🚀 @SBERLOGACOMPETE webinar on Kaggle competition by WINNER of the last year similar challenge:
👨🔬 Nikolai Russkikh "Solutions of Predict Modality and Match modality tasks at Multimodal single cell data integration challenge (NeurIPS 2021)"
⌚️ Thursday 15 September, 18.00 (Moscow Time)
Speaker: Russkikh Nikolai, Head of Data Science Department, NOVEL LLC. NOVEL - Bio IT company, developer of the software for NGS data analysis, protein analysis and medicine. Provider of professional bioinformatics services. (Nikolay's team won the first place in the Predict Modality task and took a second place overall in the Match Modality nomination).
Abstract: With recent advances in single-cell genomic technologies, the amount of data regarding different aspects of the cellular state grows rapidly. While promising impactful insights regarding human health and disease, analysis of these data is challenging due to such factors as their high dimensionality, technically imposed factors of variation and sparsity. Taking part in the NeurIPS 2021 Multimodal single cell data integration challenge, we proposed two methods for two different tracks in this competition: prediction of missing data modalities and aligning of the samples of different modalities measured independently. Our first approach took the first place in one nomination in modality prediction track, the second approach was a runner-up in all nominations, standing out of other methods along with the winner. Our modality prediction method is based on the autoencoder neural networks and LSI dimensionality reduction. The proposed method of modalities alignment is based on deep metric learning and is inspired by contrastive language-image pretraining approach. In order to obtain a valid one-to-one matching out of pairwise embedding distance matrix, we find a maximum weight bipartite matching.
📹 Video: https://youtu.be/dS0p3e-Je90
📖 Presentation: https://t.me/sberlogacompete/1925
✔️ The paper: https://www.biorxiv.org/content/10.1101/2022.04.11.487796v1
👨🔬 Nikolai Russkikh "Solutions of Predict Modality and Match modality tasks at Multimodal single cell data integration challenge (NeurIPS 2021)"
⌚️ Thursday 15 September, 18.00 (Moscow Time)
Speaker: Russkikh Nikolai, Head of Data Science Department, NOVEL LLC. NOVEL - Bio IT company, developer of the software for NGS data analysis, protein analysis and medicine. Provider of professional bioinformatics services. (Nikolay's team won the first place in the Predict Modality task and took a second place overall in the Match Modality nomination).
Abstract: With recent advances in single-cell genomic technologies, the amount of data regarding different aspects of the cellular state grows rapidly. While promising impactful insights regarding human health and disease, analysis of these data is challenging due to such factors as their high dimensionality, technically imposed factors of variation and sparsity. Taking part in the NeurIPS 2021 Multimodal single cell data integration challenge, we proposed two methods for two different tracks in this competition: prediction of missing data modalities and aligning of the samples of different modalities measured independently. Our first approach took the first place in one nomination in modality prediction track, the second approach was a runner-up in all nominations, standing out of other methods along with the winner. Our modality prediction method is based on the autoencoder neural networks and LSI dimensionality reduction. The proposed method of modalities alignment is based on deep metric learning and is inspired by contrastive language-image pretraining approach. In order to obtain a valid one-to-one matching out of pairwise embedding distance matrix, we find a maximum weight bipartite matching.
📹 Video: https://youtu.be/dS0p3e-Je90
📖 Presentation: https://t.me/sberlogacompete/1925
✔️ The paper: https://www.biorxiv.org/content/10.1101/2022.04.11.487796v1