Developing And Internally Validating AI-Based Aging Resilience Biomarkers in Non-Human Primates

This study develops and internally validates AI-based "Aging Resilience" biomarkers using longitudinal clinical data from baboons and rhesus macaques, demonstrating that non-linear machine learning models outperform linear approaches in predicting mortality by capturing biological resilience rather than just chronological age.

The Gist

The Big Idea: Measuring "Wear and Tear" Instead of Just "Time"

Imagine you have two cars.

• Car A is 10 years old, but it has been driven gently, kept in a garage, and gets regular oil changes. It still starts every morning and runs smoothly.
• Car B is also 10 years old, but it has been raced on dirt tracks, ignored by mechanics, and runs on cheap fuel. It sputters, leaks oil, and might break down tomorrow.

If you only look at the calendar, both cars are exactly the same age. But if you look at the engine, they are in completely different places.

This paper is about building a tool to measure the "engine condition" (biological health) of non-human primates (baboons and monkeys), rather than just counting how many years they have been alive. The researchers call this tool "Aging Resilience."

The Problem: We Didn't Have a Good "Check Engine" Light

For humans and mice, scientists have developed "aging clocks" that can tell you how fast your body is aging based on blood tests. But for monkeys and baboons (which are crucial for testing human medicines), we didn't have a good way to do this.

Why? Because monkey medical records are messy. They don't have perfect "normal" ranges for blood tests, and the data is collected over many years by different vets. It's like trying to solve a puzzle where half the pieces are missing and the picture on the box is blurry.

The Solution: Teaching AI to Read the Story

The researchers gathered two huge piles of data:

1. The "Big & Simple" Pile: 4,300 baboons with 19 basic health checks (like weight and blood counts).
2. The "Small & Detailed" Pile: 281 monkeys with 80 detailed health checks (including heart scans and specific metabolic markers).

They fed this data into five different types of Artificial Intelligence (AI) models. Think of these models as different types of detectives:

• The Linear Detective (Simple Math): Looks for straight lines. "If weight goes down, age goes up."
• The Complex Detective (Neural Networks): Looks for hidden, twisting patterns. "If weight goes down and blood sugar spikes and the heart rate changes in a specific rhythm, then the body is struggling."

The Big Surprise: The "Perfect" Detective Was Wrong

The researchers asked the AI: "Can you guess the animal's actual age just by looking at its health data?"

• The Result: The Simple Math detectives were amazing at this. They could guess the age almost perfectly (99% accuracy).
• The Twist: When the researchers asked, "Okay, now which animals are going to die soon?" the Simple Math detectives failed miserably. They knew the animals were "10 years old," but they couldn't tell you which one was sick.

The Analogy: The Simple Math detective is like a librarian who knows exactly how many books are on a shelf (the age), but doesn't know if the books are falling apart (the health).

The Complex Detectives (the AI that looked for twisting patterns) were slightly worse at guessing the exact age, BUT they were incredible at predicting who would die soon. They could see the subtle "cracks" in the engine that the simple math missed.

The New Metric: "Aging Resilience"

Instead of just saying "This monkey is 15 years old," the researchers created a new score called Aging Resilience (AR).

• Rate of Aging (RoA): How fast is the car rusting right now?
• Normalized Cumulative Aging (NCA): How much rust has accumulated over the entire life of the car?

The Finding: The "Total Rust" score (NCA) was the best predictor of death. It didn't matter if the car was 10 or 15 years old; if the total rust was high, the car was likely to break down soon.

Why This Matters

1. It Works on Messy Data: This method works even with imperfect, real-world veterinary records. You don't need expensive, fancy lab tests; you just need the routine blood work and check-ups that vets already do.
2. It Finds Trouble Early: Because it looks at the history of the data, it can spot a monkey that is "aging faster than it should" long before it gets sick. This is like catching a flat tire before the car breaks down on the highway.
3. It Helps Humans: Since monkeys are very similar to humans, if we can test anti-aging drugs on them using this "Resilience Score," we can be much more confident that the drug will work for us, too.

The Bottom Line

This study proved that knowing how old someone is isn't the same as knowing how healthy they are.

By using advanced AI to look at the story of a monkey's health over time, scientists can now measure "biological resilience." It's a new way to tell the difference between a healthy 15-year-old and a fragile one, giving us a powerful new tool to fight aging and disease.

Technical Summary

1. Problem Statement

The quantification of biological aging is critical for geroscience, yet existing biomarkers for Non-Human Primates (NHPs) are limited. Current challenges include:

• Lack of NHP-specific metrics: While frailty indices exist for humans and rodents, NHPs lack standardized clinical reference ranges and normative values, making traditional deficit-accumulation models difficult to apply.
• Data Limitations: NHP clinical data is often longitudinal but heterogeneous, with varying numbers of features and animals across datasets.
• The "Chronological vs. Biological" Paradox: Traditional models often excel at predicting chronological age (time) but fail to capture biological resilience (healthspan/mortality risk). In NHPs, where humane euthanasia often occurs at the first sign of serious illness, metrics must detect decline before overt morbidity.
• Need for Translational Tools: There is a need for scalable, AI-driven frameworks using routine veterinary records (analogous to human Electronic Health Records) to bridge the gap between murine models and human clinical trials.

2. Methodology

The study developed and validated a computational workflow to derive "Aging Resilience" (AR) metrics from longitudinal routine clinical data.

Datasets:
Two distinct NHP cohorts were utilized to test generalizability:

1. SNPRC (Southwest National Primate Research Center): 4,328 baboons (Papio spp.), 9,836 observations, 19 clinical features (primarily hematological and body weight).
2. RLEC (Radiation Late Effects Cohort): 281 rhesus macaques (Macaca mulatta), 1,227 observations, 80 clinical features (including cardiac imaging, DEXA scans, and metabolic markers).

Computational Pipeline:

• Preprocessing: Data was aggregated by animal ID and year. Missing values were handled via Multivariate Imputation by Chained Equations (MICE). Features were normalized (z-score), and temporal features (sequence index, relative time) were engineered for recurrent models.
• Model Training: Five computational architectures were trained to predict chronological age:
  1. Penalized Multiple Linear Regression (MLR)
  2. Linear Mixed-Effects Models (LMM)
  3. Random Forest (RF)
  4. Feed-Forward Neural Network (NN)
  5. Recurrent Neural Network (RNN)
  • Note: A Large Language Model (LLM) was tested but excluded due to poor performance and high cost.
• Derivation of Aging Resilience (AR) Metrics: Instead of using the age prediction directly, the study utilized prediction residuals (Predicted Age – Chronological Age) to calculate two categories of metrics:
  1. Normalized Cumulative Aging (NCA): The Area Under the Curve (AUC) of residuals over time, representing the cumulative burden of physiological deviation.
  2. Rate of Aging (RoA): The slope of residuals over time (Linear, LMM-based, and LMM-with-year controls), representing the velocity of aging.
• Validation: AR metrics were validated against the Age at Death in a "Natural Death" cohort (animals euthanized due to age-related decline or natural causes) using Pearson correlation.

3. Key Results

A. Chronological Age Prediction Performance:

• RLEC Cohort (Feature-Rich): Linear models (LMM and MLR) achieved near-perfect accuracy (Test $R^2 \approx 0.99$ and $0.92$, respectively). This was likely driven by strong linear correlations between physical traits (e.g., bone mass, body composition) and time.
• SNPRC Cohort (Feature-Sparse): Non-linear models performed better, with Random Forest achieving the highest Test $R^2$ (0.353). Linear models struggled ($R^2 < 0.20$) due to the lack of high-variance physical features.

B. The Critical Paradox:

• Models that best predicted chronological age (Linear models in RLEC) correlated poorly with lifespan.
• Conversely, AR metrics derived from non-linear models (RNN and RF), which were less accurate at predicting exact chronological age, showed strong predictive validity for mortality (Pearson's $r > 0.8$).

C. Feature Importance:

• SNPRC (Baboons): Aging signatures were dominated by hematological and immunological markers (Red Blood Cells, Mean Platelet Volume), suggesting "inflammaging" and declining circulatory efficiency.
• RLEC (Macaques): Aging signatures were cardiometabolic (Whole Body Mass, Alkaline Phosphatase, Left Ventricular Internal Diameter), reflecting structural heart changes and metabolic shifts.

D. Temporal Robustness:

• The AR metrics required data approaching late-life stages to achieve significant correlation with mortality. Early-life data (e.g., <30% of lifespan) did not yield robust predictive correlations, indicating these metrics capture accumulated burden rather than early developmental trajectories.

4. Key Contributions

1. First NHP Aging Resilience Framework: Established a scalable methodology to derive biological aging metrics from routine clinical records in NHPs, overcoming the lack of standardized normative values.
2. Validation of the "Non-Linear" Advantage: Demonstrated that while linear models are superior for tracking chronological time, non-linear models (RNN/RF) are superior for capturing the complex, non-linear physiological dysregulation that dictates healthspan and mortality.
3. Metric Definition: Defined and validated two distinct AR metrics: NCA (cumulative burden) and RoA (velocity). The study found NCA to be a more robust predictor of mortality than RoA in these datasets.
4. Translational Bridge: Provided a workflow that uses standard clinical data (similar to human EHRs), making the approach directly adaptable for monitoring human patient resilience and testing anti-aging interventions.

5. Significance

This study fundamentally shifts the paradigm of aging biomarkers in translational models. It proves that a "good" aging clock is not necessarily one that predicts the calendar year accurately, but one that quantifies the biological deviation from that timeline.

• For Geroscience: It offers a tool to detect subtle biological decline in "healthy" NHPs before clinical symptoms appear, allowing for earlier intervention testing.
• For Human Medicine: Since the methodology relies on routine clinical data (blood work, vitals, basic imaging), it provides a blueprint for developing similar "Resilience Clocks" for human populations without requiring expensive omics or novel biomarkers.
• Limitations: The "black box" nature of deep learning models complicates direct biological interpretation of specific feature directions, and the metrics currently require late-life data to reach peak predictive power.

In conclusion, the authors successfully established that Aging Resilience metrics derived from non-linear AI models are superior predictors of lifespan in NHPs compared to traditional chronological age predictors, offering a powerful new tool for preclinical aging research.