Single-cell hit calling in high-content imaging screens with Buscar

The paper introduces Buscar, an open-source Python method for high-content screening that improves hit calling by leveraging full single-cell heterogeneity to define morphology signatures and quantify both compound efficacy and specificity, thereby overcoming the limitations of traditional aggregate statistics.

The Gist

Imagine you are a detective trying to solve a mystery in a massive city. The city represents a petri dish full of cells, and the "mystery" is figuring out which chemical drug or genetic tweak can fix a sick cell and turn it back into a healthy one.

For a long time, scientists had a blind spot. When they looked at the city, they didn't look at individual people; they just took a census average. They'd say, "The average citizen is 5 feet 6 inches tall." But what if half the city is made of giants and the other half of dwarfs? The average hides the truth. In biology, this is called "aggregation." It assumes every cell reacts the same way to a drug, which is rarely true. Some cells might get cured, some might get worse, and some might not care at all.

Enter Buscar (which means "to look for" in Spanish). It's a new tool that changes the game by letting scientists look at every single cell individually, rather than just the average.

Here is how Buscar works, using a simple analogy:

The Setup: The "Before" and "After" Photos

Imagine you have two groups of people:

1. The Sick Group (Reference): People with a specific illness (e.g., a heart condition).
2. The Healthy Group (Target): People who are perfectly healthy.

Scientists take photos of both groups and measure everything: height, shoe size, how fast they walk, how loud they talk, etc. These are the "morphology features."

Step 1: The "On" and "Off" Lists

Buscar looks at the photos and creates two lists:

• The "On" List (The Symptoms): These are the traits that are different between the sick and healthy groups. Maybe sick people walk slowly and have red shoes, while healthy people walk fast and wear blue shoes. If a drug changes a sick person to walk fast and wear blue, that's good! This is Efficacy.
• The "Off" List (The Normal Stuff): These are traits that are the same for both groups. Maybe both groups have the same hair color and eye shape. If a drug changes a sick person's hair color to neon green, that's weird! That wasn't part of the cure; it's a side effect. This is Specificity (or lack of off-target effects).

Step 2: The Drug Test

Now, you give a new drug to the sick people. Buscar takes a photo of them after the drug and compares them to the lists.

• The "On-Buscar" Score (The Cure Score): How close do the treated people look to the "Healthy Group" regarding the "On" traits?
  • Low Score: Great! They look just like the healthy people. The drug worked.
  • High Score: Bad. They still look sick.
• The "Off-Buscar" Score (The Side Effect Score): Did the drug mess up the "Off" traits?
  • Low Score: Great. They kept their normal hair and eye color. The drug was precise.
  • High Score: Bad. They turned neon green. The drug is too messy and might be dangerous.

Why This Matters: The Three Big Wins

The paper tested this tool on three different "cities" (datasets) to prove it works:

1. The Heart City (Cardiac Fibroblasts):
   Scientists had heart cells from patients with heart failure. They tried a drug meant to fix the heart.

   • Old Way: They would have just said, "The average cell looks a little better."
   • Buscar Way: It saw that the drug actually fixed the "sick" cells (low On-Buscar score) but also noticed it made some healthy cells act weirdly (high Off-Buscar score). This tells scientists: "This drug works, but be careful, it has side effects."

2. The Gene City (MitoCheck):
   They tested thousands of genes to see which ones, when turned off, caused specific cell shapes (like a cell splitting into two heads).

   • The Result: Buscar was like a smart librarian. It correctly ranked the genes. If turning off a gene made cells look like the "two-headed" target, that gene got a top rank. If it didn't, it got a low rank. It proved the tool understands biology, not just math.

3. The Reproducibility City (CPJUMP1):
   Sometimes, if you run an experiment on Monday vs. Tuesday, or in Plate A vs. Plate B, the results change because of tiny technical differences (like the temperature of the room).

   • The Result: Buscar was rock solid. Even when the same drug was tested on different plates, it gave the same score. This means scientists can trust it to find real hits, not just lucky accidents.

The Bottom Line

Buscar is like upgrading from a blurry, black-and-white group photo to a high-definition, 4K video where you can see every person's face.

• Old Method: "The crowd looks happy." (But maybe only half are, and the other half are crying).
• Buscar: "Group A is cheering, Group B is crying, and Group C is confused. Let's find the drug that makes Group B cheer without making Group C cry."

By separating the "cure" from the "side effects" and looking at every single cell, Buscar helps scientists find better drugs faster and avoid the ones that might fail later in human trials. It's a smarter, more honest way to look for the next big medical breakthrough.

Technical Summary

1. Problem Statement

High-content screening (HCS) generates vast amounts of single-cell image-based data to quantify morphological changes induced by chemical or genetic perturbations. However, the current standard for "hit calling" (identifying effective compounds) relies on aggregating single-cell data into summary statistics (e.g., mean or median values per well).

This aggregation approach has critical limitations:

• Loss of Heterogeneity: It assumes uniform cellular responses, obscuring subpopulations (e.g., resistant vs. sensitive cells) and differential toxicity.
• Reduced Sensitivity: Subtle or heterogeneous effects are diluted, leading to high false-negative rates and missed mechanistic insights.
• Inability to Distinguish Efficacy vs. Specificity: Aggregated methods struggle to separate a compound's desired phenotypic rescue (efficacy) from unintended, broad morphological disruptions (off-target effects/specificity).

2. Methodology: The Buscar Framework

The authors introduce Buscar (Bioactive Unbiased Single-cell Compound Assessment and Ranking), a Python package that operates directly on distributions of single-cell profiles rather than aggregated summaries. The method consists of two sequential modules:

Module 1: Defining Morphology Signatures

Buscar requires two reference populations representing distinct morphological states:

1. Reference State: Typically a disease or negative control state (e.g., failing cardiac fibroblasts).
2. Target State: Typically a healthy or desired state (e.g., healthy cardiac fibroblasts).

Using non-parametric statistical tests (Kolmogorov-Smirnov test with Benjamini-Hochberg FDR correction), Buscar compares feature distributions between these two states to partition the feature space into two signatures:

• On-Morphology Signature: Features that differ significantly between the reference and target states. These represent the morphological changes required to achieve the desired phenotype.
• Off-Morphology Signature: Features that do not differ significantly. These represent stable cellular characteristics that should remain unchanged; alterations here indicate off-target effects.

Module 2: Perturbation Scoring

For every perturbation in a screen, Buscar calculates two complementary metrics:

• On-Buscar Score (Efficacy): Measures how closely the perturbed cell distribution aligns with the target state along the on-morphology signature.
  • Metric: Earth Mover's Distance (EMD/Wasserstein distance).
  • Interpretation: A score of 0.0 indicates a perfect shift to the target state (high efficacy); 1.0 indicates no shift from the reference state.
• Off-Buscar Score (Specificity): Measures the extent to which the perturbation alters features in the off-morphology signature.
  • Metric: The proportion of off-morphology features that become statistically significant (via KS test) when comparing the perturbed state to the target state.
  • Interpretation: A low score indicates high specificity (few off-target effects); a high score suggests broad, unintended morphological disruption.

3. Key Contributions

• Paradigm Shift: Moves hit calling from aggregate statistics to single-cell distribution analysis, preserving biological heterogeneity.
• Dual-Metric Scoring: Simultaneously quantifies efficacy (phenotypic rescue) and specificity (off-target effects), allowing researchers to navigate the trade-off between potency and safety.
• Open-Source Implementation: Released as a Python package compatible with the Cytomining ecosystem (Pycytominer, CytoTable), utilizing Polars for scalable data processing.
• Interpretability: Unlike "black box" deep learning embeddings, Buscar's scores are derived from explicit statistical comparisons of defined morphological features.

4. Results & Validation

The authors validated Buscar across three distinct datasets:

A. Cardiac Fibroblasts (Cell Painting Dataset)

• Context: Testing TGFβ receptor inhibitors (TGFβRi) on failing vs. healthy cardiac fibroblasts.
• Findings:
  • Buscar successfully detected phenotypic rescue in failing cells treated with TGFβRi (low on-Buscar score).
  • It identified subpopulations within the healthy control group that were obscured by aggregation.
  • It revealed context-specific off-target effects: The off-Buscar score was higher for failing cells treated with TGFβRi compared to healthy cells, suggesting secondary compensatory responses in the disease state that aggregation would have missed.

B. MitoCheck (Genome-wide RNAi Screen)

• Context: 16 manually annotated nuclear phenotypes (e.g., mitosis, apoptosis) in HeLa cells.
• Findings:
  • Using a "Leave-One-Gene-Out" (LOGO) approach, Buscar correctly ranked genes based on their ability to induce specific phenotypes.
  • Genes producing a higher proportion of cells with a specific phenotype received significantly lower on-Buscar scores (higher efficacy).
  • Cross-phenotype analysis: Hierarchical clustering of gene rankings revealed biologically coherent groups (e.g., mitotic phases clustered together), confirming that the method captures genuine morphological covariance rather than label artifacts.

C. CPJUMP1 (Small Molecules & CRISPR)

• Context: Reproducibility testing across multiple plates, cell lines (U2OS, A549), and modalities (compounds, CRISPR).
• Findings:
  • Buscar demonstrated robust cross-plate reproducibility. Paired replicates (same treatment on different plates) showed significantly tighter score distributions and lower variance compared to non-paired (randomized) controls.
  • This consistency held for both small molecules and CRISPR knockdowns, proving the method is reliable for large-scale screening campaigns.

5. Significance

• Improved Hit Identification: By accounting for cell-to-cell heterogeneity, Buscar reduces false positives and false negatives, potentially lowering the high attrition rates of compounds in clinical development.
• Mechanistic Insight: The separation of efficacy and specificity allows researchers to identify compounds that not only rescue a phenotype but do so without causing broad cellular toxicity.
• Scalability: The method is designed to handle high-dimensional single-cell data efficiently and integrates with existing bioinformatics pipelines.
• Future Directions: The authors note limitations, such as the quadratic computational complexity of EMD (suggesting subsampling or approximation methods) and the current use of univariate feature testing (suggesting future multivariate approaches like Maximum Mean Discrepancy).

In conclusion, Buscar provides a reproducible, interpretable, and statistically rigorous framework for single-cell hit calling, addressing the critical gap between high-content imaging data generation and biologically meaningful drug discovery.