Predict Modality

BABEL is a deep autoencoder that learns modality-specific encoders and decoders for paired scRNA-seq and scATAC-seq measurements. RNA is encoded/decoded with dense layers; ATAC accessibility is encoded with a chromosome-split architecture (per-chromosome sub-networks, matching genomic locality) that groups peaks by chromosome parsed from their genomic-coordinate names (e.g. "chr1:1000-2000"). The full architecture is bidirectional (RNA->RNA, RNA->ATAC, ATAC->RNA, ATAC->ATAC), trained jointly with a negative binomial reconstruction loss for RNA counts and a binary cross-entropy loss for binarized ATAC accessibility, but this component only exposes the ATAC->RNA (GEX) prediction direction. The RNA->ATAC direction was found to collapse to a per-peak base-rate prediction that ignores the RNA input -- a known failure mode of unweighted binary cross-entropy under the extreme class imbalance typical of ATAC accessibility data (~3% positive) -- confirmed via near-zero discrimination between accessible/inaccessible peaks even after training to convergence, and is intentionally disabled rather than silently returned.

CellMapper is a general framework for k-NN based mapping tasks in single-cell and spatial genomics. This variant uses CellMapper to project modalities from a reference dataset (train) onto a query dataset (test) in a PCA/CCA latent space.

CellMapper is a general framework for k-NN based mapping tasks in single-cell and spatial genomics. This variant uses CellMapper to project modalities from a reference dataset (train) onto a query dataset (test) in a modality-specific latent space computed with suitable scvi-tools models. For gene expression data, we use the scVI model on raw counts (nb likelihood), for ADT data, we use the scVI models on normalized counts (gaussian likelihood), and for ATAC data, we use the PeakVI model on raw counts. The actual CellMapper pipeline is modality-agnostic.

This is a solution developed by Team Guanlab - dengkw in the Neurips 2021 competition to predict one modality from another using kernel ridge regression (KRR) with RBF kernel. Truncated SVD is applied on the combined training and test data from modality 1 followed by row-wise z-score normalization on the reduced matrix. The truncated SVD of modality 2 is predicted by training a KRR model on the normalized training matrix of modality 1. Predictions on the normalized test matrix are then re-mapped to the modality 2 feature space via the right singular vectors.

A linear model regression method.

This method trains an encoder-decoder MLP model with one output neuron per component in the target. As an input, the encoders use representations obtained from ATAC and GEX data via LSI transform and raw ADT data. The hyperparameters of the models were found via broad hyperparameter search using the Optuna framework.

scButterfly (Basic variant, scButterfly-B) is a dual variational autoencoder with an adversarial translator that learns to convert between paired single-cell modalities. For the predict-modality task it is trained on paired Multiome cells and used to translate the held-out test modality. Chromosome grouping for the ATAC branch is parsed from peak coordinates in the feature names. Only Multiome GEX<->ATAC is supported; CITE-seq (ADT) datasets are not handled.

Two-stage method from the OpenProblems NeurIPS 2021 competition. Stage 1 trains four LightGBM models on different RNA feature representations (log-normalized, CLR-TSVD, custom sqrt-normalized, and raw counts). Stage 2 refines predictions with two neural network architectures: a bidirectional GRU with cosine-similarity loss and a dense bidirectional GRU with MSE loss. Final predictions are a weighted blend (55% cosine, 45% MSE) of per-fold averaged outputs.

This folder contains the AXX solution to the OpenProblems-NeurIPS2021 Single-Cell Multimodal Data Integration. Team took the 4th place of the modality prediction task in terms of overall ranking of 4 subtasks: namely GEX to ADT, ADT to GEX, GEX to ATAC and ATAC to GEX. Specifically, our methods ranked 3rd in GEX to ATAC and 4th in GEX to ADT. More details about the task can be found in the competition webpage.

Encoder-decoder MLP method using SVD-based dimensionality reduction for both inputs and targets, followed by batch-median correction. The encoder maps (optionally augmented) cell embeddings to a latent space; multiple decoder blocks predict target expression in the SVD-compressed space. The method was the winning solution of the NeurIPS 2021 Open Problems Multimodal Single-Cell Integration Kaggle competition.

The average difference between the expression values and the predicted expression values.

The mean of the pearson values of per-cell expression value vectors. A per-cell vector with zero variance contributes a correlation of 0 rather than NA.

The mean of the pearson values of per-gene expression value vectors. A per-gene vector with zero variance contributes a correlation of 0 rather than NA.

The mean of the spearman values of per-cell expression value vectors. A per-cell vector with zero variance contributes a correlation of 0 rather than NA.

The mean of the spearman values of per-gene expression value vectors. A per-gene vector with zero variance contributes a correlation of 0 rather than NA.

The pearson correlation between the solution and the prediction, with both matrices flattened into a single vector. A constant matrix scores 0 rather than NA.

The spearman correlation between the solution and the prediction, with both matrices flattened into a single vector. A constant matrix scores 0 rather than NA.

The square root of the mean of the square of all of the error.

Single-cell CITE-Seq data collected from bone marrow mononuclear cells of 12 healthy human donors using the 10X 3 prime Single-Cell Gene Expression kit with Feature Barcoding in combination with the BioLegend TotalSeq B Universal Human Panel v1.0. The dataset was generated to support Multimodal Single-Cell Data Integration Challenge at NeurIPS 2021. Samples were prepared using a standard protocol at four sites. The resulting data was then annotated to identify cell types and remove doublets. The dataset was designed with a nested batch layout such that some donor samples were measured at multiple sites with some donors measured at a single site.

Single-cell CITE-Seq data collected from bone marrow mononuclear cells of 12 healthy human donors using the 10X 3 prime Single-Cell Gene Expression kit with Feature Barcoding in combination with the BioLegend TotalSeq B Universal Human Panel v1.0. The dataset was generated to support Multimodal Single-Cell Data Integration Challenge at NeurIPS 2021. Samples were prepared using a standard protocol at four sites. The resulting data was then annotated to identify cell types and remove doublets. The dataset was designed with a nested batch layout such that some donor samples were measured at multiple sites with some donors measured at a single site.

Single-cell CITE-Seq data collected from bone marrow mononuclear cells of 12 healthy human donors using the 10X Multiome Gene Expression and Chromatin Accessibility kit. The dataset was generated to support Multimodal Single-Cell Data Integration Challenge at NeurIPS 2021. Samples were prepared using a standard protocol at four sites. The resulting data was then annotated to identify cell types and remove doublets. The dataset was designed with a nested batch layout such that some donor samples were measured at multiple sites with some donors measured at a single site.

Single-cell CITE-Seq data collected from bone marrow mononuclear cells of 12 healthy human donors using the 10X Multiome Gene Expression and Chromatin Accessibility kit. The dataset was generated to support Multimodal Single-Cell Data Integration Challenge at NeurIPS 2021. Samples were prepared using a standard protocol at four sites. The resulting data was then annotated to identify cell types and remove doublets. The dataset was designed with a nested batch layout such that some donor samples were measured at multiple sites with some donors measured at a single site.

Single-cell CITE-Seq data collected from bone marrow mononuclear cells of 12 healthy human donors using the 10X 3 prime Single-Cell Gene Expression kit with Feature Barcoding in combination with the BioLegend TotalSeq B Universal Human Panel v1.0. The dataset was generated to support Multimodal Single-Cell Data Integration Challenge at NeurIPS 2022. Samples were prepared using a standard protocol at four sites. The resulting data was then annotated to identify cell types and remove doublets. The dataset was designed with a nested batch layout such that some donor samples were measured at multiple sites with some donors measured at a single site.

Single-cell CITE-Seq data collected from bone marrow mononuclear cells of 12 healthy human donors using the 10X 3 prime Single-Cell Gene Expression kit with Feature Barcoding in combination with the BioLegend TotalSeq B Universal Human Panel v1.0. The dataset was generated to support Multimodal Single-Cell Data Integration Challenge at NeurIPS 2022. Samples were prepared using a standard protocol at four sites. The resulting data was then annotated to identify cell types and remove doublets. The dataset was designed with a nested batch layout such that some donor samples were measured at multiple sites with some donors measured at a single site.

Single-cell CITE-Seq data collected from bone marrow mononuclear cells of 12 healthy human donors using the 10X Multiome Gene Expression and Chromatin Accessibility kit. The dataset was generated to support Multimodal Single-Cell Data Integration Challenge at NeurIPS 2022. Samples were prepared using a standard protocol at four sites. The resulting data was then annotated to identify cell types and remove doublets. The dataset was designed with a nested batch layout such that some donor samples were measured at multiple sites with some donors measured at a single site.

Single-cell CITE-Seq data collected from bone marrow mononuclear cells of 12 healthy human donors using the 10X Multiome Gene Expression and Chromatin Accessibility kit. The dataset was generated to support Multimodal Single-Cell Data Integration Challenge at NeurIPS 2022. Samples were prepared using a standard protocol at four sites. The resulting data was then annotated to identify cell types and remove doublets. The dataset was designed with a nested batch layout such that some donor samples were measured at multiple sites with some donors measured at a single site.