💬 NLP

Exploring Anti-Aging Literature via ConvexTopics and Large Language Models

This paper proposes a convex optimization-based clustering method that guarantees global optima and produces stable, interpretable topics, demonstrating superior reproducibility and effectiveness compared to traditional approaches like K-means and LDA when applied to a large corpus of anti-aging biomedical literature.

Original authors: Lana E. Yeganova, Won G. Kim, Shubo Tian, Natalie Xie, Donald C. Comeau, W. John Wilbur, Zhiyong Lu

Published 2026-02-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Lana E. Yeganova, Won G. Kim, Shubo Tian, Natalie Xie, Donald C. Comeau, W. John Wilbur, Zhiyong Lu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you walk into a massive library that is growing so fast, it adds a new book every second. This library is full of scientific papers about how to live longer, stay young, and fight aging. The problem? There are so many books that no human can read them all, organize them, or find the hidden patterns. It's like trying to find a specific needle in a haystack that keeps growing bigger every day.

This paper introduces a new, super-smart librarian named ConvexTopics to solve this mess. Here is how it works, explained simply:

1. The Problem with Old Librarians

Previous methods for organizing these books (like K-means or LDA) are a bit like guessing games.

The "Guessing Game" Flaw: You have to tell them exactly how many piles to make (e.g., "Make 50 piles"). If you guess wrong, the piles are messy.
The "Local Optimum" Trap: Imagine you are trying to find the highest peak in a foggy mountain range. Old methods might get stuck on a small hill and think, "This is the top!" when there's a much higher mountain right next to them. They get stuck in local solutions and miss the big picture.
Inconsistency: If you run the same old method twice, you might get two different results. That's bad for science.

2. The New Solution: ConvexTopics

The authors created ConvexTopics, which is like a librarian who never gets stuck and never needs you to guess the number of piles.

The "Global Optimum" Guarantee: Instead of wandering around in the fog, this new method looks at the entire mountain range at once. It uses a special mathematical trick (convex optimization) that guarantees it will always find the absolute highest peak. It never gets stuck on a small hill.
No Guessing Required: You don't have to tell it how many topics to find. It looks at the data and says, "Okay, there are exactly 2,194 distinct themes here." It figures out the perfect number automatically.
The "Exemplar" Strategy: Instead of inventing a fake "average" topic, it picks real, actual sentences from the papers to represent each topic. It's like saying, "This specific book is the best example of 'Diet'," rather than making up a generic description.

3. Testing the Librarian

The team tested this new librarian on two types of libraries:

General News: Old news articles about politics and sports.
Biomedical Science: About 12,000 papers specifically about anti-aging.

They compared ConvexTopics against the old methods (K-means, LDA) and a fancy new AI method (BERTopic).

The Result: ConvexTopics won. It organized the books much better, matching what human experts would have done, and it did it faster. It was especially good at finding the right "buckets" for complex medical topics without needing a human to tell it how many buckets to use.

4. Putting It to Work: The Anti-Aging Map

The authors used this tool to map out the world of anti-aging research. They found hundreds of tiny, specific topics, such as:

Sarcopenia: The specific science of why our muscles shrink as we get older.
Gut Microbiota: How the bacteria in our stomachs might help us live longer.
NAD+ Supplements: A popular pill people take for energy, which the research shows might have risks we don't fully understand yet.

The "Smart Summary" Trick:
Because the computer found so many tiny topics (over 2,000!), it would be overwhelming for a human to read them all. So, they used a Large Language Model (like the AI behind this explanation) to group these tiny topics into big, easy-to-read categories like "Exercise," "Diet," and "Genetics."

Think of it like this: The computer builds the foundation and the walls of a house (the hard, math-heavy part), and then a human-friendly AI paints the rooms and puts up signs so you know where the kitchen and bedroom are.

5. Why This Matters

Trustworthy: Because the math guarantees the best answer, scientists can trust the results more than with older methods.
Transparent: You can see exactly which real papers make up a topic.
Real-Time: It's fast enough that you could build a website where a doctor or a curious person types in "anti-aging," and the system instantly shows them a clear map of what science knows (and doesn't know) about it.

In a nutshell: This paper gives us a new, mathematically perfect way to organize the explosion of medical knowledge. It stops us from getting lost in the noise and helps us find the real signals about how to live healthier, longer lives.

1. Problem Statement

The biomedical literature is expanding exponentially, creating significant challenges for organizing knowledge and detecting emerging trends. Traditional clustering and topic modeling methods (e.g., K-means, LDA, BERTopic) suffer from several critical limitations:

Sensitivity to Initialization: They often converge to local optima rather than global optima, leading to non-reproducible results depending on random seeds.
Hyperparameter Dependency: Users must predefine the number of topics ( $K$ ), a decision that is often arbitrary and lacks methodological support.
Evaluation Difficulties: Automated metrics (like coherence scores) correlate poorly with human judgment, while expert evaluation does not scale.
Lack of Interpretability: In complex domains like anti-aging research, existing methods struggle to produce fine-grained, reproducible topics that align with established biomedical categories (e.g., MeSH terms).

2. Methodology: ConvexTopics

The authors propose ConvexTopics, a novel algorithm that reformulates clustering as a convex optimization problem to guarantee a globally optimal solution.

Core Algorithmic Framework

Exemplar-Based Clustering: Adapted from Lashkari and Golland (2019), the method assumes cluster centers are not abstract vectors but actual data points (exemplars) from the dataset. This restricts the search space from infinite/continuous to finite/discrete.
Convex Objective Function: By using an exemplar-based likelihood function, the objective becomes convex, ensuring convergence to a global optimum via an efficient iterative algorithm.
Automatic Topic Determination: The number of topics is not pre-specified. Instead, the algorithm computes a probability ( $q_j$ ) for each data point being a source (exemplar). Only points with $q_j > 0$ become cluster centers, automatically determining the optimal number of topics.
NLP Adaptation:
- Vocabulary Construction: Uses single terms (minus stop words) and noun phrases. Terms are filtered using a Hypergeometric distribution p-value and Benjamini–Hochberg correction (FDR = 0.01) to ensure statistical significance.
- Similarity Measure: Replaces Euclidean distance with the Dice coefficient to measure term co-occurrence probabilities ( $s_{ij}$ ). A cutoff ( $s_{ij} < 0.05$ ) creates a sparse matrix for computational efficiency.
- Scoring: Documents are scored against topics based on term weights ( $r_{ij}$ ) and local document weights.

Evaluation Framework: MaxMAP Topic Alignment

To address the lack of scalable evaluation, the authors introduce MaxMAP (Maximum Mean Average Precision) Topic Alignment:

Ground Truth: Uses MeSH (Medical Subject Headings) terms assigned to PubMed articles as expert-curated labels.
Process: For each qualified MeSH term, the system calculates the Average Precision (AP) across the generated topics. The topic with the highest AP is assigned to that MeSH term.
Metric: The final score is the Mean Average Precision (MAP) of the top $N$ MeSH terms. This directly measures how well a topic aligns with expert-defined biomedical concepts, rather than just internal word coherence.

Integration with Large Language Models (LLMs)

Fine-Grained to Macro-Structure: ConvexTopics generates hundreds of fine-grained topics. To make this navigable, the authors use GPT-4o to group these fine-grained topics into higher-level thematic categories.
Summarization: GPT-4o generates titles and summaries for each topic based on the top 10 scoring documents, ensuring the output is human-readable while remaining grounded in the algorithmically derived data.

3. Key Contributions

ConvexTopics Algorithm: A robust, reproducible clustering method that guarantees global optimality and automatically determines the number of topics without user intervention.
MaxMAP Evaluation: A novel, scalable evaluation metric using MeSH terms that correlates strongly with expert biomedical knowledge, overcoming the limitations of traditional coherence metrics.
Hybrid Human-AI Workflow: A framework that combines the stability of convex clustering with the summarization capabilities of LLMs, mitigating LLM hallucinations by anchoring them to specific, algorithmically verified topics and documents.
Application to Anti-Aging: A comprehensive analysis of ~12,000 PubMed articles on anti-aging, revealing fine-grained insights into molecular mechanisms, dietary supplements, and lifestyle factors.

4. Results

The method was evaluated on general NLP benchmarks (Reuters-21578, 20-Newsgroups) and three biomedical datasets (Anti-Aging, Diabetes, Age-Related Macular Degeneration).

Performance on Benchmarks:
- Reuters-RCV1: ConvexTopics achieved a MaxMAP of 0.2506, outperforming LDA (0.2291), K-means (0.1637), and BERTopic (0.1640).
- 20-Newsgroups: ConvexTopics achieved 0.2759, outperforming LDA (0.2673), K-means (0.1403), and BERTopic (0.1618).
Performance on Biomedical Datasets:
- Anti-Aging: ConvexTopics scored 0.4850, significantly higher than LDA (0.3744), K-means (0.3470), and BERTopic (0.3372).
- Diabetes & AMD: Similar trends were observed, with ConvexTopics consistently achieving the highest MaxMAP scores across all datasets.
Efficiency: The algorithm executed in less than a minute for large datasets using 20 threads.
Expert Validation: Manual review confirmed that ConvexTopics topics (e.g., Sarcopenia, Gut Microbiota) aligned closely with accepted biomedical categories and captured multidimensional research trends (mechanistic, clinical, and therapeutic).

5. Significance and Future Work

Scalability and Reproducibility: The method provides a foundation for building real-time, web-accessible tools for knowledge discovery in PubMed, removing the "black box" nature of heuristic-based clustering.
Scientific Utility: The application to anti-aging literature successfully identified nuanced topics (e.g., the difference between limb muscle and diaphragm aging, the role of NAD+ supplements, and the limitations of FMT in longevity) that are difficult to surface with traditional methods.
Limitations & Future Directions:
- The method can generate a large number of narrowly focused clusters, requiring LLMs for aggregation.
- Occasional generation of large, less meaningful clusters based on specific exemplar terms.
- Future work aims to quantify the impact of LLM-generated summaries on sentiment and detail accuracy and to expand the tool to allow researchers to query any collection of biomedical articles in real-time.

In conclusion, ConvexTopics represents a significant advancement in biomedical text mining by offering a mathematically rigorous, reproducible, and interpretable approach to organizing vast scientific literature, bridging the gap between raw data and actionable scientific insight.