.jpg)
The identification of unknown compounds in biological samples is one of the greatest challenges of modern metabolomics. When thousands of molecules are analyzed using tandem mass spectrometry (MS/MS), most remain unidentified because they are not present in spectral libraries. These uncharacterized metabolites and natural products are precisely the ones of greatest interest in areas such as drug discovery, biomarker research, and exposomics. Traditionally, computational methods for structural elucidation required prior knowledge of the molecular formula, a piece of data that is unavailable for genuinely novel compounds and, when predicted, carries significant error.
This is where MARLIN comes into play, an artificial intelligence approach that resolves molecular structures directly from the spectrum, without needing the molecular formula at any stage of the process. MARLIN employs a self-supervised encoder that extracts a molecular fingerprint from the raw peaks of the spectrum, and a block-based diffusion language model that generates candidate structures conditioned solely on that fingerprint and the precursor mass measured by the instrument. A mass-shell constraint ensures each candidate is consistent with the measured mass, without predetermining the atomic inventory, and only those that match exactly in parts per million are accepted. Additionally, it incorporates a symmetric noise objective to absorb encoder errors and a diversity mechanism that prevents candidates from collapsing into a single structure.
Results on the NPLIB1 benchmark demonstrate that MARLIN outperforms any other method evaluated without a known formula, both in exact match accuracy and in structural distance and fingerprint similarity. It is even capable of recovering the correct molecular formula as a byproduct, with a frequency comparable to that of a specialized predictor, but without ever having used one. This breakthrough opens the door to reliable de novo elucidation in the realistic discovery regime, where the molecular formula is unknown.
The application of techniques like MARLIN in business and research environments requires robust technological solutions. Companies working with massive spectrometry data need artificial intelligence for businesses that can be integrated into customized workflows. At Q2BSTUDIO, we offer custom software development services to implement AI models in analytical processes, as well as tailored applications that adapt to the specific needs of each laboratory or R&D department. Furthermore, we deploy these solutions on scalable cloud infrastructure, whether on AWS and Azure cloud services, ensuring performance and security.
Cybersecurity is also a critical aspect when handling sensitive research data or intellectual property information. Therefore, our services include security audits and pentesting to protect systems. Likewise, if the goal is to extract value from the data generated by these analytical techniques, we offer business intelligence services with Power BI and customized dashboards, enabling teams to make data-driven decisions in real time. And for repetitive processes, automation through AI agents and process automation solutions can significantly reduce analysis times.
MARLIN's ability to work without a prior molecular formula represents a qualitative leap. In the near future, we will see how these models are integrated into drug discovery platforms, environmental exposure assessment, and clinical diagnostics. The collaboration between advanced artificial intelligence and custom software development is the key to transforming complex data into actionable knowledge.
