The ImmuneWatch Logo

Our algorithms

Making the T-cell receptor repertoire clinically actionable by applying machine learning in immunology.

TCR repertoire analyses lack annotations

Current analysis techniques give insights in the dynamics and size of the T-cell receptor (TCR) repertoire. They can reveal what kind of TCRs are clonally expanded and what the total diversity of the repertoire is. However, these analyses lack the most crucial step to make the data useful in any clinical setting: annotation of the pathogens (epitopes/antigens) that each T-cell receptor recognises.

Schematic overview of T-cell receptor repertoire annotation using ImmuneWatch's techology. It brings machine learning in immunology
Schematic overview of T-cell receptor repertoires. In an unannotated repertoir it is not possible to know the epitopes, antigens or pathogen the T-cell receptor will recognise (top panel). However, by annotating the T-cell receptor repertoire, the receptors will reveal their target, leading to allowing clinically actionable insights.

Make your data clinically actionable

ImmuneWatch offers two ways to add meaningful annotations to your datasets.

Database annotation

We have built a comprehensive database of T-cell receptors and the epitopes they recognise. The database contains peer-reviewed published data as well as in-house-generated datasets.

Our data is curated by a trained biodata-curation team to ensure it is well annotated. In addition, we have designed several metrics that can be used to evaluate the quality of TCR-epitope pairs.

Match your dataset with our database to generate new insights!

Artificial intelligence image

AI-based annotation

Due to the large diversity in TCRs and epitopes, it is impossible to build a database for each theoretically possible TCR-epitope combination. Therefore, in order to annotate unseen TCR-epitope pairs, we need to turn to clever, trustable algorithms that can learn the underlying rules of TCR-epitope binding.

Thankfully artificial intelligence (AI) has proven its usefulness to tackle these kind of bio-molecular binding challenges. As a University of Antwerp spin-off company we can make use of the expertise in this field gathered by our academics. Our models, trained on our comprehensive well-curated database, enable machine learning in immunology. Therefore, ImmuneWatch can offer unprecedented insights into immune responses.

Algorithm portfolio

ImmuneWatch has multiple algorithms in the pipeline that bring machine learning into immunology. Our algorithms can be divided into three categories.

All of them are under active development and are updated through a continuous improvement and continuous development (CI/CD) strategy.

Known-epitope models

This category of algorithms contains AI-models that can predict TCRs that bind to epitopes for which we have training data available in our database, i.e. previously seen epitopes. These are our best performing algorithms as the classification problem is less complex than the other category of algorithms (epitope-agnostic models). The main disadvantage is that if the epitope is not yet covered in our database, predictions are not possible and this data first needs to be generated with additional experiments.

An example of such a model in our portfolio is TCRex (published in Frontiers in Immunology). Further internal development have improved the performance of alternative versions of this algorithm since its publication (unpublished proprietary data).

The logo of the TCRex tool
TCRex, an example of a "known-epitope model" in ImmuneWatch's portfolio.
Graphical abstract of the SARS-CoV2 omicron white paper
Graphical abstract of the white paper which uses the "Epitope-agnostic model" Parzival to evaluate the effect of mutations within the SARS-CoV-2 omicron variant on T-cell immunity.

Epitope-agnostic models

In contrast to the “known-epitope models” the epitope-agnostic models do not require an epitope to be present in the database. The user can feed in any epitope of interest and explore whether there are TCRs in their dataset that would are able to recognise this epitope. In general, development of models within this category of algorithms is more challenging as large training datasets are required to be able to deduce the underlying rules of TCR-epitope binding.

Being at the forefront of this research field, a large chunk of ImmuneWatch’s research efforts is allocated to development of these kind of algorithms. Continuous improvements and an ever growing training database will ensure that the performance of these algorithms will be state-of-the-art.

Our current portfolio of algorithms in this category entails ImRex (published in Briefings in Bioinformatics). In addition, Parzival, which we used to predict the effect of the large number of mutations in the SARS-CoV-2 Omicron variant on T-cell immunity, belongs to this category of algorithms.

Clustering of TCRs

Another useful category of algorithms in T-cell receptor repertoire analysis are clustering algorithms. They allow the user to identify TCRs that are similar to each other, based on e.g. sequence similarity. Under the hypothesis “similar TCRs will bind similar epitopes” this category of algorithms can be used to infer the annotation of a TCR based on the cluster it belongs to.

Our flagship algorithm in this category is ClusTCR (published in Bioinformatics). ClusTCR offers a drastic improvement in clustering speed, which allows the clustering of millions of TCR sequences in just a few minutes.

Contact us now for more information, collaborations and partnerships


Connect with us!

Business [at]

Privacy and cookie policy

Copyright © 2021 ImmuneWatch BV. ImmuneWatch is a registered trademark.

ImmuneWatch BV | BE0774.472.942 | Eénmeilaan 52, 3010, Leuven, Belgium