In this competition, co-occuring with the NeurIPS 2021 conference, three problems were presented related to multimodal single cell RNA-seq sequencing data. A first-of-its-kind dataset was compiled to compare participants' solutions.
Participants could choose to solve any of the three problems:
- Given one modality, predict the other. For example, based on the availability of chromatin measured in tens of thousands of cells, one needs to predict RNAexpression for each cell.
- Modality matching. The task is to establish pairwise correspondences between the data of two different modalities, measured in the same cells.
- Representation learning. The task is to reduce the dimension of multimodal data so as to preserve as much as possible information about the biological properties of cells, annotated by experts.
The NOVEL team participated in the first two tasks. Fully connected neural networks were used in the modality prediction task; for the task of modality matching, ideas from the CLIP model developed by OpenAI to match images with their descriptions, were used.The solution to the first problem, proposed by our team, brought victory in the prediction of RNA expression by protein expression subtask. In the match modality task, our team was runner-up in all subtasks.
Congratulations to our team on the victory!
About the competition: As part of the 35th conference NeurIPS 2021 (Conference on Neural Information Processing Systems), an international competition was announced for solving problems in the field of bioinformatics and applied genomics.
Organizers and sponsors of the competition: Cellarity, Chan Zuckerberg Initiative and Chan Zuckerberg BIOHUB, Yale University and the Helmholtz Institute (Munich).
The competition was attended by 276 teams from around the world. The results of the competition are published on the website of the competition and reported at the conference.