Computational models capable of forecasting an individual’s immune response to the influenza vaccine are needed to provide mechanistic insights and pave the way for personalized vaccination strategies. In this challenge we provide training data collected pre- and post-vaccination from historical cohorts, as well as a testing (a.k.a. challenge) dataset comprising pre-vaccination data from the 2025-26 season. This challenge dataset comes from an annual influenza vaccine study run at the La Jolla Institute for Immunology. These data are multi-modal, consisting of cytokine measures, transcriptomics, genetic information, and cell frequencies and phenotypes. Participants are asked to train models that predict post-vaccination responses and apply them to the challenge dataset.

We invite you to take part by heading over to our competition site on Kaggle :
https://www.kaggle.com/competitions/cmi-flu-first-prediction-challenge

The overall goal of the CMI-Flu prediction challenge is two-fold. We want to:

1. Establish a community platform to test and compare computational models of immunity in vaccination.
2. Better understand vaccine-induced immunity to influenza. By establishing and testing computational models that attempt to predict the cascade of events that follow vaccination, we will improve our understanding of the mechanisms underlying these events, with the ultimate goal of identifying what variables induce a strong and durable recall response.

The CMI-Flu prediction challenge is an exciting opportunity to explore the complex world of systems vaccinology by utilizing longitudinal immune response data obtained through cutting-edge multi-omics experiments. With this challenge, the CMI-Flu prediction contest hopes to foster a collaborative research community, addressing challenges and advancing scientific knowledge more rapidly than any individual or research group could achieve alone.

Your task during this challenge is to predict outcomes, showcasing your intuition and analytical skills. To participate in the Prediction Challenge, utilize the provided training datasets to build your computational models and predict the post-vaccination outcomes of newly tested individuals (challenge dataset).

We have selected different tasks to assess different aspects of immunity. These tasks will be released in 2 parts where participants can work alone or in teams to predict unseen post-vaccination data. In a third part of the competition, contestants will work together to build retrospective models after the post-vaccination data are revealed.

This competition is about exploring system-wide responses. We provide a broad set of readouts, and we're excited to see how far you can push by integrating them to discover signals that emerge only when the pieces are viewed together. We encourage you to explore as many assays as possible for any given task.

Part 1 (current competition)

You will build models to predict post-vaccination values based on baseline data and demographics. Data will be released on 23 July 2026, and this part of the challenge will close on 6 October 2026. For this part you will predict systems responses (both innate and adaptive) and antibody responses.

Part 2 (future competition)

You will build models to predict post-vaccination values based on baseline data and demographics as well as innate post-vaccination data. This challenge will launch on 20 October 2026 and close on 19 January 2027. For this part you will predict antibody responses.

Part 3 (collaborative model building)

After the post-vaccination data have been revealed, contestants will have an opportunity to join forces with other contestants to collectively build the best-in-class model for this year. All participants will be invited to be co-authors on a collaborative manuscript.

Specific details of each task, including descriptions of the feature to predict, the assay involved, and any transforms needed, may be found here.

Systems Responses

As sketched out in Figure 2, immediately after influenza vaccination, an innate immune response is activated that includes monocytes, dendritic cells (DCs), and natural killer (NK) cells intended to inhibit pathogens and stimulate adaptive immune responses. Antigen-presenting cells (APCs) take up influenza antigens at the vaccination site and migrate to the local draining lymph nodes, where they present antigens to resident T and B cells and those recruited via innate activation signals. In the following prediction tasks, we explore different aspects of this cascade. In this part, we are interested in identifying the baseline features that determine responsiveness.

Tasks:

1.1 : CXCL10 (IP10) response at Day 1.

1.2 : Monocyte frequency at Day 1.

1.3 : Plasmablast frequency at Day 7.

1.4 : antigen-specific T cell frequency at Day 7.

Antibody responses

Influenza vaccines have historically been evaluated by measuring antibody titers through the hemagglutination inhibition (HAI) assay, with higher titers representing a stronger response. This remains the only known correlate of protection, where an HAI titer≥40 correlates with ≈50% protection. While prior analyses often investigated this antibody response in isolation, the CMI-Flu challenge will examine whether a systems immunology approach can better predict the magnitude, breadth, and durability of antibody titers.

Tasks:

2.1 : antibody magnitude at Day 28.

2.2 : antibody breadth at Day 28.

2.3 : antibody durability at Day 365.

Figure 1 : HAI titers

Antibody responses based on the post-vaccination innate response (Part 2)

In the second part of the competition, we will explore the contributions of the innate immune system after vaccination to antibody response.

Tasks:

3.1 : antibody magnitude at Day 28.

3.2 : antibody breadth at Day 28.

3.3 : antibody durability at Day 365.

Data may be accessed through the Kaggle competition site. Contestants are provided with:

• Training data. Public datasets that have been collected and standardized, containing paired pre- and post-vaccination measurements. Contestants use these to build and validate their models.
• Challenge data (2025LJI). Generated at the La Jolla Institute for Immunology. Pre-vaccination measurements are released so participants can generate predictions; the corresponding post-vaccination responses are withheld and constitute the prediction target. They will be released once the competition closes.
• Reference files. A set of supporting files accompanies the training and challenge data. Each is described in more detail on the relevant data preparation pages.

We encourage you to follow the links to find out more about how data was:

• collected,
• generated, and
• prepared and standardized for the competition. Contestants will find a lot of useful information about the data available for training on these pages.

Challenge data will also be made available through our CMI-Flu central repository.

Submissions

For each task, submit your predictions as values predicted by your model, for each donor in the challenge dataset. An example of the submission format can be found in the sample_submissions_part1.csv file in the data folder. We suggest you download the sample submission file and fill this in for your submissions.

The submission file must be a .csv file with the first column ('Participant_ID') listing all 40 donors, followed by 1 column for each task (named "Task_x.x").

Evaluation

1. Kaggle's public leaderboard: Models will be evaluated through-out the competition using Kaggle's public leaderboard, which relies on 30% percent of the withheld data. Contestants can submit multiple models and track their progress up until the close of the competition, at which point, the top 4 models from each participant will be selected to go through to the final evaluation. Contestants may select the models they would like to enter, alternatively the top models will be automatically selected by Kaggle.
2. Final evaluation: The final models will be evaluated on the remaining 70% percent of the withheld data.

Evaluation metric

For Kaggle's public leaderboard, each task will be evaluated using a Spearman correlation, and a mean will be taken across all tasks. If a user does not attempt a task, the task is assigned a correlation score of 0 for that task.

For the final evaluation, models will be scored using both a spearman correlation and a root mean squared error (RMSE) for each task individually. Here, only the model with the best score from each contestant/team will be evaluated. Winners will be asked to submit details of their models, including methods used, data sources etc.

We only evaluate one model from each contestant. This ensures that a model that does well cannot dominate the rankings if submitted multiple times with minor tweaks. However we value the submission of widely different models. So while contestants can submit multiple models from the same account in Kaggle, we also encourage contestants to submit different models from different accounts so that all of these models may be evaluated.

We hold regular informational sessions through-out the competition. Timings for these sessions are shown on the challenge timeline. Below are recordings and slides from previous sessions.

Don't forget to check out the Kaggle discussion forum where you can view previous questions and post your own.

First informational session (31st July 2026)

• slides
• recording

Second informational session (28th August 2026)

• slides
• recording

Figure 2: Our prediction challenges

You can find out more about our previous CMI-PB challenges here.