PACMON: Pathway-guided Multi-Omics data integration for interpreting large-scale perturbation screens

The paper introduces PACMON, a scalable Bayesian latent factor model that integrates multi-omics data from large-scale perturbation screens to interpret gene function by jointly inferring pathway-level programs and their specific modulation by experimental perturbations.

The Gist

Imagine you are a detective trying to solve a massive mystery in a bustling city called The Cell.

In this city, there are millions of workers (genes and proteins) doing their jobs. Sometimes, you want to see what happens if you fire a specific worker, give them a new tool, or change their shift schedule. In science, this is called a perturbation.

For years, scientists have been able to fire thousands of workers at once and watch the city's reaction using high-tech cameras (single-cell sequencing). But here's the problem: The data they get back is a chaotic, overwhelming mess. It's like watching a stadium of 100,000 people and trying to figure out exactly who changed their mind, who started a chant, and why, just by looking at a blurry photo of the whole crowd.

Existing tools tried to solve this by either:

1. Grouping workers into random clusters without knowing why they grouped together.
2. Trying to guess the cause of the chaos without using a map of the city's rules.

Both approaches left scientists confused. They had lists of affected genes, but no clear story about biological pathways (the organized teams or departments within the cell that actually do the work).

Enter PACMON: The "Smart City Planner"

The paper introduces PACMON, a new computer program designed to make sense of this chaos. Think of PACMON as a super-intelligent city planner who doesn't just look at the crowd, but knows the city's blueprints (biological pathways) and can instantly tell you which "departments" were affected by your changes.

Here is how PACMON works, using simple analogies:

1. The "Department" System (Pathway-Guided)

Instead of looking at individual genes one by one, PACMON looks at them as teams.

• The Analogy: Imagine the cell isn't a list of 20,000 individual workers, but a company with departments like "Security," "Construction," "Energy," and "Marketing."
• How it works: PACMON uses a pre-existing map (like a Hallmark directory of known biological pathways) to say, "Okay, these 50 genes work in the 'Immune Response' department."
• The Magic: When you disrupt the cell, PACMON doesn't just say "Gene X changed." It says, "The Immune Response Department was turned off." This makes the results instantly understandable.

2. The "Noise-Canceling" Headphones (Structured Sparsity)

Real-world data is messy. Sometimes a gene in the "Immune Department" gets a little noisy, or a gene from the "Construction Department" accidentally joins the Immune team.

• The Analogy: Imagine you are trying to hear a specific conversation in a loud bar. Most people are shouting (noise), but you know exactly who is sitting at the "Immune Table."
• How it works: PACMON wears "noise-canceling headphones." It focuses heavily on the genes it expects to be in a team (based on the map). If a gene outside that team starts acting like it belongs there, PACMON listens closely. If the evidence is strong, it says, "Wait, maybe this gene should be in this team!" It refines the map as it goes, making the team definitions more accurate than before.

3. The "Multimodal" Detective (Seeing the Whole Picture)

Old tools often only looked at one type of data, like the "RNA" (the blueprints). But cells also have "Proteins" (the actual machinery).

• The Analogy: Imagine trying to understand a car engine. Looking only at the blueprints (RNA) is helpful, but looking at the actual gears turning (Proteins) tells you if the engine is actually running.
• How it works: PACMON is a multimodal detective. It looks at the blueprints and the gears simultaneously. It can tell you, "The blueprint for the immune response is loud, and the immune gears are spinning fast." This gives a much clearer, more reliable picture of what's happening.

4. The "Speed Demon" (Scalability)

The biggest breakthrough in this paper is speed.

• The Analogy: Previous tools were like a librarian trying to sort 100 million books by hand. It would take them years.
• How it works: PACMON is like a high-speed sorting robot. It uses a mathematical trick called Variational Inference (think of it as a smart shortcut that guesses the answer very quickly and gets it right) to handle 100 million cells in a single run.
• The Result: The authors tested this on the Tahoe-100M dataset (a massive library of 100 million cells and 1,000 drug combinations). No other tool could handle this size. PACMON mapped out how different drugs affect different "departments" of the cell, revealing patterns that were previously invisible.

The Real-World Wins

The paper shows PACMON solving three major mysteries:

1. The Simulation Test: They created a fake city with a known answer. PACMON found the answer almost perfectly, beating all other tools in both accuracy and speed.
2. The Melanoma Mystery: They looked at skin cancer cells. They discovered that when they blocked a specific signal (the "IFN-gamma" signal), the cancer cells stopped showing their "immune flags" (proteins like PD-L1). PACMON connected the dots between the genetic switch and the protein flag instantly.
3. The Drug Atlas: They tested 1,000 drugs. PACMON showed that a "JAK inhibitor" drug specifically shuts down the "Interferon Department," while an "mTOR inhibitor" shuts down the "Growth Department." It even showed how changing the dose of the drug changes the intensity of the shutdown.

The Bottom Line

PACMON is a new tool that turns a chaotic, overwhelming list of genetic changes into a clear, organized story about biological teams.

• It uses maps to know what the teams are.
• It uses smart shortcuts to handle massive amounts of data (100 million cells!).
• It looks at multiple types of evidence (RNA and proteins) at the same time.

In short, PACMON helps scientists stop drowning in data and start understanding the story of how our cells react to drugs and diseases. It turns a blurry, noisy photo of a crowd into a high-definition movie of a coordinated team effort.

Technical Summary

1. Problem Statement

High-throughput perturbation screens (e.g., CRISPR-Cas9, drug screens) coupled with single-cell multi-omics profiling (Perturb-seq, Perturb-CITE-seq) generate massive datasets. However, interpreting these data remains challenging due to three main limitations in existing methods:

• Interpretability: Traditional latent factor models (e.g., PCA, standard NMF) identify gene modules but do not inherently link them to known biological pathways, requiring cumbersome post-hoc annotation.
• Perturbation Modeling: Existing approaches often treat perturbations as covariates in a second step or fail to jointly model the perturbation effects on latent programs, limiting the ability to quantify specific pathway modulation.
• Scalability: State-of-the-art methods like GSFA (Guided Sparse Factor Analysis) rely on Markov Chain Monte Carlo (MCMC) inference, which is computationally prohibitive for modern atlas-scale datasets (millions of cells).
• Modality Limitations: Many tools are restricted to transcriptomic data and cannot naturally integrate multimodal measurements (e.g., RNA + surface proteins).

2. Methodology: PACMON

The authors introduce PACMON (Pathway-guided Multi-Omics data integration for interpreting large-scale perturbation screens), a Bayesian latent factor model designed to jointly infer pathway-level programs and their modulation by experimental perturbations.

Core Model Architecture

• Generative Framework: PACMON decomposes high-dimensional molecular measurements from $M$ modalities (e.g., RNA, proteins) into a shared low-dimensional latent space.
  • Latent Factors ($Z$): Represent pathway-level gene programs.
  • Loadings ($W$): Modality-specific matrices mapping latent factors to observed features.
  • Perturbation Coefficients ($B$): A matrix quantifying how each perturbation ($C$) activates or represses each latent factor ($K$).
• Perturbation-Driven Priors: Unlike unsupervised models, PACMON models latent factors as a function of the perturbation design matrix ($P$):
  $$z_i | p_i, B \sim \mathcal{N}(B^\top p_i, I)$$
  This directly links experimental conditions to biological programs.
• Structured Sparsity Priors (Pathway Integration):
  • The model employs a regularized horseshoe prior on the loading matrices ($W$).
  • Informed Priors: Pathway annotations (e.g., Hallmark, Reactome, GO) are encoded as a scaling matrix $\Omega$. Features within a pathway receive a weak prior (allowing contribution), while features outside receive strong shrinkage.
  • Data-Driven Refinement: Crucially, the horseshoe prior allows genes outside the annotated pathway to acquire non-zero loadings if strongly supported by the data, enabling the refinement of pathway definitions.
• Non-Negativity: To ensure biological interpretability (additive gene programs), PACMON applies a ReLU transformation to factor scores and loadings during inference.
• Inference: The model uses Stochastic Variational Inference (SVI) with a mean-field approximation. This replaces slow MCMC sampling with gradient-based optimization, enabling scalability to millions of cells.
• Significance Assessment: Effects are assessed using a unified framework combining the Variational Local False Sign Rate (vLFSR) and a Region-of-Practical-Equivalence (ROPE) criterion to control for false positives and negligible effect sizes.

3. Key Contributions

1. Unified Framework: PACMON is the first perturbation-aware latent factor model that jointly integrates multimodal data (RNA + proteins) with prior pathway knowledge and perturbation design.
2. Scalability: By utilizing stochastic variational inference, PACMON scales to atlas-level datasets (e.g., 100 million cells), a regime where MCMC-based methods like GSFA fail.
3. Interpretability by Design: Instead of inferring abstract factors and annotating them later, PACMON aligns latent factors with specific biological pathways a priori while allowing for data-driven discovery of novel pathway members.
4. Refinement Capability: The model can extend curated pathway definitions by identifying high-weight genes not originally included in the pathway annotation.

4. Results

The authors evaluated PACMON in three settings of increasing complexity:

A. Synthetic Data Benchmarking

• Accuracy: On synthetic data with known ground truth, PACMON achieved near-perfect recovery of pathway structures (AUPR = 0.99) and perturbation effects (AUPR = 0.94), significantly outperforming GSFA (AUPR $\approx$ 0.60–0.65).
• Robustness: PACMON maintained high accuracy even when pathway priors were corrupted with 50% noise.
• Scalability: PACMON demonstrated superior computational efficiency. For 100,000 samples, PACMON took ~1.7 hours, whereas GSFA required ~20.5 hours on the same hardware.

B. Multimodal Perturb-CITE-seq (Melanoma)

• Dataset: ~150,000 cells with RNA and surface protein measurements under CRISPR perturbations, T-cell co-culture, and IFN$\gamma$ treatment.
• Findings:
  • Recovered coherent Interferon Signaling and Cell Cycle programs spanning both RNA and protein modalities.
  • Identified that disruption of the IFN$\gamma$-JAK/STAT cascade (e.g., STAT1, JAK1 KO) specifically suppressed interferon programs, consistent with known immune evasion mechanisms.
  • Demonstrated that CD274 (PD-L1) knockout reduced interferon program activity but modulated broader cytokine signaling, revealing nuanced biological states.
  • Successfully integrated surface protein data (e.g., PD-L1, MHC components) to refine pathway interpretations.

C. Tahoe-100M Atlas

• Dataset: ~100 million cells across 1,135 drug-dose combinations.
• Findings:
  • Produced the first pathway-level latent factor analysis at this scale.
  • Revealed coherent drug-response landscapes: JAK/STAT inhibitors clustered in the interferon-suppressed region, while mTOR inhibitors clustered in the mTORC1-suppressed region.
  • Identified dose-dependent stress and apoptosis signatures for proteasome and translation inhibitors.
  • Refined Hallmark gene sets by adding data-supported genes (e.g., CCNB1 in G2M checkpoint) that were missing from original annotations.

5. Significance

PACMON represents a paradigm shift in the analysis of large-scale perturbation screens.

• From Genes to Pathways: It shifts the analytical focus from individual differentially expressed genes to interpretable, perturbation-modulated biological programs.
• Atlas-Scale Feasibility: It makes it possible to analyze "giga-scale" single-cell atlases (100M+ cells) with biological interpretability, a task previously impossible due to computational constraints.
• Multimodal Integration: It provides a rigorous statistical framework for combining transcriptomics and proteomics to gain a more holistic view of cellular responses.
• Drug Discovery: The ability to map drug-dose combinations to specific pathway landscapes offers a powerful tool for understanding mechanisms of action and identifying off-target effects in large-scale screening campaigns.

In conclusion, PACMON provides a scalable, unified, and interpretable framework that bridges the gap between massive perturbation datasets and biological mechanism discovery.