsoft and trained on more than one billion mass spectra — the largest AI system ever built for identifying molecules from nature — and how Enveda finds drugs. Last month, the company reported positive Phase 1 results for ENV-308, the first oral candidate medicine designed from the chemistry of exercise; earlier this year, ENV-294, a once-daily oral therapy for atopic dermatitis, produced an 85% mean reduction in disease severity in a Phase 1b study. Building PRISM taught Enveda what the field lacks: open training data at scale and a benchmark that can tell a real advance from a model that has quietly seen the answers.

“There are thousands of AI and machine learning researchers who have never heard of mass spectrometry,” said David Healey, Ph.D., Founding Data Science Fellow at Enveda and lead organizer. “But the methods they develop daily across the tech world could advance one of the most exciting challenges in science and continue to revolutionize our understanding of biology.”

CASMI 2026 Competition Details

CASMI 2026 opened Sept. 14, 2026, and will close Dec. 14, 2026.

The competition is free to enter and will award $50,000 in total prize money across the top five teams, including $16,000 for first place.

The competition is hosted on Kaggle and organized by Enveda, with support from CASMI founders Emma Schymanski and Steffen Neumann as well as the Kaggle organizers.

Hosted at: https://www.kaggle.com/competitions/enveda-CASMI26-molecule-id-mass-spectra/

Contact: [e-mail]

About Enveda

Enveda is a biotechnology company that learns from life’s chemistry to create better medicines faster. Enveda uses AI-powered tools to identify and characterize molecules produced by living organisms, the vast majority of which have never been explored by science, creating a database of chemical biodiversity: the library of life. By growing, organizing, translating, and searching this unique library, Enveda learns from life’s evolved solutions to address today’s most pressing medical needs. For more information, visit enveda.com.

Forward-looking statements

This press release contains forward-looking statements, including statements regarding the potential safety and therapeutic benefits of ENV-294 and ENV-308, and the anticipated milestones for our clinical programs. These statements are based on management’s current expectations and are subject to substantial risks and uncertainties that could cause actual results to differ materially. These risks include the inherent uncertainties of clinical development, the risk that preclinical data may not predict clinical results, and regulatory review processes. ENV-294 and ENV-308 are investigational agents and are not approved for use by any regulatory authority. Enveda undertakes no obligation to update these statements.

¹ Gauglitz JM, West KA, Bittremieux W, et al. Enhancing untargeted metabolomics using metadata-based source annotation. Nat Biotechnol. 2022;40:1774–1779. doi:10.1038/s41587-022-01368-1.