Request access to the SEPIQ 2026 training data

Access is provided for approved SEPIQ Challenge participants only. Please complete the form below so the organizing team can review your request.

By requesting access, you agree to use this dataset only for SEPIQ Challenge participation and approved research purposes. Redistribution, rehosting, selling, or sharing the data without written permission from the organizers is not allowed.

Log in or Sign Up to review the conditions and access this dataset content.

SEPIQ 2026 Training Data

Sequence-Based Epitope Prediction Intelligent Query

Gated training data repository for the SEPIQ 2026 Challenge.


Overview

SEPIQ 2026 is a community challenge for benchmarking computational methods that predict antibody–antigen interaction sites from VHH nanobody sequence data.

This repository will provide access to the training data approved for release to accepted SEPIQ Challenge participants. Access is gated so that the organizing team can review requests before data access is granted.

Metadata description

This dataset consists of 1,843,875 unique VHH sequences, together with IE-scores (Interaction Effect), antigen-antibody interaction status labels, cluster annotations, and amino acid sequences.

The IE-score indicates the degree of enrichment or dilution of each VHH clone before and after target-specific panning. The method for calculating the IE-score follows the previously published AVIDa-hIL6 dataset methodology [1]. In brief, a chi-square test is performed on the change in abundance ratio before and after panning for each VHH clone. The natural logarithm of the p-value is then calculated. Extremely low p-values that result in negative infinity during computation are substituted with -500 for convenience. Values for diluted clones are converted to positive values.

A lower IE-score indicates a higher likelihood of binding to the target, whereas a higher IE-score indicates a lower likelihood of binding.

The following algorithm was used to assign antigen-antibody interaction status labels:

Condition
1 If the target-IE score is equal to or greater than the positive threshold (1.3), label it as Not (non-binder).
2 If the background IE-score is equal to or less than the adjusted negative threshold (-3), label it as noise (noise).
3 If the background IE-score is negative and the difference between the background IE-score and the target IE-score is greater than 2, label it as Yes (binder).
4 If the background IE-score is negative but does not meet condition 3, label it as non-sig. (non-significant).
5 If the background IE-score is positive and the target IE-score is at or below the negative threshold (-1.3), label it as Yes (binder).
6 For all other cases (where the background IE-score is positive and the target IE-score is not significant), label it as non-sig. (non-significant).

Main columns

Column Description
#ONU ID VHH clone number. Clone numbers are assigned in descending order based on total read counts.
Total Total number of NGS reads across all libraries.
Sequence Amino acid sequence of each VHH clone.
Label Antigen-antibody interaction status label assigned using an in-house algorithm. Possible categories include Yes(binder), Not(non-binder), noise, and non-sig.(non-significant).
Binary Label Binary label where 1 indicates a binder and 0 indicates a non-binder. Other labels are treated as pending and are not assigned a value.
Cluster Cluster name assigned to each VHH clone. Unique VHH sequences are clustered based on 93% identity.
#Cluster Numerical part of the cluster name.
Cluster size Number of clones in the cluster.

Sublibrary columns

Columns with names containing the prefix Marin represent data obtained from sublibraries under different experimental conditions. The prefix corresponds to the name of the alpaca from which the data was obtained.

The suffix indicates the type of Mother Library. In this dataset, the suffix consists of B and C, and the order of the letters indicates repeated immunostimulation.

The second-to-last segment of the hyphenated section indicates the experiment number. The experiment was conducted in two main phases, 1st and 2nd. Since biological experiments often include background variation, dividing the experiment into phases helps reduce noise.

The third-to-last segment of the hyphenated section indicates the experimental temperature. Experiments were performed at either 4°C or 37°C to provide information on temperature-related effects in antigen-antibody reactions.

Target and control information

Column group Description
NC1 Sublibrary generated by panning using only Protein G beads. Since this library does not contain HER2 protein, it is considered a HER2-free negative control.
NC2 Sublibraries generated by binding anti-HA antibodies to Protein G beads, incubating them with HER2 proteins, and then panning. Ideally, HER2 protein should not bind to the beads, making this a negative control. However, nonspecific HER2 adsorption cannot be fully excluded.
NC3_C1, NC3_C2 Sublibraries generated by amine coupling bovine serum albumin (BSA) to beads, followed by panning. These are HER2-free negative controls with biological replicates.
Comparator_C1, Comparator_C2 Sublibraries generated by amine coupling alternative comparator proteins to beads, followed by panning. These are HER2-free negative controls with biological replicates.
Her2, Her2_C1, Her2_C2 Sublibraries generated by amine coupling HER2 proteins to beads, followed by panning. These libraries may include signals from VHH clones that bind nonspecifically to Protein G or to proteins other than HER2.
Her2_Trastuzumab_C1, Her2_Trastuzumab_C2, Her2_Trastuzumab_C3 Sublibraries generated by binding Trastuzumab to Protein G beads, incubating them with HER2 proteins, and then panning. These experiments include biological replicates.
Her2_Pertuzumab_C3 Sublibrary generated by binding Pertuzumab to Protein G beads, incubating them with HER2 proteins, and then panning. This experiment was conducted under the same conditions as the Her2_Trastuzumab_C3 library.
Her2_RabMAb_C3 Sublibrary generated by binding anti-HER2 rabbit monoclonal antibodies to Protein G beads, incubating them with HER2 proteins, and then panning. This experiment was conducted under the same conditions as the Her2_Trastuzumab_C3 library.

In sublibraries containing Trastuzumab, Pertuzumab, and rabbit monoclonal antibodies, signals from VHH binders competing for these epitopes may be reduced. This information may therefore be useful for epitope prediction. However, competition data are available only for Mother Library B, so competition cannot be determined for VHHs that show signals only in Mother Library C.

Aggregated IE-scores

Column Description
Background IE-score Statistically aggregated signal from negative control columns, including NC1, NC2, NC3, and Comparator libraries. If any value is negative, the smallest value is used. If all values are positive, the largest value is used.
Target IE-score Statistically aggregated signal from target-related columns, including Her2, Her2_Trastuzumab, Her2_Pertuzumab, and Her2_RabMAb libraries. If any value is negative, the smallest value is used. If all values are positive, the largest value is used.

Cluster analysis

Cluster analysis was performed using USEARCH [2]. Unique VHH sequences were clustered based on 93% identity and numbered in descending order of clone size. Due to the continuous nature of sequence homology, the order of cluster numbers may not always correspond exactly to cluster size.

References

[1] Tsuruta, H., Yamazaki, H., Maeda, R., Tamura, R., Wei, J.N., Mariet, Z., Phloyphisut, P., Shimokawa, H., Ledsam, J.R., Colwell, L., Imura, A. AVIDa-hIL6: A large-scale VHH dataset produced from an immunized alpaca for predicting antigen-antibody interactions. Advances in Neural Information Processing Systems 36, 2023.

[2] Edgar, R. C. Search and clustering orders of magnitude faster than BLAST. Bioinformatics 26, 2460-2461, 2010.

Materials

Trastuzumab, Anti-erbB-2 (Her-2/neu) (4D5-8), Absolute Antibody.

Pertuzumab, Anti-HER2, #A2008, Selleck Chemicals.

Anti-Her2 Rabbit Monoclonal Antibodies, Her2/ErbB2 (D8F12), #4290, Cell Signaling.

Access

This dataset is available by request only.

To request access, users must be logged into a Hugging Face account and submit the access request form. Hugging Face collects the user’s account information and email address by default. The SEPIQ access request form also asks for:

  • full name;
  • affiliation, organization, or independent researcher status;
  • participant type;
  • application type;
  • team or participant name;
  • intended use;
  • whether they would like to be considered for additional GPU resources for Track B, if available;
  • agreement to the SEPIQ data use terms.

Submitting a request does not automatically grant access. Access requests are reviewed manually by the SEPIQ organizing team.

If you are applying as an independent researcher and do not have an institutional or company affiliation, please write Independent researcher in the affiliation field.

If you are applying as an individual participant and do not have a team name, please write your own name in the Team or participant name field.

What this repository may contain

This repository is intended only for approved training data that the organizers agree can be shared with accepted participants.

Possible contents may include:

Data type Status
HER2 VHH training sequences Uploaded
Metadata and labels Uploaded
Starter/example files In preparation
Submission templates To be added

What must NOT be uploaded here

This repository must not contain hidden evaluation data.

The following data should remain private on the organizer side and should not be uploaded to this repository:

  • hidden Cryo-EM structures;
  • final test labels;
  • blind evaluation ground truth;
  • final epitope/paratope annotations used for scoring;
  • any data that should not be accessible to approved participants.

Data use terms

By requesting access, users agree to:

  • use the dataset only for SEPIQ Challenge participation and related approved non-commercial research;
  • not redistribute, rehost, sell, publish, or share the dataset or substantial subsets of it;
  • not attempt to access, infer, reconstruct, or redistribute hidden evaluation labels or Cryo-EM ground truth;
  • not use the dataset for commercial purposes without written permission from the organizers;
  • cite or acknowledge SEPIQ/COGNANO where applicable in reports, publications, or challenge submissions.

Challenge platform

The main SEPIQ Challenge page, rules, timeline, submission instructions, and leaderboard will be hosted separately on Hugging Face Spaces.

This dataset repository is only for gated access to approved training data.

Status

This dataset repository is currently being prepared by the SEPIQ organizing team.

Downloads last month
90

Space using sepiq-2026/SEPIQ-2026-training-data 1