RNA Surface Segmentation Dataset

This dataset contains 640 triangulated surface meshes of the molecular envelopes for RNA molecules, gathered from the PDB database. Each vertex is assigned a ground-truth segmentation label according to ~120 functional categories (see the quoted blurb below for a full description and collection process). We use this data as a representative task for 3D machine learning on surfaces, predicting the functional segmentation from only the molecule surface shape.

From: Effective Rotation-invariant Point CNN with Spherical Harmonics Kernels, Adrien Poulenard, Marie-Julie Rakotosaona, Yann Ponty, Maks Ovsjanikov, in 3DV 2019.

Please cite this dataset as:

@inproceedings{poulenard2019effective,
  title={Effective rotation-invariant point cnn with spherical harmonics kernels},
  author={Poulenard, Adrien and Rakotosaona, Marie-Julie and Ponty, Yann and Ovsjanikov, Maks},
  booktitle={2019 International Conference on 3D Vision (3DV)},
  pages={47--56},
  year={2019},
  organization={IEEE}
}

In this repository

• off/ The surface meshes composing the dataset, in the .off file format.
• labels/ Ground-truth segmentations. One text file per mesh in the dataset, with one integer per line giving segmentation IDs per-vertex. The ordering of the lines corresponds to the ordering of vertices in the .off file.
• train.txt / test.txt Lists of filenames giving a recommended 80/20 train/test split, sampled randomly.

Notes about the data

• The segmentation indices range from [-1,258] (a very small number of vertices are labelled as -1). For simplicity, we treat -1 as an additional class, yielding 259 classes for prediction. (note that adding 1 to the class targets yields integers on [0,259] ready for training)

• These meshes are not necessarily nice meshes! They have many disconnected components, poor-quality elements, etc. The meshes are not necessarily in any canonical alignment. These are common challenges of real data, which are important to represent in the benchmark.

• In some cases, you may want to remesh/resample the meshes for facilitate an experiment, such as uniformly-sampling a point cloud from the meshes. New training labels can be collected by nearest-neighbor labels from the original mesh. We strongly encourage that results be projected back to the original meshes to report accuracy, though not all past work has followed this procedure.

Full description

(the following description is quoted from Poulenard et. al. above, please contact the authors therein for further questions)

We also applied our approach to tackle the challenging task of segmenting molecular surfaces into functionally homologous regions within RNA molecules. We considered a family of 640 structures of 5s ribosomal RNAs (5s rRNAs), downloaded from the PDB database [3]. A surface, or molecular envelope, was computed for each RNA model by sweeping the volume around the atomic positions with a small ball of fixed radius. This task was performed using the scripting interface of the Chimera structure-modelling environment [20]. Individual RNA sequences were then globally aligned, using an implementation of the Needleman-Wunsch algorithm [19], onto the RFAM [12] reference alignment RF00001 (5s rRNAs). Since columns in multiple sequence alignments are meant to capture functional homology, we treated each column index as a label, which we assigned to individual nucleotides, and their corresponding atoms, within each RNA. Labels were finally projected onto individual vertices of the surface by assigning to a vertex the label of its closest atom. This results in each shape represented as a triangle mesh consisting of approximately 10k vertices, and its segmentation into approximately 120 regions, typically represented as connected components. Shapes in this dataset are not pre-aligned and can have fairly significant geometric variability arising both from different conformations as well as from the molecule acquisition and reconstruction process (see Fig. 5 for an example).

Papers using this dataset

(create a pull request to add more!)

• "Effective Rotation-invariant Point CNN with Spherical Harmonics Kernels" by Poulenard et al., 3DV 2019
• "DiffusionNet: Discretization Agnostic Learning on Surfaces" by Sharp et al., ACM ToG 2021