Beyond Raw Data: Automating Intact Mass Analysis with Domino and Multi-modal LLMs

In the world of BioPharma R&D, data is abundant, but insights are often buried under layers of complex outputs. One of the most common bottlenecks in our lab was the post-processing of Intact Mass data from Byos. While Byos does an excellent job with the initial analysis, the raw Excel exports often require significant manual intervention before they are ready for cross-functional stakeholders.

To solve this, I’ve developed an automation pipeline and published it as a Domino Model Product, allowing my colleagues to transform raw mass spec data into comprehensive, AI-annotated reports with a single click.

The Problem: The “Data Fog”

A typical plate-based intact mass analysis generates hundreds of protein species and impurities. For an end-user—who might be focusing on process development or formulation—digesting a list of every detected peak is overwhelming. They need to know the “big picture” of their samples, not just the raw intensity of every minor variant.

Step 1: Intelligent Summarization & Classification

The first layer of the automation script focuses on data reduction. Instead of presenting a raw list of peaks, the script categorizes detected species into logical groups:

• Main Protein Species: Highlighting the primary product.
• Specific Impurities: Grouping known variants (e.g., glycosylation patterns, truncations).
• Unknowns: Flagging unexpected peaks for further review.

This classification allows end-users to understand the sample composition instantly without needing to be mass spec experts themselves.

Step 2: Spatial & Statistical Visualization

Visualizing data across a plate is critical because sample distribution is rarely random; patterns often emerge based on plate layout (e.g., edge effects, gradient shifts). My tool generates three primary visualizations:

1. Plate Heatmaps: These provide an immediate “at-a-glance” view of the entire plate. By mapping attributes like purity or concentration to the physical layout, we can quickly identify spatial trends or systematic errors.
2. Violin Plots: These help us understand the distribution and density of specific attributes across different groups or conditions.
3. Native Excel Stacked Bar Charts: I integrated these directly into the output Excel file. They provide a clear view of the relative composition of each sample, showing the ratio of main products to various impurities.

Step 3: “Giving the AI Eyes” (LLM Annotation)

The most innovative part of this workflow is the integration of GPT-5.1 for multi-modal annotation. Traditionally, automated reports are limited to processing text and numbers. However, by sending the actual mass spectra images along with metadata to the LLM, we effectively give the AI “eyes.”

• Multi-modal Analysis: GPT-5.1 doesn’t just look at intensity values; it evaluates the quality of the spectrum itself—checking for signal-to-noise ratios, peak shape, and baseline stability.
• Contextual Annotation: The AI follows specific biological and analytical rules to write structured annotations for each sample, providing a human-like summary of the data’s reliability and significance.

Step 4: Cost Efficiency via Azure OpenAI Caching

Deploying LLMs at scale can be expensive. To keep costs down, I utilized Azure OpenAI API’s cache discount features. By structuring the prompt endpoints and metadata consistently, we maximize the cache hit ratio. This strategy significantly reduces the cost per sample, making it feasible to run large-scale plate analyses through the LLM without breaking the budget.

Step 5: Democratizing the Tool via Streamlit & Domino

By leveraging Streamlit for the frontend and publishing it as a Domino Model Product, I’ve transformed a complex script into a user-friendly web application. The UI is designed to be intuitive while offering deep flexibility:

• Optional Features: Users have full control over the workflow. Whether it’s generating Plate Heatmaps (which requires uploading a platemap CSV from our LIMS), performing AI Annotations, or using Spectrum Visual Analysis, everything is optional based on the specific needs of the experiment.
• Parallelized Processing: For large datasets, the AI annotation step can be parallelized. Users can specify the number of concurrent processes to significantly speed up the reporting phase.
• Flexible Data Export: Once processing is complete, a clear success notification informs the user. They can then choose to download the entire package or select specific components, such as high-resolution plots, images, the processed Excel report, or the detailed processing logs.

This integration of traditional automation, spatial visualization, and cutting-edge multi-modal AI is a glimpse into the future of data-driven bioprocessing.

Generated by Gemini CLI - Empowering Biopharma Workflows.