Team Smith Roster
| Role | Participant | Affiliation |
|---|---|---|
| Team Lead | Todd Smith, PhD | Digital World Biology, LLC |
| Tech Lead | Herminio Vazquez | Copado Inc. |
| Flex | Zainab Adenaike | NIH/NLM/NCBI |
| Flex | Jake Lance | student, University of Toronto |
| Flex | Mohsen Sharifi Renani | Spotify AB |
| Writer | Stephen Panossian | Unaffiliated |
The project focused on developing resources and documentation for teacing data science and machine learning / artificial intelligence (ML/AI) cocepts related to antibody engineering. Immune profiling (immunoprofiling) datasets were used as a source of antibody sequneces for both data science and ML. The team develope Jupyter notebooks to undertake comparative analyses of iReceptor datasets, and then incorporate the AbLang2 antibody-specific language model to characterize data from CoV-AbDab. A dictionary and glossary of terms defining essential computer and biology terms related to the computations processed within the Jupyter notebook were also developed.
- CoV-AbDab database in csv format. CoV-AbDab is a public database to document all published/patented antibodies and nanobodies able to bind to coronaviruses, including SARS-CoV2, SARS-CoV1, and MERS-CoV. The codathon used the Feb 8, 2024 release containing 12,916 entries. Entries are highly annotated and indicate neutralizing ability, kind of receptor (antibody, nanobodie), where data are pair (heavy and light chaing, just heavy), epitope bound, if a stucture exists, virus reactivitiy among others.
- iReceptor (free account required) lymphoma dataset uptained with the following filters: Study ID: PRJEB1289; Study type Case Control (Ontology ID): NCIT:C15197; Filter by Sample > PCR target: IGH or IGK or IGL
- Immune Profiling: See notebooks for details: Key python libraries include Pandas for structuring and manipulating data, json for reading metadata, Matplot lib for graphing and Seaborn for exploring correlations between data in columns.
- Machine learning: AbLang2
The following diagrams represent the high-level methods employed in Data Science and Bioinformatics
The common source for antibody seqeunce data comes from immune profiling experiements and assays.
flowchart TD
A[Collect Samples] --> B[Isolate DNA / RNA->cDNA] --> C[PCR] -- V-gene, C-gene primers --> D[Sequence DNA] -- NGS - massively parallel --> E[IgBLAST] -- Vh Dh Jh, Vl Jl, Vk Jk references --> F[Immune Profile Dataset];
F -- repeat --> A
F --> G[Explore data, analyze];
F --> H[Machine learning];
High level data science workflow.
flowchart LR
A[Collect] --> B[Profile]
B -->C{complete?}
C-->|Yes|D[Exploration]
C-->|No|A
D --> E[Charts]
D --> F[Impute]
E --> G[Aggregate]
F --> G
G --> H[Model]
H --> I[Feature Engineering]
I --> J[Train/Test]
I --> K[Tune]
K --> J
J --> L[Predict]
L --> M[Operationalize]
M --> N[Monitor]
See mermaid to learn about making the figure. Mermaid.org, and flow charts provide complete documentation.
The team used software tools including Amazon Web Service (AWS) cloud computing accounts, Jupyter notebooks, and datasets from both iReceptor and SAbDab (The Structural Antibody Database) from the Oxford Protein Information Group (OPIG). The general workflow is: 1) create an AWS instance, 2) step through the enclosed Jupyter notebook, and 3) analyze the antibody results. Minor experimentation was done with Docker containers.
Prior work illustrates this approach:
Example: Covid not Covid
- https://github.com/AntibodyEngineers/covid-not-covid
- notebook: https://github.com/AntibodyEngineers/covid-not-covid/blob/main/ab_predict_neutralising.ipynb
- datafile: covabdab_all.csv
Example: Immune Profiling
- https://github.com/AntibodyEngineers/immune-profiling
- data: see Methods above.
2024 ML/AI Codeathon Log
| Date | Issues |
|---|---|
| 26-Feb-2024 | Several issues utilizing Docker |
| 27-Feb-2024 | Team accessed AWS account and Jupyter notebook; runtime challenges |
| 28-Feb-2024 | None reported |
| 29-Feb-2024 | None reported |
| 01-Mar-2024 | Final Presetation |
Many jupyter notebooks and notebook fragements were created. All are in the notebooks folder. The most instructive notebooks are:
- ab_predict_neutralising.ipynb the notebook from Covid-not-Covid that was used as a starting point for this work.
- ab_predict_neutralising_final.ipynb includes working code plus descriptions of the machine learning process and rational for certain choices. In addition code is included for exploring the dataset that was used for training and testing the model.
- Herminio.ipynb provides a few simple examples of using markdown in jupyter notebooks.
- ireceptor.ipynb was a notebook from Immune Profilling and was used as a starting point for this work.
- ireceptor-herminio.ipynb demonstrates several pyton libraries and code that are used to explore a large dataset and introduces parquet files as a way to efficiently work with large datasets.
- ireceptor-mohsen.ipynb add PCA plotting.
Project materials will create a resource with instruction and hands-on examples that can demystify ML/AI for many scientists and students who need greater awareness of the data, steps, and practicalities. The focus on antibodies supports work in basic research and biotechnology. Digital World Biology's Antibody Engineering Hackathons are creating materials for course-base undergraduate research in community colleges (https://antibody-engineers.org/).
The resulting work will be used in Digital World Biology's National Science Foundation funded summer hackathon (August 2024) on antibody engineering. In the project we will consider ML applications for predicting antibody antigen recognition, genetic contributions to antibody expression, and de novo antibody design. Work will identify one or two examples that include specific datasets, workflows, an appropriate ML method, and tests. The examples will then be used to create instructions and explanations that can be used in classroom settings, starting points for undergraduate research, and scientists wishing they had ways to better understand ML.
This software was created as part of an NCBI codeathon, a hackathon-style event focused on rapid innovation. While we encourage you to explore and adapt this code, please be aware that NCBI does not provide ongoing support for it.
For general questions about NCBI software and tools, please visit: NCBI Contact Page