keju: powerful and accurate inference in Massively Parallel Reporter Assays

The paper introduces **keju**, a hierarchical statistical model that significantly improves the sensitivity and false positive rate control of Massively Parallel Reporter Assay (MPRA) inference by accounting for batch-specific and modality-specific uncertainties through conditioning on DNA counts.

The Gist

The Big Picture: Testing Thousands of "Genetic Switches"

Imagine you are a master electrician trying to figure out which of 10,000 light switches in a massive, dark warehouse actually turn on the lights. You can't just flip them one by one; it would take forever. Instead, you build a machine that flips all 10,000 switches at once and measures how bright the room gets.

In biology, this is what scientists do with MPRAs (Massively Parallel Reporter Assays). They test thousands of tiny DNA snippets (the "switches") to see if they turn on genes (the "lights"). They do this by measuring two things:

1. The DNA: How many switches did you actually put in the room? (The input).
2. The RNA: How many lights actually turned on? (The output).

The goal is to calculate the ratio: RNA / DNA. If you have a lot of DNA but very little RNA, that switch is broken. If you have a lot of RNA, that switch is powerful.

The Problem: The "Noisy" Mess

The problem is that this experiment is incredibly messy. It's like trying to hear a whisper in a hurricane.

1. The "Input" is Clean, The "Output" is Messy: Counting the DNA switches is easy and precise (like counting bricks in a pile). But counting the RNA lights is noisy because biology is chaotic (like trying to count how many people are dancing in a crowded, flashing disco).
2. The "Batch" Effect: Sometimes, the experiment is run on a Tuesday, and sometimes on a Friday. Maybe the lab temperature changed, or the machine was slightly different. These "batches" introduce extra noise that confuses the results.

The Old Way: Previous computer programs (like MPRAnalyze) tried to solve this by using a "one-size-fits-all" rule. They assumed the noise in the DNA was the same as the noise in the RNA, and that the noise was the same for every batch. This is like assuming the static on a radio is the same whether you are listening to a classical station or a rock station, or whether you are in a quiet library or a busy street. It's a bad guess, and it leads to false alarms (thinking a switch works when it doesn't) or missed opportunities (missing a switch that actually works).

The Solution: Enter "Keju"

The authors created a new tool called Keju (named after the Danish word for "cheese," perhaps implying it's a "grating" tool that separates the good from the bad, or just a catchy name).

Think of Keju as a super-smart, custom-tailored detective that looks at the data differently:

1. The "Fixed Anchor" Strategy

Keju realizes that counting the DNA switches is so precise that we can treat it as a fixed anchor. It says, "I trust the DNA count completely. I don't need to worry about its noise." Instead, it focuses 100% of its energy on modeling the messy RNA noise.

• Analogy: Imagine you are trying to measure how much water a sponge holds. You know exactly how heavy the dry sponge is (DNA). You don't need to guess that weight; you just focus entirely on measuring the water (RNA) accurately.

2. The "Batch-Specific" Glasses

Keju puts on different glasses for every "batch" of the experiment. It realizes that the Tuesday batch might be noisier than the Friday batch. By modeling them separately, it stops the noise from one batch from ruining the results of another.

• Analogy: If you are judging a singing contest, you don't use the same volume setting for a singer in a small room as you do for a singer in a stadium. Keju adjusts the volume (noise level) for every specific room (batch).

3. The "Group Hug" (Shrinkage)

Keju notices that some switches are very similar (they look for the same pattern, called a "motif"). If one switch is a bit weird, Keju says, "Hey, you look like your neighbor. Let's borrow a little bit of their data to make our guess more stable."

• Analogy: If you are trying to guess the average height of a group of basketball players, and one guy is wearing giant platform shoes, you don't just ignore him. You look at the other players to get a better sense of the "real" average, then adjust your guess for the guy in shoes. This prevents one weird data point from ruining the whole analysis.

Why Does This Matter? (The Results)

The authors tested Keju against the old methods using fake data (simulations) and real lab data.

• Sensitivity (Finding the Good Stuff): Keju found 59% of the real "active switches." The old methods only found 31% and 9%.
  • Translation: Keju is much better at finding the "golden needles" in the haystack.
• False Positives (Avoiding False Alarms): The old methods were very noisy. They thought 34% of the broken switches were working! Keju only thought 6.8% were working.
  • Translation: Keju is much more reliable. It doesn't waste your time chasing ghosts.

The "Promoter" Twist

The paper also looked at how different "bases" (called minimal promoters) affect the switches. They found that some bases act like a "volume booster." Keju is smart enough to realize, "Oh, this switch looks loud, but that's just because it's sitting on a loud speaker (the promoter), not because the switch itself is better." It separates the volume of the speaker from the quality of the switch.

The Bottom Line

Keju is a new, smarter way to analyze genetic experiments. By admitting that DNA is clean, RNA is messy, and every experiment batch is unique, it cuts through the noise.

• Old Way: "Let's guess the noise level for everything and hope for the best."
• Keju Way: "Let's measure the noise for every specific part of the experiment, group similar things together, and get a much clearer picture."

This means scientists can now discover new genetic switches that were previously hidden in the noise, helping us understand diseases and design better medicines with more confidence.

Technical Summary

1. Problem Statement

Massively Parallel Reporter Assays (MPRAs) are high-throughput experiments used to quantify the regulatory function of thousands of genetic elements by linking DNA input counts to RNA output counts via unique barcodes. However, statistical inference in MPRAs is complicated by several sources of uncertainty:

• Disparate Uncertainty Levels: DNA counts (representing transfection efficiency) generally have lower uncertainty than RNA counts (representing transcription, degradation, and other biological noise).
• Batch Effects: Variability exists between experimental batches (e.g., different sequencing runs or treatment conditions).
• Limitations of Existing Methods: Current state-of-the-art tools like MPRAnalyze and BCalm often fail to account for these nuances. Specifically, MPRAnalyze shares a single overdispersion parameter across both DNA and RNA counts and across all batches, leading to suboptimal variance estimation. This results in reduced statistical power (missing true effects) and poor false positive rate (FPR) control (calling noise as significant).

2. Methodology: The keju Model

The authors present keju, a Bayesian hierarchical statistical model designed specifically to address the unique noise structure of MPRA data.

Core Assumptions & Architecture:

1. Fixed DNA Offsets: keju treats DNA counts as fixed offsets with negligible uncertainty. It models uncertainty only in the RNA counts. This simplifies the model from a two-GLM structure (used by MPRAnalyze) to a single Negative Binomial Generalized Linear Model (GLM) on RNA counts.
2. Modality-Specific Uncertainty: Unlike previous methods that pool uncertainty across modalities, keju estimates overdispersion parameters specifically for RNA counts, acknowledging that biological noise is higher in RNA than in DNA.
3. Batch-Specific Overdispersion: The model estimates distinct overdispersion parameters for each experimental batch, capturing batch-level variability rather than assuming a global variance.
4. Mean-Count Grouping (Binning): To stabilize overdispersion estimation, keju bins enhancers with similar mean RNA counts (default bin size $G=50$) and shares a single overdispersion estimate ($\vartheta_g$) within each bin. This addresses the mean-variance trend common in count data.
5. Hierarchical Shrinkage:
   • Motif-Level Shrinkage: Enhancers targeting the same transcription factor motif are shrunk toward a shared motif-level mean and variance, improving power for weak effects.
   • Promoter-Level Modeling: For experiments using multiple minimal promoters, keju can model promoter-specific intercepts and slopes to disentangle promoter effects from motif effects, allowing for the prediction of unseen promoter-motif combinations.
6. Covariate Correction: The model uses experimental negative controls to set covariate-specific baselines, correcting for systematic biases (e.g., differences caused by minimal promoter choice).

Implementation:
The model is implemented using Stan with Hamilton Monte Carlo (HMC) sampling (NUTS algorithm).

3. Key Contributions

• Novel Statistical Framework: Introduced a hierarchical Bayesian model that decouples DNA and RNA uncertainty and explicitly models batch-specific variance, a feature missing in leading competitors.
• Improved Variance Estimation: Demonstrated that modeling modality-specific and batch-specific overdispersion leads to more accurate uncertainty quantification compared to shared-parameter models.
• Flexible Experimental Design Handling: The model supports paired and pooled designs, multiple minimal promoters, and motif-level regularization, making it adaptable to complex synthetic biology experiments.
• Open Source Tool: Released as an R package (keju) with code and data available for reproducibility.

4. Results

The authors benchmarked keju against MPRAnalyze and BCalm using 19 case-control comparisons from the Zahm et al. dataset (involving 6,144 candidate enhancers) and extensive simulations.

A. Simulation Performance (Power):

• keju demonstrated significantly higher sensitivity in detecting true positive effects.
• Recall Rates: keju recalled 59.1% of ground truth significant calls, compared to 31.1% for MPRAnalyze and 9.2% for BCalm.
• Ablation studies confirmed that both motif-level shrinkage and mean-count grouping are critical drivers of this increased power.

B. False Positive Rate (FPR) Control:

• Using masked negative controls (scramble sequences), keju showed superior calibration and robustness.
• Average FPR: keju had an average FPR of 6.8%, compared to 34.2% for MPRAnalyze and 12.2% for BCalm.
• Robustness: While MPRAnalyze and BCalm exhibited extreme FPR outliers (calling >50% of negative controls significant in some datasets), keju maintained a stable FPR (never exceeding 14% median) across all 19 datasets.

C. Biological Insights:

• The model successfully identified that the minCMV minimal promoter drives higher baseline transcription rates and "stretches" effect sizes compared to minTK and minProm.
• By modeling promoter-specific effects, keju allowed for the separation of promoter-level and motif-level contributions, enabling the prediction of transcription rates for unseen combinations.

5. Significance

• Reliability in Synthetic Biology: keju provides a more reliable framework for identifying viable genetic elements for synthetic enhancer construction, reducing the risk of false positives that could derail downstream experimental validation.
• Detection of Weak Effects: The increased statistical power allows researchers to detect subtle regulatory effects that were previously masked by noise in existing methods.
• Methodological Advancement: The paper establishes that treating DNA counts as fixed and modeling batch/modality-specific variance is a superior approach for MPRA data analysis, offering a new standard for high-throughput functional genomics inference.
• Scalability: While currently MCMC-based (taking ~1 day for large datasets), the authors suggest future implementation of variational inference to improve speed for massive datasets.

In conclusion, keju represents a significant leap forward in MPRA analysis by rigorously addressing the specific statistical challenges of the data type, resulting in a tool that is both more powerful in discovery and more robust in error control than its predecessors.