BioTrendFinder - an interactive web tool for exploring functional drivers in gene- and protein-level bulk omics data

BioTrendFinder is an interactive web tool that integrates sample-ranking strategies with statistical, metadata, and functional network data to enable the efficient identification and prioritization of functional drivers and candidate targets in bulk gene- and protein-level omics datasets.

The Gist

Imagine you have a massive library of books (your biological data, like RNA or protein levels) representing thousands of different cells. Traditionally, scientists would sort these books into two piles: "Sick" and "Healthy," then look for the few books that are different between the two piles. It's a bit like looking for a needle in a haystack by only checking the top of the haystack.

BioTrendFinder is a new, interactive digital tool that changes the game. Instead of just sorting books into two static piles, it lets you arrange the entire library along a sliding scale (a spectrum) and watch how the books change as you move from one end to the other.

Here is a simple breakdown of how it works, using everyday analogies:

1. The Problem: The "Snapshot" vs. The "Movie"

Standard analysis takes a snapshot. It asks, "What is different between Group A and Group B?"
BioTrendFinder takes a movie. It asks, "If we slowly move from a 'very sick' state to a 'very healthy' state, how do the molecules change along the way?"

2. The Core Concept: The "Trendline"

Imagine you have a long line of people (your samples) arranged from shortest to tallest.

• The Old Way: You just count how many people are wearing red shirts in the "short" group vs. the "tall" group.
• BioTrendFinder's Way: You watch the line of people walk by. You look for specific people (molecules) who are wearing red shirts that get brighter and brighter as the people get taller. Or, you look for people whose shirts get darker and darker.

These changing patterns are called "Trendlines." BioTrendFinder finds the molecules that have the strongest, most consistent "story" as you move along your line of samples.

3. The Workflow: How the Tool Works

The tool guides you through a few simple steps, like a GPS for your data:

• Step 1: The Map (Upload & Rank): You upload your data. The tool creates a map (like a scatter plot) of your samples. You can draw a line through this map to decide the order of your samples. Maybe you want to rank them from "Low Stress" to "High Stress," or "Healthy" to "Obese."
• Step 2: The Detective Work (Analyze): The tool draws a line for every single molecule, showing how its level changes as you move along your ranking. It then groups them:
  • Set 1: Molecules that go UP steadily (like a rising sun).
  • Set 2: Molecules that go DOWN steadily (like a setting sun).
• Step 3: The Filter (Statistics): Not every rising or falling line is important. Some are just random noise. The tool acts like a sieve, filtering out the weak signals and keeping only the "statistically significant" trends.
• Step 4: The Social Network (Functional PPI): This is the coolest part. The tool takes the important molecules it found and asks, "Who do these molecules know?" It connects them to a giant social network (called STRING) where molecules that work together are friends.
• Step 5: The Team Huddle (Functional Module): Finally, it groups these "friends" into teams based on what they do (e.g., "The Inflammation Team" or "The Fat-Burning Team"). It then ranks the members of the team to tell you: "Who is the Captain?" (The most important driver) and "Who is the Lieutenant?"

4. Why This Matters: Finding the "Needle"

In the paper, the authors used this tool to study fat tissue (adipose tissue) in two different scenarios:

1. Secreted Proteins: They looked at proteins released by fat cells. They found that as cells were stressed (treated with norepinephrine), some proteins dropped out of the mix while others appeared. This revealed a hidden "stress response" pattern that standard methods missed.
2. Obesity: They looked at humans ranging from "Metabolically Unhealthy Obese" to "Metabolically Healthy Lean." They found specific genes that act like a switch, turning on as you get healthier.
   • The Result: They identified specific molecules (like SPX and AZGP1) that could be potential drug targets to help treat obesity.

The Big Takeaway

Think of BioTrendFinder as a smart spotlight.

• Standard tools shine a light on the whole room and say, "There's a mess here."
• BioTrendFinder shines a spotlight on the movement of the mess. It tells you exactly which items are being thrown out, which are being picked up, and which specific item is the "mastermind" behind the change.

It helps scientists move from asking "What is different?" to asking "What is driving the change?" This makes it much easier to find the best candidates for new medicines or therapies without getting lost in the noise.

Technical Summary

1. Problem Statement

Standard analytical workflows for bulk omics data (e.g., RNA-seq, proteomics) typically follow a linear, disjointed sequence: dimensionality reduction (PCA/UMAP), group-wise statistical comparisons (differential expression), functional enrichment analysis, and network mapping.

• Limitations: These steps are often performed independently, limiting the ability to integrate high-level patterns with specific molecular drivers.
• Gap: Conventional methods struggle to efficiently identify functional drivers and candidate targets that emerge from continuous expression trends across sample compositions rather than just discrete group contrasts. There is a need for a tool that integrates dimensionality reduction, sample ranking, statistical filtering, and functional network analysis into a cohesive, interactive workflow.

2. Methodology

BioTrendFinder is an interactive web-based tool (built with R/Shiny and Python) designed to explore functional drivers by analyzing molecular "trendlines" across ranked samples.

Core Workflow

The tool operates through a sequential pipeline of tabs:

1. Data Upload & Preparation:
   • Supports three tracks: (1) Main data + pre-computed dimensionality reduction, (2) Main data only (with internal PCA/UMAP), or (3) Pre-ranked ID & score data.
   • Automatically filters data (removes Ensembl version numbers, maps to STRING IDs, handles NAs/duplicates).
2. Sample Ranking:
   • Plot-based: Users click an axis on a dimensionality-reduced plot (PCA/UMAP) to generate a sample ranking.
   • Manual/Alternative: Users can manually rank samples, rank based on specific molecule expression, or rank based on group means/medians.
3. Trendline Generation & Analysis:
   • For every molecule, a trendline is generated tracing expression across the ranked samples.
   • Metrics: Trendlines are annotated with Spearman correlation, slope (modified linear model fit), mean expression, and a Combined_score (product of selected metrics).
   • Set Definition: Users define "Set 1" (increasing trends) and "Set 2" (decreasing trends) using slider thresholds for correlation, slope, and combined scores.
4. Statistical Filtering:
   • Segment-based: Compares left vs. right end-segments of trendlines using Welch's t-test.
   • Metrics: Computes p-values, FDR, log2 fold change (log2FC), and a pi-score ($-\log_{10}(p) \times \log2FC$).
   • Outlier Removal: Iteratively removes extreme values to ensure trendline segments do not overlap.
   • Group Division: Calculates a "group division score" to quantify how well sample groups cluster at the low or high ends of the trend.
5. Functional Integration (PPI & Enrichment):
   • PPI Network: Molecules are mapped to the STRING v12 database.
   • Enrichment: Performs enrichment analysis across 11 bio-ontologies (GO, KEGG, Reactome, UniProt Keywords, DISEASES, etc.).
   • Graph Enrichment: Significant ontology terms are used to add edges between nodes in the PPI network, creating a functional graph.
6. Functional Module Construction:
   • Generates a ranked list of molecules (a "Functional Module") based on a composite FM_score.
   • FM_score components: Degree centrality, edge weight fraction, node importance, trendline statistics (Spearman, pi-score), and FDR.

Key Algorithms

• Ranking: Uses a half-circle construction around 2D coordinates to derive ranking axes.
• Confidence Score: Quantifies sample group clustering along a ranking axis using a Dirichlet model-based null distribution.
• Community Detection: Uses the Louvain algorithm (via NetworkX) to identify functional communities.
• Graph Construction: Builds hierarchical ontology graphs using Dijkstra's algorithm to connect significant terms via shortest paths.

3. Key Contributions

• Novel Analytical Paradigm: Shifts focus from static group contrasts to continuous sample-ranking strategies, allowing users to interrogate data along specific biological gradients (e.g., disease severity, treatment response).
• Integration of Multi-Omics Layers: Uniquely combines expression dynamics (trendlines), statistical significance, sample metadata, and functional connectivity (STRING + 11 ontologies) into a single interactive framework.
• Entity-Ranked Functional Modules: Instead of just listing enriched terms, BioTrendFinder outputs a ranked list of specific molecules (genes/proteins) that are central to the identified functional modules, prioritizing candidates for validation.
• Flexibility: Supports both hypothesis-driven (manual ranking) and exploratory (plot-based) analyses.

4. Results & Validation

The authors validated BioTrendFinder using two previously published datasets:

Case Study 1: Proteomics Secretome (Adipose Tissue)

• Context: Differentiated adipocytes from Brown (BAT) and White (WAT) adipose tissue, with/without norepinephrine (NE) treatment.
• Finding: Identified a transition from a classical secretory profile to a stress/remodeling signature.
  • Set 1 (Increasing): Enriched in intracellular proteins (e.g., CCT8, RPL3, HSP90AA1, NAMPT) suggesting stress response and browning.
  • Set 2 (Decreasing): Enriched in secreted adipokines (e.g., AZGP1, TF, MFAP5) associated with basal homeostasis.
• Insight: Revealed a pattern of "intracellular leakage" or stress-induced secretion that was not obvious in standard group contrasts.

Case Study 2: Transcriptomics (Obesity Phenotypes)

• Context: Subcutaneous adipose tissue from Metabolically Unhealthy Obesity (MUO), Metabolically Healthy Obesity (MHO), and Metabolically Healthy Lean (MHL).
• Finding: Ranked genes from MUO $\to$ MHO $\to$ MHL to find "obesity ameliorating" factors.
  • Set 1 (Increasing toward MHL): Identified SPX (suppresses appetite) and AZGP1 (leptin sensitivity) as potential therapeutic targets.
  • Set 2 (Decreasing toward MHL): Identified GAL (increases fat intake) and CHIT1 (inflammation) as targets for inhibition.
• Insight: Successfully narrowed the search space to specific secreted factors with clear therapeutic potential.

Cross-Study Validation

• Both studies identified AZGP1 and MFAP5, demonstrating the tool's robustness.
• The tool highlighted context-dependent regulation (e.g., AZGP1 transcript levels increased in healthy states, but secretion decreased in specific stress conditions), illustrating the value of multi-level analysis.

5. Significance

• Target Prioritization: BioTrendFinder effectively reduces the "search space" for downstream experimental validation by ranking molecules based on a composite of expression trends, statistical significance, and network centrality.
• Hypothesis Generation: It enables researchers to ask "What happens as samples move along this axis?" rather than just "What is different between Group A and B?"
• Accessibility: As a free, interactive web tool (Shiny), it lowers the barrier for biologists to perform complex, integrative network analyses without requiring extensive coding skills.
• Biological Insight: By integrating functional networks with expression trends, it uncovers "functional drivers" that might be statistically significant but biologically isolated in standard analyses, or conversely, highlights biologically coherent modules that lack extreme statistical significance in single-molecule tests.

Availability:

• Web Tool: https://cphbat.shinyapps.io/biotrendfinder/
• Data/Code: https://github.com/scheelelab/BioTrendFinder_results_and_data