Statistical end-to-end analysis of large-scale microbial growth data with DGrowthR

The paper introduces DGrowthR, a flexible, non-parametric statistical framework and no-code application designed to efficiently analyze large-scale microbial growth data and identify differential growth patterns without relying on restrictive parametric assumptions.

The Gist

Imagine you are a detective trying to solve a mystery: How do bacteria react when you throw different things at them?

Sometimes you throw antibiotics, sometimes food additives, and sometimes other drugs. To solve this, you need to watch the bacteria grow over time. In the past, scientists used to look at these growth patterns like a rigid checklist. They assumed bacteria always grow in a perfect "S" shape (slow start, fast middle, slow end). If the bacteria acted weirdly—maybe they grew fast then crashed, or grew slowly then sped up—the old tools would get confused or just say, "I can't measure this."

Furthermore, modern robots can now test thousands of chemicals at once, generating massive amounts of data. The old tools were like trying to count every grain of sand on a beach with a single spoon; they were too slow and too rigid for the job.

Enter DGrowthR. Think of this as a super-smart, flexible robot assistant for biologists. Here is how it works, broken down into simple concepts:

1. The "Shape-Shifter" (Gaussian Process Regression)

Old tools tried to force every growth curve into a pre-made mold (like trying to fit a square peg in a round hole). If the bacteria didn't fit the mold, the tool broke.

DGrowthR is different. Imagine it's a clay sculptor. Instead of forcing the data into a box, it gently molds itself around the actual shape of the bacteria's growth, no matter how weird or wiggly it is. It doesn't guess what the curve should look like; it just learns what it actually looks like. This allows it to spot subtle changes that other tools miss, like a bacteria that grows for a bit, stops, and then starts again.

2. The "Crowd Control" (Clustering and Visualization)

When you have 20,000 or even 100,000 growth curves, it's like looking at a stadium full of people all moving at once. You can't see individual patterns.

DGrowthR acts like a smart crowd manager. It uses a technique called "dimensionality reduction" (think of it as taking a 3D movie and projecting it onto a 2D screen without losing the story). It groups similar bacteria movements together.

• Group A: Bacteria that are thriving.
• Group B: Bacteria that are dying off.
• Group C: Bacteria that are acting weird (maybe they paused and then restarted).

This helps scientists instantly see, "Oh, look! All the bacteria treated with this specific drug are in the 'weird' group," which is a huge clue.

3. The "Judge" (Statistical Testing)

Once the robot has grouped the bacteria, it needs to decide: "Is this difference real, or just a fluke?"

Imagine you are a judge in a courtroom. You have a defendant (the bacteria treated with a drug) and a control group (bacteria with just water). The judge needs to know if the drug actually changed the outcome.

• Old way: The judge might just look at one number, like "How tall did they get?"
• DGrowthR way: The judge looks at the entire movie of their growth. It asks, "If we shuffled the labels (pretending the drug was actually water), how often would we see this result by pure luck?"

It runs this "shuffling" test thousands of times in seconds. If the result is still rare even after shuffling, the judge declares: "Guilty! This drug definitely changed the bacteria's behavior."

What Did They Discover?

The authors used this tool to solve three real-world mysteries:

1. The Drug Screen: They tested thousands of chemicals on two types of dangerous bacteria. They found that some drugs didn't just kill the bacteria; some made them grow faster or act strangely. DGrowthR spotted these "non-canonical" behaviors that other tools would have ignored.
2. The Genetic Mystery: They looked at a specific bacteria (Vibrio cholerae) that had a part of its DNA removed. They found that without this DNA, the bacteria became resistant to certain antibiotics. DGrowthR helped pinpoint exactly how the bacteria's growth slowed down differently compared to the normal version.
3. The Cocktail Effect: They tested mixing two drugs together. They confirmed that mixing Vanillin (the flavor in vanilla) with an antibiotic made the antibiotic work better. But mixing Caffeine with another antibiotic made it work worse. DGrowthR proved these interactions were real and significant.

The Bottom Line

DGrowthR is a bridge between raw data and biological discovery. It takes the messy, overwhelming flood of data from modern robotic labs and turns it into clear, actionable stories.

• Before: Scientists had to guess, force data into boxes, and often miss the weird but important stuff.
• Now: With DGrowthR, they have a flexible, automated, and statistically rigorous "detective" that can find the hidden patterns in the chaos, helping us find new antibiotics and understand how bacteria fight back.

It's essentially giving biologists a pair of super-vision glasses to see the true story of how life grows and reacts.

Technical Summary

1. Problem Statement

The quantitative analysis of microbial growth curves is critical for understanding bacterial responses to environmental and chemical cues. However, current methodologies face significant limitations when applied to modern, high-throughput data generated by automated robotics platforms:

• Parametric Assumptions: Traditional tools (e.g., Sicegar, Growthcurver) rely on parametric models (Logistic, Gompertz) that assume sigmoidal or double-sigmoidal curve shapes. These fail when growth dynamics are distorted by chemical perturbations or genetic factors, leading to inaccurate parameter estimation.
• Lack of Flexibility: Existing flexible tools (e.g., grofit, QurvE) often rely on splines, making them sensitive to outliers and technical noise.
• Insufficient Statistical Testing: Most tools lack rigorous statistical frameworks to compare entire growth dynamics across conditions. Many rely on summary statistics (like Area Under the Curve, AUC) which lose temporal information, or lack principled hypothesis testing for differential growth (DG).
• Scalability and Usability: There is a lack of integrated, user-friendly software (specifically in R) that handles end-to-end analysis of massive datasets (tens to hundreds of thousands of curves) while offering exploratory data analysis (EDA), non-parametric modeling, and statistical inference.

2. Methodology: The DGrowthR Framework

The authors introduce DGrowthR, an R package and standalone desktop application designed for the integrative, end-to-end analysis of large-scale microbial growth data. The framework consists of four main modules:

A. Data Pre-processing and Standardization

• Input: Handles raw Optical Density (OD) data from automated plate readers.
• Cleaning: Removes initial noisy time points, applies log-transformation to OD measurements, and performs baseline correction (setting initial measurements to zero) to account for instrument fluctuations and varying starting populations.

B. Exploratory Functional Data Analysis (EDA)

• Dimensionality Reduction: Utilizes Functional Principal Component Analysis (FPCA) and Uniform Manifold Approximation and Projection (UMAP) to embed high-dimensional growth curves into low-dimensional spaces.
• Clustering: Applies density-based clustering (DBSCAN) on UMAP embeddings to identify distinct growth patterns (clusters) and detect outliers without prior knowledge of the curve shapes.

C. Non-Parametric Modeling via Gaussian Process (GP) Regression

• Modeling: Uses Gaussian Process (GP) regression (specifically the laGP package) to model growth curves. This non-parametric Bayesian approach does not assume a fixed functional form, allowing it to capture complex, non-canonical dynamics.
• Replication: Can pool replicate curves for joint modeling, reducing the influence of outliers.
• Parameter Extraction: Derives key biological parameters from the GP model and its derivatives:
  • First Derivative: Maximum growth rate ($\alpha_{max}$), maximum decay rate ($\alpha_{decay}$), and doubling time.
  • Second Derivative: Identification of lag phase end and stationary phase start.
  • Direct Metrics: Maximum OD, AUC, and growth loss.

D. Differential Growth (DG) Statistical Testing

• Hypothesis Testing: Compares growth curves between a treatment condition and a control using a Bayes Factor approach.
  • Null Hypothesis ($H_0$): All curves are modeled as a function of time only.
  • Alternative Hypothesis ($H_a$): Curves are modeled as a function of time and treatment condition.
• Permutation Testing: To determine significance, treatment labels are permuted to generate a null distribution of Bayes Factors.
• Gamma Approximation: To handle the computational cost of large-scale permutation testing (where $p$-values can be extremely small), the distribution of permuted Bayes Factors is approximated using a Gamma distribution. This allows for the accurate estimation of small $p$-values with fewer permutations, significantly speeding up computation.
• Correction: Adjusts for multiple comparisons using the Benjamini-Hochberg procedure to control the False Discovery Rate (FDR).

3. Key Contributions

1. End-to-End R Framework: Provides the first comprehensive, no-code (via standalone app) and code-based R solution for the full pipeline of large-scale growth curve analysis, from raw data import to statistical inference.
2. Non-Parametric Flexibility: Replaces rigid parametric models with GP regression, enabling accurate modeling of distorted, non-sigmoidal, and complex growth dynamics that traditional tools miss.
3. Rigorous Statistical Inference: Introduces a computationally efficient, permutation-based DG testing framework that compares entire growth trajectories rather than isolated parameters, providing robust evidence for differential growth.
4. Scalability: Demonstrates the ability to process datasets containing over 100,000 growth curves efficiently through parallelization and Gamma approximation.

4. Results and Validation

The authors validated DGrowthR using three distinct large-scale datasets:

• Case 1: In-house Chemical Screen (20,000+ curves)

  • Data: Salmonella enterica and Campylobacter jejuni exposed to 2,415 compounds.
  • Findings: FPCA/UMAP revealed diverse growth dynamics beyond simple "growth/no-growth" binary classifications. DG analysis identified compounds with significant inhibitory effects (e.g., Auranofin reducing S. enterica growth) and compounds inducing non-canonical dynamics (e.g., Pimavanserin increasing C. jejuni growth).
  • Comparison: DGrowthR successfully modeled curves that parametric tools (Growthcurver) failed to fit or fitted inaccurately.

• Case 2: Genetic Perturbation (Brenzinger et al.)

  • Data: Vibrio cholerae wild-type vs. $\Delta$CBASS (anti-phage system deleted) strains screened against 94 compounds.
  • Findings: Recapitulated known findings that the $\Delta$CBASS strain is less susceptible to antifolate antibiotics (Sulfamethoxazole, Trimethoprim).
  • Novel Insight: Identified a decreased decay rate in the $\Delta$CBASS strain specifically in response to cell-wall synthesis inhibitors (Amoxicillin, Meropenem, Penicillin G), suggesting a new mechanistic link between CBASS and cell wall integrity.

• Case 3: Drug Combination Screen (Brochado et al.)

  • Data: E. coli BW treated with ~3,000 pairwise drug combinations.
  • Findings: Recovered known synergistic (Vanillin + Spectinomycin) and antagonistic (Caffeine + Amoxicillin) interactions.
  • Novel Insight: Identified a novel antagonistic interaction between Vanillin and D-Cycloserine (contrary to Vanillin's typical antagonistic behavior) and highlighted the broad antagonistic role of Vanillin across many antibiotic classes.

5. Significance

• Paradigm Shift: DGrowthR shifts microbial growth analysis from parametric, summary-statistic-based approaches to a functional data analysis paradigm, treating growth curves as continuous functions rather than discrete points.
• Reproducibility: By providing an open-source R package with a standardized workflow, it addresses the reproducibility crisis in high-throughput screening, allowing researchers to compare results across different labs and conditions.
• Biological Discovery: The framework enables the detection of subtle, complex growth phenotypes (e.g., altered decay rates, non-canonical dynamics) that are often missed by traditional methods, leading to new hypotheses regarding drug mechanisms and genetic interactions.
• Accessibility: The inclusion of a standalone desktop app with a no-code interface lowers the barrier to entry for microbiologists who may not be proficient in programming, while the R package offers flexibility for bioinformaticians.

In conclusion, DGrowthR provides a robust, statistically rigorous, and scalable framework that is essential for the next generation of high-throughput microbial phenotyping and drug discovery.