scMultiPreDICT: A single-cell predictive framework with transcriptomic and epigenetic signatures

The paper introduces scMultiPreDICT, a computational framework that systematically benchmarks single-cell transcriptomic and epigenetic features to reveal that while RNA-derived data generally offers superior predictive power for gene expression, the added value of multimodal integration is gene-specific and context-dependent.

The Gist

Imagine a cell as a busy, high-tech factory. Inside this factory, there are two main types of instruction manuals that tell the machines what to build and how to behave:

1. The "Active Blueprint" (RNA): This is the current list of orders on the factory floor. It shows exactly what proteins are being made right now.
2. The "Master Archive" (Chromatin/ATAC): This is the library in the basement. It contains all the potential blueprints, but some are locked in vaults (closed off) and others are sitting on open desks (accessible). Just because a blueprint is in the library doesn't mean it's being used today.

For a long time, scientists thought that if you knew the Active Blueprint (RNA), you could predict exactly what the factory would do next. But recently, we've realized the Master Archive (Chromatin) might be secretly pulling the strings, deciding which blueprints can even be used before the factory floor ever sees them.

The Problem

Scientists have a new tool that lets them read both manuals at the same time for every single cell. But they faced a big question: Which manual is actually more important for predicting the factory's future?

• Is it the current orders (RNA)?
• Is it the open vaults in the library (Chromatin)?
• Or do we need to read both together to get the perfect prediction?

Existing computer programs were good at mixing these two data sources, but they didn't really tell us which source was doing the heavy lifting for specific genes.

The Solution: scMultiPreDICT

The authors of this paper built a new computer framework called scMultiPreDICT. Think of it as a super-smart "Crystal Ball" simulator for cells.

Here is how they tested their crystal ball:

1. The Setup: They took data from three different types of "factories" (Mouse Stem Cells and Human Immune Cells).
2. The Test: For hundreds of specific genes, they tried to predict the future activity using three different strategies:
   • Strategy A (RNA Only): "I'll guess the future based only on what's happening on the factory floor right now."
   • Strategy B (Chromatin Only): "I'll guess the future based only on which blueprints are unlocked in the library."
   • Strategy C (The Combo): "I'll guess the future by reading both manuals together."
3. The Models: They used six different types of "predictors" (from simple math equations to complex AI neural networks) to see which one worked best.

The Surprising Results

1. The "Active Blueprint" (RNA) is the MVP.
When they looked at the results, the RNA-only strategy was the clear winner. Knowing what genes are currently active was the strongest predictor of what a gene would do next. It's like saying, "If the factory is currently building cars, it's highly likely it will keep building cars."

2. The "Master Archive" (Chromatin) is a supporting actor.
Reading only the library (Chromatin) gave a decent guess, but it wasn't as accurate as reading the factory floor. It's like trying to guess what a factory will build just by looking at its library; you might get the general idea, but you'll miss the specific details of the current orders.

3. The "Combo" isn't always better.
The biggest surprise? Reading both manuals together didn't always make the prediction better.

• For some genes, adding the library info helped a little.
• For many others, it didn't help at all.
• For a few, it actually made the prediction slightly worse (because the library info was "noise" or confusing).

The Analogy: Imagine trying to predict if it will rain tomorrow.

• RNA is looking at the clouds right now. (Very accurate).
• Chromatin is looking at the weather forecast for next week. (Helpful, but not as immediate).
• The Combo is looking at both.
• The Finding: Sometimes, looking at next week's forecast doesn't help you predict tomorrow's rain any better than just looking at the clouds right now. In fact, for some specific days, the forecast is just wrong or irrelevant.

The "Gene-Specific" Twist

The most important discovery is that every gene is different.

• For some genes (like Etv6 in stem cells), the factory floor (RNA) is the only thing that matters.
• For other genes (like RUNX3 in immune cells), the library (Chromatin) plays a huge role. The "open vaults" in the library are just as important as the current orders.

Why Does This Matter?

This framework is like a diagnostic tool for drug developers.

If a scientist wants to stop a disease-causing gene, they need to know where to intervene.

• If the gene is driven mostly by RNA, they should try to block the factory floor (stop the protein production).
• If the gene is driven by Chromatin, they need to go to the library and lock the vaults (change the accessibility).

scMultiPreDICT tells scientists exactly which "switch" to flip for each specific gene, saving time and money on failed experiments. It proves that biology isn't "one size fits all"; sometimes you need to change the current orders, and sometimes you need to reorganize the library.

In a Nutshell

The authors built a tool to test if knowing the "library" (chromatin) helps us predict the "factory floor" (gene expression) better than just looking at the floor alone. They found that the floor is usually the best predictor, but for specific genes, the library matters a lot. Their tool helps scientists figure out which rule applies to which gene, guiding them to better treatments.

Technical Summary

1. Problem Statement

Understanding how cellular responses to genetic perturbations are determined requires dissecting the relative contributions of transcriptional programs (gene expression) and the epigenetic landscape (chromatin accessibility). While single-cell multiomics technologies (scRNA-seq and scATAC-seq) allow for simultaneous profiling of these layers, current computational methods face two main limitations:

• Lack of Systematic Comparison: Existing approaches focus on data integration or network inference but do not systematically compare the predictive power of transcriptional versus epigenetic features on a gene-by-gene basis.
• Unclear Value of Multimodality: It remains unclear whether integrating both modalities (RNA + ATAC) consistently improves gene expression prediction accuracy over single-modality approaches, or if the benefit is context-dependent.

The authors aim to answer: Which regulatory layer (transcriptional or epigenetic) is most informative for individual genes, and does multimodal integration provide added value?

2. Methodology: scMultiPreDICT Framework

The authors developed scMultiPreDICT (single-cell Multimodal Predictor of Gene Expression via Integrated Chromatin and Transcriptome), a computational framework implemented in R. The workflow consists of the following key steps:

A. Data Acquisition and Preprocessing

• Datasets: Three paired single-cell multiomics datasets were used:
  1. Mouse Embryonic Stem Cells (mESC) at day 7.5 (two biological replicates).
  2. Human Peripheral Blood Mononuclear Cells (PBMC), specifically T-cell subsets.
• Quality Control (QC): Independent QC for RNA and ATAC modalities to filter low-quality cells.
• Data Splitting: Cells were randomly partitioned into 70% training, 20% validation, and 10% test sets at the cell level to prevent information leakage.
• Denoising (Metacell Construction): To address sparsity, the authors used k-Nearest Neighbor (kNN) smoothing to construct "metacells." This involved:
  • Dimensionality reduction (PCA for RNA, LSI for ATAC, or joint strategies like WNN, scVI+PeakVI, MultiVI).
  • Averaging gene expression and chromatin accessibility signals across k-nearest neighbors (k=20).
  • Normalizing counts to CPM (Counts Per Million).

B. Feature Engineering

For each target gene, three distinct feature sets were constructed to enable comparative modeling:

1. RNA-only: Expression of the top 1,000 Highly Variable Genes (HVGs) from the training set (excluding the target gene if present).
2. ATAC-only: Chromatin accessibility peaks located within ±250kb of the target gene's Transcription Start Site (TSS), capturing proximal promoters and distal cis-regulatory elements.
3. Multimodal: A concatenation of the RNA-only and ATAC-only feature sets.

C. Model Training and Evaluation

• Algorithms: Six machine learning models were benchmarked:
  • Linear: Ordinary Least Squares (OLS), Ridge, Lasso, Elastic Net.
  • Non-linear: Random Forest (RF), Deep Neural Network (DeepNN).
• Target Genes: Predictions were made for two sets: HVGs and randomly selected non-HVGs.
• Metrics: Performance was evaluated on the held-out test set using:
  • Spearman's rank correlation coefficient.
  • Coefficient of determination ($R^2$).
  • Root Mean Squared Error (RMSE).
• Feature Importance: To handle the imbalance in feature counts (1,000 RNA features vs. ~60 ATAC peaks), the authors introduced a Selection Rate metric. This normalizes importance by calculating the percentage of top 30 features derived from a specific modality relative to the total available features for that modality.

3. Key Results

A. Predictive Performance by Modality

• RNA-only: Consistently provided strong predictive power across all datasets and models. Median Spearman correlations ranged from 0.60 to 0.78 (highest in mESC Replicate 1).
• ATAC-only: Yielded modest performance (Median Spearman correlation: 0.38 to 0.60). While better than random, it significantly underperformed compared to RNA-only models.
• Multimodal Integration: Contrary to the hypothesis that integration would universally improve accuracy, multimodal models did not consistently outperform RNA-only models.
  • The benefit of adding ATAC data was gene-specific and context-dependent.
  • For the majority of genes, adding chromatin features provided little to no improvement, and in some cases, slightly reduced performance.
  • Different integration strategies (PCA+LSI, WNN, scVI+PeakVI, MultiVI) showed similar results, suggesting the integration method itself was not the limiting factor.

B. Model Performance

• Random Forest consistently achieved the highest predictive performance across all datasets and feature sets.
• DeepNN generally outperformed regularized linear models, likely due to its ability to capture non-linear regulatory associations.
• OLS (unregularized) performed the worst, highlighting the need for regularization in high-dimensional genomic data.

C. Feature Importance and Biological Insights

• Dominance of Transcriptional Features: For most genes (especially in mESC), RNA-derived features dominated the top predictors.
• Context-Specific Epigenetic Contribution: In T cells, ATAC-derived features contributed significantly more, reaching levels comparable to RNA features for certain genes.
• Case Studies:
  • Etv6 and Tbx3 (mESC): Predicted primarily by transcriptional regulators (e.g., Pbx1, Prdm6).
  • RUNX3 (T cells): Showed a mixed signature where a local ATAC peak was the second most important predictor alongside transcription factors (LEF1, TSHZ2), indicating a strong role for cis-regulatory elements in this specific cellular context.

4. Key Contributions

1. Systematic Benchmarking: Provides the first systematic, gene-by-gene comparison of transcriptional vs. epigenetic predictive power using diverse machine learning models.
2. Quantification of Regulatory Layers: Demonstrates that while transcriptional data is the primary driver of gene expression prediction, chromatin accessibility provides meaningful, albeit selective, predictive power for specific genes in specific cell types.
3. Novel Metric (Selection Rate): Introduces a method to fairly compare feature importance between modalities with vastly different feature counts.
4. Open-Source Framework: Releases scMultiPreDICT as an R package to facilitate the identification of regulatory drivers for targeted perturbation studies.

5. Significance and Implications

• Therapeutic Prioritization: The framework helps researchers decide whether to target gene regulatory networks (transcriptional) or chromatin accessibility (epigenetic) for specific therapeutic interventions. If a gene's expression is best predicted by ATAC peaks, modulating chromatin accessibility may be more effective.
• Perturbation Study Design: Guides the design of CRISPR or small-molecule experiments by identifying which regulatory layer is most informative for a target gene.
• Biological Understanding: Confirms that chromatin accessibility is a permissive but not deterministic factor for gene expression. The "time lag" between chromatin opening and transcriptional activation, along with other unmeasured epigenetic factors (e.g., methylation), likely explains why ATAC-only models underperform.
• Data Integration Strategy: Suggests that simply concatenating modalities does not guarantee better predictions; the value of multimodal data is highly specific to the biological context and the gene of interest.