bio.rodeo
HomeCompetitorsLeaderboardOrganizations
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

© 2026 bio.rodeo. All rights reserved.
Imaging

TopCUP

Chan Zuckerberg Initiative

First-place solution from the CZII CryoET Object Identification Kaggle competition; an ensemble of 3D EfficientNet-encoder U-Nets for multi-class protein particle picking.

Released: 2025

Overview

TopCUP — Top CryoET U-Net Picker — is the first-place solution from the CZII CryoET Object Identification Kaggle competition, a machine learning challenge organized by the Chan Zuckerberg Imaging Institute (CZII) from November 2024 to February 2025. The competition attracted 1,135 participants and tasked them with automating the detection of protein complexes in cryo-electron tomography (cryo-ET) volumes, using a benchmark dataset comprising both simulated and experimental tomograms. TopCUP is hosted on the CZ Virtual Cells Platform (v1.0.1) and integrated with the copick dataset API to enable direct application to any cryo-ET dataset in the CZ ecosystem.

Particle picking — identifying the 3D positions of macromolecular complexes within cellular cryo-ET tomograms — remains one of the primary bottlenecks in structural cell biology. Template matching, the conventional approach, requires a known structural template and becomes unreliable for flexible or heterogeneous complexes. TopCUP addresses this with a heatmap-based segmentation approach: rather than directly predicting particle coordinates, the model generates a 3D density map in which elevated values indicate positions likely to contain a target particle, and coordinates are subsequently extracted by peak detection. This representation is differentiable, robust to partial occlusion, and naturally handles particles of varying orientations without template requirements.

The winning solution was achieved by a competitive machine learning team working outside the cryo-ET domain, illustrating the competition's goal of bringing general ML expertise to bear on biological imaging challenges. The code and trained model weights were subsequently incorporated into the CZII codebase and made available through the Virtual Cells Platform.

Key Features

  • FlexibleUNet with 3D EfficientNet-B3 encoder: Uses the MONAI FlexibleUNet architecture with a 3D EfficientNet-B3 backbone pretrained on natural image data as the encoder, providing rich multi-scale feature representations that generalize to cryo-ET tomographic volumes.
  • Reduced decoder for efficiency: The decoder is trimmed to three upsampling levels rather than the full symmetric U-Net depth, reducing parameter count and inference memory while retaining sufficient spatial resolution for accurate heatmap prediction.
  • Three-model ensemble: Trains three models separately on 6, 12, or 24 experimental tomograms respectively, and combines their heatmap predictions by averaging, achieving an ensemble score of 0.774 on the competition private leaderboard — the top result among all 1,135 entrants.
  • Heatmap-based particle localization: Predicts a continuous 3D probability density rather than discrete bounding boxes or coordinates, making the approach robust to particle overlap and compatible with multi-class simultaneous prediction across six protein species.
  • Copick integration: Fully integrated with the copick storage-agnostic cryo-ET API, enabling inference directly on any tomogram registered in a copick project without custom data loading code.
  • Transfer from simulated to experimental data: Trained using the phantom dataset available through the CZ cryo-ET Data Portal, demonstrating that models trained on realistically simulated cryo-ET data can generalize to experimental tomograms with different noise characteristics and missing wedge geometries.

Technical Details

TopCUP is built on the MONAI deep learning library for medical imaging. The core architecture is the MONAI FlexibleUNet, which decouples the encoder selection from the decoder design. The encoder is a 3D EfficientNet-B3, adapted from its original 2D natural image classification context by inflating 2D convolutions to 3D and fine-tuning on the cryo-ET training data. EfficientNet-B3 provides an efficient compound scaling of network width, depth, and resolution relative to parameter count, making it suitable for 3D volumetric data where memory constraints are more stringent than in 2D imaging.

The decoder uses three upsampling stages with transposed convolutions and skip connections from the encoder at matching resolutions, outputting a multi-channel heatmap volume where each channel corresponds to one of the six target protein species: apoferritin, beta-amylase, beta-galactosidase, 80S ribosome, thyroglobulin, and virus-like particle (VLP). Particle centers are extracted from the predicted heatmaps using local maxima detection with a minimum distance threshold derived from the known approximate sizes of each target complex.

Three models were trained on subsets of the experimental phantom dataset (6, 12, and 24 tomograms respectively), with the full 27-tomogram simulated phantom dataset used for pretraining initialization. Ensemble predictions are computed by averaging heatmaps across all three models before center extraction. The ensemble achieved a final private leaderboard score of 0.774 under the F-beta metric (beta=4, emphasizing recall over precision) used to evaluate particle picking accuracy in the competition.

Applications

TopCUP is most directly applicable to cryo-ET laboratories running subtomogram averaging pipelines for structural determination of macromolecular complexes. Because it was trained specifically on the CZII phantom dataset covering six reference complexes, it is immediately useful for studies involving ribosomes, apoferritin (a common fiducial/calibration standard), beta-galactosidase, thyroglobulin, beta-amylase, or VLPs. For other targets, TopCUP's pretrained weights provide a strong initialization for fine-tuning with a small number of annotated experimental tomograms. The copick integration makes it straightforward to slot TopCUP into existing CZII-ecosystem pipelines alongside tools such as Octopi, SABER, and the CZ cryo-ET Data Portal. The tutorial on the Virtual Cells Platform walks users through the full workflow from data loading to coordinate output in an immediately runnable notebook environment.

Impact

TopCUP demonstrates that state-of-the-art computer vision architectures — when applied with thoughtful problem framing via heatmap regression and ensemble training — can match or exceed domain-specialized cryo-ET pickers on competitive benchmarks, with a first-place score of 0.774 among 1,135 competition entrants. The competition from which it emerged has been credited with substantially accelerating the development of open cryo-ET particle picking tools by engaging the broader ML community. By hosting TopCUP on the Virtual Cells Platform alongside the benchmark dataset and copick integration, CZII makes the winning approach immediately reproducible and usable by the structural biology community without requiring participants to re-implement competition code. As of early 2026, TopCUP has not yet been the subject of a standalone peer-reviewed publication; competition results were documented in an accompanying biorxiv preprint covering lessons learned from the challenge.

Tags

particle pickingsegmentationCNNtransfer learningcryo-ETstructural biology

Resources

GitHub RepositoryOfficial Website