CARTiBASE: an interactive knowledge base for CAR sequence retrieval and similarity analysis

CARTiBASE is a freely available, interactive web platform that consolidates and standardizes over 10,000 Chimeric Antigen Receptor (CAR) sequences from diverse sources to enable efficient retrieval, comparative analysis, and exploration of design trends in cellular immunotherapy.

The Gist

Imagine you are a master chef trying to invent the perfect spicy soup. You know that the secret to a great soup lies in the specific combination of ingredients: the broth, the spices, the vegetables, and the garnish. Over the last decade, thousands of chefs around the world have been inventing their own versions of this "spicy soup" (which, in the world of medicine, is called a CAR-T cell therapy). These therapies are designed to teach a patient's immune cells how to hunt down and destroy cancer.

However, there's a huge problem: every chef writes their recipe down differently. Some scribble it on a napkin, some type it in a fancy notebook, some hide it in a patent document, and some just list the ingredients without saying how much of each to use. If you want to compare two recipes to see which one is better, or if you want to steal a great idea from another chef to improve your own, it's a nightmare. You can't find the recipes, they don't match up, and you can't easily see how one is different from another.

Enter CARTiBASE.

Think of CARTiBASE as the world's first giant, organized recipe library specifically for these cancer-fighting cell therapies.

What is it exactly?

The researchers behind this project went out and collected over 10,000 different "recipes" (CAR sequences) from scientific papers, patents, and other sources. They didn't just copy-paste them; they cleaned them up and standardized them.

Imagine taking 10,000 messy, handwritten recipes and rewriting them all into a single, perfect format where every ingredient is clearly labeled. Now, instead of a soup, the "ingredients" are parts of a protein:

• The Antenna (Signal Peptide): How the cell knows to start building the tool.
• The Radar (Antigen-Binding Domain): The part that actually spots the cancer cell.
• The Hinge: A flexible joint that lets the radar move.
• The Anchor (Transmembrane): How the tool sticks to the cell wall.
• The Engine (Intracellular Signaling): The part that tells the cell to attack once it spots the enemy.

What can you do with it?

The paper describes a website (like a Google for cell therapies) where you can do three main things:

1. The "Find My Twin" Search:
   Imagine you have a new recipe you invented, but you want to know if someone else has already made something similar. You can type your recipe (or upload a file) into CARTiBASE. The system acts like a super-fast librarian, scanning all 10,000 recipes and saying, "Hey, this one you made is 95% similar to this one from 2018, but you changed the hinge length!" This helps scientists avoid reinventing the wheel and spot trends.

2. The "Recipe Map":
   When you look at a specific therapy in the database, it doesn't just show you a long string of letters (which looks like gibberish). It breaks it down visually, showing you exactly where the "Radar" ends and the "Anchor" begins. It's like having a diagram of a car engine where you can click on the spark plugs to see exactly what kind they are.

3. The "Trend Report":
   The researchers used this library to look at the big picture. They found some fascinating patterns:

   • The "Standard Anchor": They discovered that most chefs (scientists) use the exact same "Anchor" part for their recipes. It's like how almost every car uses the same type of tire. One specific tire design is used in over 4,000 recipes!
   • The "Radar Families": For certain cancers (like CD19 or BCMA), there are a few "parent" radar designs that everyone copies and tweaks slightly. It's like how almost all modern smartphones look similar because they all evolved from the same original design. But for other cancers, the designs are all over the place, with no single dominant style.

Why does this matter?

Before CARTiBASE, trying to compare these therapies was like trying to compare languages where everyone speaks a different dialect. It was slow, confusing, and prone to errors.

Now, CARTiBASE is like a universal translator. It allows scientists to:

• Reuse good ideas: If a recipe works great, they can find it instantly and use it as a base for a new one.
• Spot mistakes: If everyone is using a specific part that causes side effects, they can see that pattern immediately.
• Invent faster: By seeing what has already been tried, they can focus on the parts of the recipe that haven't been explored yet.

In a Nutshell

CARTiBASE is a free, online tool that organizes the chaotic world of cancer-fighting cell therapies into a neat, searchable library. It turns thousands of messy, scattered recipes into a clear map, helping scientists understand what works, what doesn't, and how to build the next generation of life-saving medicines faster and smarter.

You can visit their library at cartibase.org to explore these "recipes" yourself!

Technical Summary

1. Problem Statement

Chimeric Antigen Receptors (CARs) are modular synthetic constructs critical for cellular immunotherapy. While their clinical success has led to an explosion in new designs, the field faces significant data fragmentation challenges:

• Dispersed Sources: CAR sequences are scattered across heterogeneous sources (publications, patents, supplementary files) rather than a centralized repository.
• Inconsistency: Data is reported in inconsistent formats, including partial sequences, mixed annotations, and non-machine-readable files.
• Analytical Barriers: This fragmentation hinders reproducibility, prevents systematic comparative analysis, and limits the ability to explore how design variations (e.g., in hinge length or signaling domains) impact therapeutic performance.

2. Methodology

The authors developed CARTiBASE, a web-based platform designed to unify, standardize, and analyze CAR sequences.

A. Data Curation and Storage

• Scale: The database curates and standardizes over 10,000 CAR sequences (specifically 10,979 in the first deployment).
• Data Model: Internally, data is stored in a curated CSV format where each row represents a single construct. Columns capture:
  • Full amino acid sequences.
  • Segmented domain sequences and boundaries.
  • Target antigen information.
  • Source metadata (authors, institutions, patents).
• Updates: The database is updated monthly to incorporate new literature and patents.

B. Domain Segmentation and Annotation Pipeline

CARTiBASE employs a stepwise, hybrid annotation strategy to segment CARs into four components: Signal Peptide, Extracellular Region, Transmembrane (TM) Domain, and Intracellular Region.

1. Topology Prediction: DeepTMHMM (a deep learning-based predictor) identifies signal peptides, membrane-spanning helices, and cytosolic regions with per-residue probability labels.
2. Domain Identification: HMMER is used against curated profile Hidden Markov Models (HMMs) to identify specific domains:
   • Extracellular: scFv-related VH/VL motifs.
   • Intracellular: Signaling domains (e.g., CD28, 4-1BB, CD3ζ).
   • Only high-confidence matches (based on bit scores and E-values) are retained.
3. Antigen Extraction: A rule-based workflow using keyword detection (e.g., "anti-...", "CD\d+") and regular expressions extracts target antigens from titles and metadata, normalizing them to a pre-defined vocabulary.

C. Similarity Search Engine

• Algorithm: The platform utilizes a DIAMOND-compatible workflow for fast protein similarity search.
• Functionality: Users can query with full-length CARs or specific domain sequences (minimum 15 amino acids).
• Output: The system returns ranked hits with identity, coverage, E-values, and alignment coordinates, visualized in sortable tables with links to detailed domain maps.

D. Technical Architecture

• Backend: Python/Flask API.
• Frontend: Vue.js.
• Access: Freely available at https://www.cartibase.org without mandatory registration.

3. Key Contributions

• Standardized Resource: Creation of the first large-scale, harmonized database of >10,000 CAR sequences with consistent domain-level annotation.
• Interactive Analysis Tools: Development of a web interface allowing for global filtering, domain boundary inspection, and similarity-based retrieval.
• Automated Annotation Pipeline: A robust workflow combining DeepTMHMM and HMMER to automatically segment and annotate complex CAR architectures from raw sequences.
• Open Data Access: Provision of downloadable FASTA files for both full-length constructs and segmented domains, facilitating downstream computational and experimental reuse.

4. Results and Findings

The authors leveraged CARTiBASE to analyze the diversity of the public CAR landscape, revealing distinct design trends:

A. Target Antigen Distribution

• Dominant Targets: BCMA and CD19 are the most frequently targeted antigens, followed by CD22, CD123, and CD20.
• Data Source: These frequencies are derived from metadata extraction, not sequence inference.

B. Transmembrane (TM) Domain Diversity

• Conservation: The TM landscape is dominated by two highly conserved super-families rather than a uniform distribution.
  • Family 1: The motif IYIWAPLAGTCGVLLLSLVITLYCKRGR appears 4,016 times, with minor C-terminal variants forming a tight cluster.
  • Family 2: A second major family centered on FWVLVVVGGVLACYSLLVTVAFIIFWVRS(K/R) (965 occurrences).
• Implication: A small number of canonical TM motifs serve as the backbone for the majority of CARs, while alternative architectures remain underexplored.

C. Antigen-Binding Domain (scFv) Clustering

• BCMA: Displays multiple distinct lineages. Some are tightly conserved clusters (likely from single parental antibodies), while others form a large, heterogeneous "super-family" with moderate similarity, suggesting a common scaffold undergoing extensive affinity maturation.
• CD19: Characterized by several dense, high-similarity clusters (e.g., FMC63-derived sequences) indicating strong conservation of specific lineages, alongside a diverse background of engineered variants.
• CD123: Dominated by two large, highly conserved families (likely from few parental antibodies) and a broad, heterogeneous pool of divergent variants.

5. Significance

• Reproducibility: By providing standardized, machine-readable sequences, CARTiBASE eliminates the ambiguity caused by inconsistent reporting in literature, enabling direct comparison of constructs.
• Design Optimization: The similarity search and clustering analysis allow researchers to identify "nearest neighbors" and common design patterns (e.g., specific hinge lengths or co-stimulatory combinations), accelerating the engineering of next-generation CARs.
• Resource Efficiency: The platform saves researchers time by aggregating dispersed data and offering pre-computed domain annotations, removing the need for manual sequence parsing.
• Future Directions: The database highlights that while some domains (like TM regions) are highly conserved, others (like scFvs) show vast diversity, guiding future exploration into underutilized architectural variants.

In summary, CARTiBASE transforms fragmented CAR data into a structured, interactive knowledge base, serving as a critical infrastructure for the advancement of CAR-T cell therapy research and development.