MORPHIS (MORPHological Interpretable Signature) captures heterogeneous treatment- and aging-related responses of single cells

The paper introduces MORPHIS, a machine learning framework that utilizes explainable, analytically rich features to robustly quantify and interpret heterogeneous single-cell morphological responses to diverse perturbations, including pharmacological treatments and aging, across various biological contexts.

The Gist

Imagine you are a detective trying to figure out what happened to a city by looking at the shape of its buildings. Usually, when scientists study cells (the tiny building blocks of life), they take a "census" of the whole neighborhood. They might say, "On average, the buildings look a bit squished." But this misses the story: maybe 90% of the buildings are fine, but 10% are completely collapsed, or maybe every building has a tiny crack in a different spot.

This paper introduces a new detective tool called MORPHIS (MORPHological Interpretable Signature). Here is how it works, explained simply:

1. The Problem: The "Blurry Photo" vs. The "High-Res Portrait"

Traditionally, scientists look at cells in two ways:

• The "Gut Feeling" Method: They look at a few pictures and say, "Hmm, these cells look sick." This is subjective and misses details.
• The "Black Box" AI Method: They use powerful computer programs (Deep Learning) that can spot sick cells with amazing accuracy. But these programs are like a magic 8-ball: they give you the answer ("Sick!"), but they won't tell you why or how they decided. It's like a judge giving a verdict without explaining the evidence.

2. The Solution: MORPHIS (The "Smart Detective")

MORPHIS is a new computer program that acts like a super-smart detective who not only solves the case but also writes a clear report explaining exactly what clues led to the conclusion.

• It takes a "Morphological Fingerprint": Instead of just looking at the whole cell, MORPHIS breaks the cell down into 41 specific, easy-to-understand measurements. Think of it like measuring a person's height, shoe size, how fast they walk, and the texture of their skin.
• It uses "Explainable AI": When MORPHIS says, "This cell is reacting to Drug A," it points to the specific clues: "Because the nucleus (the cell's brain) is wider at the edges, and the skeleton (actin) looks messy." It doesn't hide the logic.

3. What Did They Discover?

The researchers tested MORPHIS on cells in a petri dish and even on tiny worms (C. elegans) to see how it handles different "attacks" on the cells.

• The "Drug Test": They hit cells with eight different substances (some meant to help drugs get inside, some meant to kill cancer).

  • Result: MORPHIS could tell the difference between them perfectly.
  • The "Why": It showed that a drug meant to kill cancer (Doxorubicin) made the cell's nucleus look like a stretched-out balloon with a messy skeleton. A drug meant to open cell doors (a permeation enhancer) just made the nucleus look slightly flatter.
  • The Analogy: If the cell is a house, Doxorubicin is like a tornado that blows the roof off and twists the walls. The permeation enhancer is just someone gently opening a window. MORPHIS can tell the difference between a tornado and an open window.

• The "Fractional Response" (The Hidden Story):

  • Usually, scientists say, "The drug worked on 50% of the cells."
  • MORPHIS goes deeper. It realized that cells aren't just "On" or "Off." Some cells are very affected, some are mildly affected, and some are unaffected.
  • The Analogy: Imagine a classroom where a teacher asks a question. The old method says, "50% of the class raised their hands." MORPHIS says, "Actually, 10 students raised their hands high, 20 raised them halfway, and 10 didn't raise them at all." This helps scientists understand that not every cell reacts the same way, which is crucial for understanding why some people get sick and others don't.

• The "Aging" Test:

  • They also used MORPHIS on old worms to see how they age.
  • Result: It found that as worms get old, their cell nuclei get wrinkly and uneven, like a grape turning into a raisin. It quantified this "wrinkling" perfectly, proving the tool works on living animals, not just lab dishes.

4. Why Does This Matter?

MORPHIS is a bridge between two worlds:

1. Accuracy: It is as good at spotting changes as the most complex, "black box" AI systems.
2. Understanding: It gives humans a clear, readable explanation of what changed and why.

In a nutshell:
If cell biology is a mystery novel, old methods gave you the solution but hid the clues. Deep learning gave you the solution but spoke in a language you couldn't understand. MORPHIS gives you the solution and lays out all the evidence on the table in plain English, helping scientists understand exactly how drugs, aging, and diseases change the shape and soul of a cell.

Technical Summary

1. Problem Statement

Cell morphology is a critical indicator of physiological states, including disease, drug response, and aging. However, current methods for assessing morphological changes face significant limitations:

• Qualitative Bias: Traditional assessments often rely on subjective visual inspection of representative images.
• Loss of Heterogeneity: Ensemble-level assays (e.g., bulk viability tests) mask cell-to-cell variability, failing to capture "fractional responses" where only subsets of a genetically identical population respond to a perturbation.
• Interpretability Gap: While deep learning (DL) models (e.g., CNNs) achieve high classification accuracy, they function as "black boxes," extracting latent features that lack biological interpretability. Conversely, classical feature-based methods (e.g., CellProfiler) often rely on generic geometric descriptors that fail to capture complex phenotypic signatures or link them to specific mechanisms.

There is a need for a framework that balances predictive accuracy with biological interpretability while operating at the single-cell resolution to quantify heterogeneous responses.

2. Methodology: The MORPHIS Framework

The authors introduce MORPHIS (MORPHological Interpretable Signature), a machine learning pipeline designed to extract compact, interpretable feature signatures from fluorescence microscopy images.

Key Components:

• Input Data: Dual-channel fluorescence images (nucleus and cytosol/actin) paired with segmentation masks.
• Feature Extraction: Instead of using hundreds of generic features or raw pixel data, MORPHIS extracts a curated set of 41 biologically interpretable features. These cover:
  • Shape and size (e.g., area, circularity).
  • Intensity distributions (e.g., mean, median, standard deviation).
  • Texture and spatial organization (e.g., Shannon entropy, homogeneity).
  • Compartment-specific metrics (e.g., nuclear edge-to-center intensity ratio, nuclear z-stack height).
• Classification Engine: The framework employs XGBoost (a gradient-boosted decision tree algorithm) trained via 5-fold stratified cross-validation. XGBoost was selected for its robustness and ability to handle non-linear relationships in tabular data.
• Interpretability (SHAP): To decode the "black box" of the classifier, MORPHIS utilizes SHAP (SHapley Additive exPlanations) values. This ranks features by importance, revealing which specific morphological attributes drive the classification of a specific treatment or condition.
• Heterogeneity Quantification:
  • UMAP Embedding: The feature space is projected using Uniform Manifold Approximation and Projection (UMAP) to visualize continuous morphological transitions rather than discrete clusters.
  • Probability-Based Stratification: Instead of binary classification, the model outputs a prediction probability ($p$). Cells are stratified into low ($p < 0.35$), mid ($0.35 \le p < 0.7$), and high ($p \ge 0.7$) responders, allowing the quantification of "fractional responses" within a population.

3. Key Contributions

1. Interpretable ML Framework: MORPHIS bridges the gap between classical feature engineering and deep learning by using a small, curated set of analytically rich features that yield high accuracy while remaining biologically traceable.
2. Quantification of Heterogeneity: It provides a novel metric for "fractional response," enabling researchers to determine not just if a treatment works, but what fraction of the population responds and how strongly.
3. Generalizability: The framework is demonstrated to be cell-type and perturbation-agnostic, successfully applied to:
   • In Vitro: Caco-2 (confluent monolayer) and HeLa (non-confluent) cells treated with 8 distinct compounds (permeation enhancers, surfactants, anticancer drugs).
   • In Vivo: Aging nuclei in C. elegans.
4. Benchmarking: The method was rigorously benchmarked against state-of-the-art tools (Vampire, CAJAL) and deep learning models (scDINO), showing comparable or superior accuracy with significantly higher interpretability.

4. Key Results

A. Benchmarking and Validation

• Binary Classification: In distinguishing control vs. Doxorubicin (DOX)-treated Caco-2 cells, MORPHIS (XGBoost) achieved 80.02% accuracy, outperforming or matching other classifiers (SVM, Random Forest, Logistic Regression).
• Comparison with SOTA:
  • Vampire: 57.20% (contour features) to 73.26% (morphological features).
  • CAJAL: 71.68%.
  • scDINO (Deep Learning): 78.65%.
  • Conclusion: MORPHIS achieves performance on par with deep learning (scDINO) but retains full feature interpretability, whereas scDINO's attention maps were globally distributed and difficult to interpret biologically.

B. Treatment-Specific Signatures (In Vitro)
MORPHIS successfully distinguished between 8 mechanistically distinct compounds in Caco-2 and HeLa cells:

• Doxorubicin (DOX): Induced a senescence signature characterized by nuclear enlargement, peripheral nuclear enrichment, and increased cytoskeletal heterogeneity.
• Triton X-100: Caused massive cytotoxicity/apoptosis, identified by nuclear signal centralization and high heterogeneity (93.8% of cells showed a strong response).
• L-penetramax (px): A permeation enhancer causing subtle nuclear flattening with low response heterogeneity (only 38.2% strong responders).
• Mechanistic Clustering: Multiclass analysis revealed that compounds with similar mechanisms (e.g., lipidic permeation enhancers C10 and SNAC) clustered together in the feature space, while those with different mechanisms (e.g., surfactants vs. intracellular DNA intercalators) showed distinct, anticorrelated feature patterns.

C. Cell-Type Specificity

• Caco-2 vs. HeLa: While some trends were conserved (e.g., C10/SNAC similarity), the dominant morphological descriptors differed. Caco-2 responses were driven largely by nuclear intensity features, whereas HeLa responses were more sensitive to actin and global shape metrics. This highlights MORPHIS's ability to detect cell-type-specific phenotypic shifts.

D. In Vivo Aging Application

• Applied to C. elegans nuclei (LMN-1::GFP reporter) comparing Day 1 (young) vs. Day 12 (aged).
• Achieved 81.57% accuracy.
• Identified age-associated features: lower nuclear edge-to-center intensity ratio, reduced solidity (envelope folding/lobulation), and higher intensity standard deviation (uneven lamina distribution).
• Captured the stochastic nature of aging, showing a continuum of nuclear deterioration rather than a binary young/old state.

5. Significance

• Mechanistic Insight: MORPHIS moves beyond simple classification to provide mechanistic insights. By linking specific features (e.g., "nuclear edge-to-center ratio") to specific treatments, it helps hypothesize the underlying biological pathways (e.g., DNA damage vs. membrane permeabilization).
• Drug Discovery & Toxicology: The ability to quantify fractional responses is crucial for drug development. It allows researchers to identify sub-populations that are resistant or uniquely sensitive to a drug, which is often missed by bulk assays.
• Aging Research: The successful application to C. elegans demonstrates the framework's utility in studying intrinsic biological processes like aging in vivo, offering a quantitative phenotyping tool for nuclear aging.
• Balancing Act: MORPHIS offers a "middle ground" in computational biology, avoiding the oversimplification of classical metrics and the opacity of deep learning, making it a versatile tool for diverse biological contexts.