Carafe2 enables high quality in silico spectral library generation for timsTOF data-independent acquisition proteomics

The paper introduces Carafe2, a deep learning-based tool that generates high-quality, experiment-specific in silico spectral libraries directly from native timsTOF DIA raw data by fine-tuning retention time, fragment ion intensity, and ion mobility prediction models, thereby outperforming existing DDA-trained models and enabling superior peptide detection across diverse proteomic applications.

The Gist

Imagine you are trying to find a specific needle in a massive, chaotic haystack. But this isn't just any haystack; it's a haystack made of millions of tiny, glowing threads, and you need to find specific patterns of light to identify them. This is essentially what scientists do when they study proteomics—the study of all the proteins in a biological sample like blood or cells.

For a long time, scientists used a method called DDA (Data-Dependent Acquisition). Think of this like a security guard at a club who only lets in the loudest, flashiest people (the most abundant proteins) and ignores the quiet ones in the back. It's fast, but you miss a lot of the crowd, and the guard's choices are a bit random, so you might not get the same people in every visit.

Then came DIA (Data-independent Acquisition). This is like a security guard who scans everyone in the club, regardless of how loud they are. It's much more thorough and reproducible. However, because everyone is being scanned at once, the data becomes a giant, confusing soup of overlapping signals. To make sense of this soup, scientists need a Spectral Library.

The Problem: The "Recipe Book" Mismatch

Think of a Spectral Library as a recipe book or a Wanted Poster. It tells the computer exactly what a specific protein "looks like" (its mass, how long it takes to travel through a machine, and how it breaks apart).

• The Old Way: Scientists used recipe books written for the old "security guard" (DDA). But the new machine (DIA) works differently. Using an old recipe book for a new machine is like trying to bake a cake using a recipe for bread; the ingredients are right, but the instructions don't quite fit, leading to a messy result.
• The New Machine: The timsTOF is a super-advanced machine that adds a third dimension to the search. It doesn't just look at weight and time; it also measures how "bouncy" or "sticky" a molecule is as it moves through air (Ion Mobility). It's like adding a "texture" check to your security scan.
• The Gap: Existing recipe books didn't know how to describe this new "texture" dimension, or they were written for the old machine, so they didn't match the new data perfectly.

The Solution: Carafe2 (The "Smart Tutor")

Enter Carafe2. The authors created a software tool that acts like a smart tutor or a personalized chef.

Instead of using an old, generic recipe book, Carafe2 looks at the specific data from your experiment and learns how to write a brand new, perfect recipe book just for you.

Here is how it works, using a simple analogy:

1. The Training Phase: Imagine you have a student (the AI) who is good at math but hasn't taken your specific class yet. Carafe2 takes a small sample of your data (a few "practice tests") and teaches the student the specific quirks of your machine. It says, "Hey, in this lab, with this machine, the proteins move a little faster and break apart slightly differently than the textbooks say."
2. The Fine-Tuning: The AI adjusts its internal "recipe book" to match your specific experiment. It learns the exact timing, the exact "bounciness" (ion mobility), and the exact intensity of the signals.
3. The Result: Now, the AI has a customized library that fits your data perfectly. It knows exactly what to look for, down to the smallest detail.

Why is this a Big Deal?

The paper shows that when scientists use this new, custom-tailored library (Carafe2) instead of the old, generic ones:

• They find more needles: They detect significantly more proteins (up to 12-13% more in some cases). It's like the security guard suddenly noticing people who were previously hiding in the shadows.
• They are more accurate: The measurements of how much of each protein is present are more precise.
• They work better with complex samples: Even in messy samples like human blood plasma (which is full of thousands of different proteins), Carafe2 outperformed the old methods.

The Extra Tools: TimsQuery and Timsviewer

To make this easy for everyone, the team also built two helper tools:

• TimsQuery: A tool that lets the software read the raw data files directly, skipping the need to convert them into messy intermediate formats. It's like having a universal translator that speaks the machine's native language instantly.
• Timsviewer: A visual tool that lets scientists "see" the data. It's like a magnifying glass that lets you zoom in on a specific protein and see if the "Wanted Poster" matches the person you found.

The Bottom Line

Carafe2 is a game-changer because it stops scientists from trying to force old tools to work on new, high-tech machines. Instead, it uses Artificial Intelligence to learn the specific "personality" of your experiment and builds a custom guidebook. This leads to finding more proteins, getting more accurate results, and ultimately helping us understand diseases and biology better.

In short: Carafe2 turns a generic map into a GPS that knows exactly where you are, ensuring you never miss a single protein in your search.

Technical Summary

1. Problem Statement

Data-Independent Acquisition (DIA) proteomics, particularly when combined with Trapped Ion Mobility Spectrometry (TIMS) on the timsTOF platform (diaPASEF), offers high reproducibility and deep proteome coverage. However, analyzing DIA data typically relies on spectral libraries containing expected peptide properties: retention time (RT), fragment ion intensity, and ion mobility (IM).

Existing challenges include:

• Mismatch in Training Data: Most in silico library generation tools rely on models trained on Data-Dependent Acquisition (DDA) data. This creates a mismatch when applied to DIA data, failing to capture experiment-specific biases.
• Lack of Ion Mobility Support: While deep learning has improved RT and intensity predictions, many existing tools do not support the prediction of ion mobility (collisional cross-section), a critical dimension for separating co-eluting interferences in timsTOF data.
• Preprocessing Overhead: Current workflows often require converting native timsTOF raw data (.d directories) into intermediate formats (e.g., mzML), which is computationally expensive and loses the hierarchical structure of the data.
• Generalizability: Pretrained models often fail to adapt to specific instrument configurations, LC gradients, or sample types, leading to suboptimal peptide detection.

2. Methodology

The authors introduce Carafe2, an extension of the original Carafe software, designed specifically to generate experiment-specific in silico spectral libraries for timsTOF DIA data.

• Direct Native Data Access: Carafe2 operates directly on native Bruker .d directories using a new standalone Rust-based tool called TimsQuery. This eliminates the need for conversion to mzML, allowing efficient extraction of chromatographic and spectral data across retention time and ion mobility dimensions.
• Deep Learning Fine-Tuning:
  • Carafe2 utilizes deep learning models (based on architectures from AlphaPeptDeep and the original Carafe) for predicting RT, fragment ion intensity, and IM.
  • Instead of relying solely on pretrained DDA models, Carafe2 fine-tunes these models using a specific DIA run from the target experiment (e.g., human data) to generate a library for a different condition (e.g., yeast data) or to refine the library for the same condition.
  • The IM model predicts inverse reduced ion mobility ($1/K_0$), which can be converted to Collisional Cross Section (CCS).
• Workflow and Integration:
  • A Graphical User Interface (GUI) was developed to support three workflows: library generation from existing DIA-NN results, end-to-end library generation with DIA-NN search, and full DIA analysis.
  • Carafe2 is integrated into Skyline for targeted proteomics.
  • A new visualization tool, Timsviewer (also written in Rust), was developed to inspect timsTOF raw data alongside Carafe2-generated libraries, allowing users to verify peptide detections without third-party SDKs.

3. Key Contributions

1. Carafe2 Software: A tool that generates high-quality, experiment-specific in silico libraries by fine-tuning RT, intensity, and IM models directly on timsTOF DIA data.
2. TimsQuery: A lightweight, cross-platform Rust tool for direct, efficient access to native timsTOF raw data, bypassing the need for format conversion.
3. Timsviewer: A standalone visualization tool for inspecting timsTOF DIA data and validating spectral library predictions.
4. Integration: Seamless integration with DIA-NN and Skyline, making the tool accessible to the broader proteomics community.
5. Open Source: All tools (Carafe2, TimsQuery, Timsviewer) are open source (Apache 2.0 license).

4. Results

The authors evaluated Carafe2 across diverse datasets (global proteome, phosphoproteome, plasma, and cancer samples) using timsTOF Ultra 2, Pro, HT, and SCP instruments.

• Prediction Accuracy:
  • Fragment Intensity: Fine-tuning significantly improved Spearman correlations between observed and predicted intensities. For phosphoproteomics, 79.92% of peptides showed improved correlation after fine-tuning compared to pretrained DDA models.
  • Retention Time: Fine-tuned models showed higher $R^2$ values and better linear agreement, particularly at the extremes of the chromatographic run where pretrained models exhibited nonlinearity.
  • Ion Mobility: Fine-tuning reduced the median absolute prediction error and corrected for experiment-specific shifts in IM values that pretrained models missed.
• Peptide Detection (DIA-NN):
  • Using fully fine-tuned libraries (RT + Intensity + IM) increased peptide precursor detection by 12.70% compared to the baseline DDA-trained library on global proteome datasets.
  • In plasma proteome analysis, Carafe2 detected 13.1% more precursors than experimental libraries derived from 63 DDA runs and 5.8% more than DIA-NN's built-in in silico library.
  • In a triple-negative breast cancer dataset, Carafe2 detected 45.5% more precursors than the DDA experimental library.
• Quantification:
  • Carafe2 maintained comparable quantification precision (Coefficient of Variation) to other methods.
  • It demonstrated superior accuracy in recovering expected mixture ratios in mixed-species datasets.
  • In a lung cancer plasma study, Carafe2 identified 16.3% more significantly regulated precursors than the DIA-NN built-in library, demonstrating increased statistical power for differential expression analysis.
• FDR Control: Evaluation using an entrapment strategy confirmed that Carafe2 fine-tuned libraries do not inflate False Discovery Rates; estimated FDPs remained under 1% at the 1% FDR threshold.

5. Significance

• Overcoming the DDA-DIA Mismatch: Carafe2 solves the critical issue of applying DDA-trained models to DIA data by enabling direct fine-tuning on the specific DIA experiment, capturing unique instrument and acquisition biases.
• Leveraging Ion Mobility: By accurately predicting ion mobility, Carafe2 fully exploits the additional separation dimension provided by timsTOF, leading to better discrimination of co-eluting peptides and higher detection sensitivity.
• Efficiency and Accessibility: The ability to process native raw data without conversion, combined with a user-friendly GUI and integration into Skyline, lowers the barrier to entry for high-quality DIA analysis.
• Clinical and Biological Impact: The demonstrated improvements in detecting low-abundance peptides and differentially expressed proteins in complex clinical samples (plasma, cancer tissues) suggest that Carafe2 can significantly enhance the sensitivity and reliability of biomarker discovery and systems biology studies.

In conclusion, Carafe2 represents a significant advancement in DIA proteomics by providing a robust, open-source framework for generating high-fidelity, experiment-specific spectral libraries that outperform both empirical DDA libraries and existing in silico alternatives.