The End of Aging Clocks: Training Foundation Models to Reason in Aging and Longevity

This paper introduces Longevity-LLM v0.1, a fine-tuned Qwen3-14B foundation model that unifies and surpasses specialized aging clocks across multiple data modalities by achieving superior accuracy in age prediction and demonstrating versatile capabilities in tasks like proteomic profile generation.

The Gist

Imagine you have a library full of different experts. One expert is a master at reading DNA to tell how old you are. Another is a wizard at analyzing blood proteins. A third is a detective who looks at clinical charts to predict your health risks.

For years, scientists have been using these "experts" (called aging clocks) to study how we get older. But there's a big problem: they can't talk to each other. If you want to know your age based on DNA, you need Expert A. If you want to know based on blood proteins, you need Expert B. They are like silos; they only speak one language and only look at one type of data. If you want to combine them, you have to build a whole new machine.

The Big Idea: The "Super-Intern"
The researchers at Insilico Medicine asked a bold question: What if we didn't need a library of separate experts? What if we could train one single, super-smart intern to learn everything all of them know?

They created Longevity-LLM, a 14-billion-parameter Artificial Intelligence (AI) model. Think of it not as a calculator, but as a super-intelligent student who has read every single textbook, research paper, and dataset about aging ever written.

Here is how they did it, using some simple analogies:

1. The Training: From "Textbook" to "Practice"

The AI started as a general language model (like a very smart chatbot) that knew how to write poems and answer questions, but knew nothing about biology.

• Supervised Fine-Tuning (SFT): Imagine giving this student a massive stack of flashcards. On one side is a piece of biological data (like a list of DNA markers), and on the other side is the answer (like "This person is 50 years old"). The student memorized these patterns. They learned to read DNA, proteins, and blood test results just like they read sentences.
• Reinforcement Fine-Tuning (RFT): This is the "coach" phase. The student was given a test. If they guessed the age correctly, the coach gave them a high-five (a reward). If they were wrong, the coach said, "Try again, but think harder about why." The AI was forced to "think out loud" (reasoning) before giving an answer, which helped it understand the logic of aging, not just memorize numbers.

2. The Results: Beating the Specialists

Once trained, they put Longevity-LLM to the test against the old "experts" and the world's most advanced AI chatbots.

• The DNA Test: The old gold-standard DNA clock (Horvath clock) is like a veteran detective who has solved thousands of cases. Longevity-LLM, after its training, actually solved the case better than the veteran, predicting biological age with an error of only about 4.3 years.
• The Protein Test: The AI was asked to look at blood protein levels and guess a person's age. It did just as well as the specialized machines built just for that job.
• The "Creative" Test: Here is the magic trick. The researchers asked the AI: "If a 60-year-old man has these specific proteins, what would his blood look like if he were 40?" The AI didn't just guess a number; it invented a new, realistic protein profile for a younger version of that person. No other AI could do this. It's like asking a painter to not just describe a sunset, but to paint a brand new, beautiful sunset that has never existed before.

3. Why This Matters: The "Swiss Army Knife" vs. The "Specialized Tool"

Before this, if a scientist wanted to study aging, they had to use a different tool for every type of data. It was like having a screwdriver for screws, a hammer for nails, and a saw for wood, and having to switch tools constantly.

Longevity-LLM is the Swiss Army Knife of Aging.

• It can look at DNA, proteins, or medical records.
• It can predict your age.
• It can guess if you might get sick.
• It can even imagine what your body could look like if you were younger.

The Bottom Line

The authors believe this is just the beginning. They aren't just trying to build a better calculator; they are building a research partner.

In the future, instead of a scientist spending months analyzing data, they could just chat with Longevity-LLM: "Hey, look at this patient's DNA and blood work. What's their biological age? Why are they aging faster than their calendar age? And what drug might slow them down?"

The AI would reason through the answer, connecting the dots between different types of data, effectively becoming a single, unified brain that understands the complex story of human aging. This could speed up the discovery of life-extending medicines from decades to years.

Technical Summary

`).
* Thinking Length Reward: Incentivizing substantive reasoning up to 5,000 characters.
* Regression Correctness Reward: Normalized absolute error against ground truth.
* Penalty: Applied for predictions outside validity bounds.
* Constraints: A Kullback-Leibler (KL) penalty prevented policy collapse by keeping the model close to the SFT reference.

3. Key Contributions

• Aging Clock Distillation: Demonstrated that a single LLM can internalize the functionality of dozens of specialist aging clocks, eliminating the need for external tool integration.
• Unified Multimodal Reasoning: The model can process DNA methylation, proteomics, transcriptomics, and clinical data within a single framework without architectural changes.
• Generative Capabilities: Beyond prediction, the model can generate biologically plausible proteomic profiles conditioned on age and sex, a task where it significantly outperformed other frontier LLMs.
• Interpretability via Reasoning: By forcing the model to output reasoning traces (<think>), the system provides a pathway to understanding why a biological age prediction was made, moving beyond "black box" regression.

4. Key Results

The model was evaluated against 15 frontier commercial LLMs and established specialist clocks on the Longevity Bench and specific holdout datasets.

• Longevity Bench Performance:
  • Longevity-LLM ranked 1st in 4 out of 7 tasks, including RNA-based cancer prognosis (0.77 accuracy) and 10-year mortality prediction (0.89 accuracy).
  • It significantly outperformed the base Qwen3-14B (which failed to produce valid outputs) and other frontier models like Gemini 3 Pro.
• Epigenetic Age Prediction (DNAm):
  • SFT Stage: Achieved a Mean Absolute Error (MAE) of 5.91 years.
  • RFT Stage: Reduced MAE to 4.34 years, significantly outperforming the established Horvath multi-tissue clock (MAE 4.61) on the same holdout set (p < 0.05).
• Proteomic Age Prediction:
  • Achieved an MAE of 7.9 years on 94 holdout samples, performing within the accuracy range of specialist proteomic clocks trained on much larger datasets (e.g., UK Biobank).
• Proteomic Profile Generation:
  • When asked to predict the 25 most abundant plasma proteins based on partial profiles, the model achieved a Jaccard index of 0.072, significantly higher than any other tested frontier model.

5. Significance and Future Outlook

• Paradigm Shift: This work challenges the necessity of specialized, single-modality aging clocks. It suggests that a "modestly sized" (14B parameter) LLM can replace a suite of specialist tools, offering a "prompt-to-drug" pipeline.
• Monolithic AI for Drug Discovery: By unifying disparate data modalities into a single model, the authors aim to create a "pharmaceutical superintelligence" capable of handling the entire drug development lifecycle—from target identification to clinical trial design—while maintaining internal biological consistency.
• Scalability and Generalization: The success of the model on proteomic data (despite limited training data) suggests that knowledge learned from one modality (e.g., DNAm) can transfer to and improve performance in underrepresented modalities.
• Future Directions: The authors plan to scale the RFT component, incorporate more diverse datasets, and refine the model's ability to articulate the biological logic behind its predictions, aiming for a "companion" in aging research rather than just a prediction endpoint.

Conclusion:
Longevity-LLM v0.1 represents a proof-of-concept that foundation models can internalize complex biological signals across multiple omics layers. By distilling the knowledge of specialist aging clocks into a conversational LLM, the authors have demonstrated a path toward a unified, interpretable, and generative AI system for biogerontology and drug discovery.