A Cross-Study Multi-Organ Cell Atlas ofMacaca fascicularis Informed by Human Foundation Model Annotation: A Resource for Translational Target Assessment

This study presents the largest harmonized single-cell transcriptomic atlas of *Macaca fascicularis*, integrating over 2.5 million cells across 43 organs and leveraging a human foundation model to enable scalable cross-species comparisons that improve target qualification, mechanistic toxicity interpretation, and the reduction of non-human primate use in preclinical research.

The Gist

Imagine you are trying to build a bridge between two very different islands: Human Island and Monkey Island.

In the world of medicine, scientists need to test new drugs (especially complex ones like antibodies) on animals before they can test them on people. The Cynomolgus monkey (Macaca fascicularis) is the "gold standard" bridge builder because these monkeys are genetically very similar to us. However, there's a problem: we don't have a complete, detailed map of Monkey Island. We have scattered, blurry snapshots of different neighborhoods (organs), but no unified guidebook.

This paper is like a team of cartographers who decided to fix that. They built the first-ever, high-definition, multi-organ atlas of the Cynomolgus monkey, covering over 2.5 million individual cells.

Here is how they did it and why it matters, explained simply:

1. The Problem: A Patchwork Quilt vs. A Master Blueprint

Before this study, if a scientist wanted to know how a specific drug target worked in a monkey's eye or skin, they had to hunt through dozens of different research papers. Each paper was like a different person drawing a map of the same city using different symbols, scales, and colors. It was messy, inconsistent, and hard to compare.

Furthermore, there is a growing ethical and legal push to reduce animal testing. Governments and companies want to use fewer monkeys if possible, but they can't stop using them until they are 100% sure the monkeys will react the same way humans do.

2. The Solution: The "Universal Translator" (UCE)

To solve the "messy map" problem, the researchers used a clever piece of AI technology called Universal Cell Embeddings (UCE).

Think of UCE as a universal translator or a Rosetta Stone for biology.

• Imagine you have a dictionary of human cells (the Tabula Sapiens, a famous human cell map).
• You have a pile of messy monkey cell notes.
• The AI takes the monkey notes and translates them into the human language. It says, "Ah, this monkey cell looks just like this human skin cell," or "This monkey immune cell matches this human blood cell."

By forcing both species into the same "language," the scientists could finally compare them side-by-side, cell-for-cell, across 43 different organs.

3. What Did They Find? (The Detective Work)

Once they had this unified map, they used it to solve three major mysteries in drug safety:

A. The "Skin Rash" Mystery (EGFR & NECTIN4)

Many cancer drugs cause terrible skin rashes. Scientists knew the drugs targeted certain proteins, but they didn't understand why the skin reacted so badly.

• The Old Way: Looking at the whole skin as one big lump.
• The New Way: Using the atlas, they zoomed in. They found that the drug targets were actually hiding in specific "rooms" of the skin house: the sebocytes (oil glands) and keratinocytes (skin cells).
• The Analogy: It's like realizing a fire alarm isn't going off because the whole building is on fire, but because a specific, tiny smoke detector in the kitchen is sensitive. This explains exactly why the rash happens and confirms that monkeys are a good model for predicting this human side effect.

B. The "Eye Trouble" Mystery (ADCs)

New cancer drugs called Antibody-Drug Conjugates (ADCs) are like "smart missiles" that carry poison to cancer cells. But sometimes, they accidentally hit the eyes, causing dryness or vision loss.

• The researchers used the atlas to look inside the monkey eye, cell by cell.
• They found that some drug targets were present in the monkey's eye cells (just like in humans), while others were missing.
• The Result: This helps scientists predict before a drug is even tested in a monkey whether it might hurt the eyes. If the target isn't there, they might not need to test that specific drug in a monkey at all, saving the animal and saving money.

C. The "CD28" Trap (The TGN1412 Lesson)

There was a famous medical disaster years ago where a drug caused a massive immune reaction in humans that never happened in monkeys. Why? Because the drug targeted a specific type of T-cell that exists in humans but not in monkeys.

• Using their new atlas, the scientists checked the monkey's immune system again. They confirmed that monkeys lack this specific "dangerous" T-cell subtype.
• The Lesson: This atlas acts as a safety scanner. It can tell researchers, "Hey, this drug targets something monkeys don't have. If you test this in a monkey, you won't see the human reaction. You need a different approach."

4. Why This Matters for You

This paper isn't just about monkeys; it's about better, safer, and more ethical medicine.

• Fewer Animals: By using this digital map, scientists can decide in advance which drugs actually need to be tested in monkeys and which ones don't. This helps reduce the number of animals used.
• Safer Drugs: It helps predict human side effects (like skin rashes or eye damage) much earlier in the process.
• Faster Cures: Instead of guessing, researchers can make evidence-based decisions, speeding up the path to new treatments.

The Bottom Line

Think of this paper as building a high-definition GPS for the monkey body. Before, scientists were driving blind, hoping the monkey road looked like the human road. Now, they have a detailed map that shows exactly where the roads match and where they diverge. This allows them to navigate the complex world of drug development with confidence, keeping both patients and animals safer.

Technical Summary

1. Problem Statement

Non-human primates (NHPs), specifically the cynomolgus monkey (Macaca fascicularis), are the gold standard for preclinical safety assessment of biologics due to their high genetic and physiological similarity to humans. However, the field faces three critical challenges:

• Regulatory Pressure: There is increasing global pressure (e.g., FDA 2.0 Act, EFPIA guidelines) to reduce, refine, and replace (3Rs) animal testing, necessitating better in silico and in vitro alternatives.
• Data Fragmentation: Existing single-cell transcriptomic data for cynomolgus monkeys is fragmented across numerous organ-specific studies. These datasets vary significantly in methodology, annotation granularity, and accessibility, lacking a unified, harmonized resource.
• Annotation Gap: Unlike human (e.g., Tabula Sapiens) and murine (e.g., Tabula Muris) atlases, there is no comprehensive, well-annotated single-cell reference for cynomolgus monkeys. This limits the ability to perform cross-species comparisons, qualify drug targets, and mechanistically interpret toxicity.

2. Methodology

The authors developed a scalable computational framework to integrate heterogeneous single-cell datasets and harmonize cell type annotations using a foundation model approach.

• Data Curation:

  • NHP Data: Aggregated 30 publicly available studies (as of Jan 2025) covering 57 anatomical regions, 43 organs, and 14 physiological systems.
  • Human Data: Utilized the Tabula Sapiens V2 (28 organs) as the high-quality reference atlas.
  • Preprocessing: Applied rigorous quality control (QC), including ambient RNA correction (SoupX), doublet removal (Scrublet), and filtering based on mitochondrial gene expression and gene counts.

• Cross-Species Integration via Universal Cell Embeddings (UCE):

  • Instead of traditional ortholog-based mapping, the study employed Universal Cell Embeddings (UCE), a self-supervised foundation model trained on multi-species data without relying on cell type labels.
  • Both human and cynomolgus datasets were embedded into a shared 1280-dimensional latent space. This allows new cells from any species to be projected into the reference space without retraining the model.

• Cell Type Annotation Strategy:

  • Human-to-NHP Transfer: For 27 organs with human counterparts, a k-Nearest Neighbors (KNN) classifier (k=20) was trained on human UCE embeddings to predict cell types in cynomolgus data.
  • NHP-to-NHP Transfer: For organs lacking human counterparts, labels were transferred from other well-annotated cynomolgus studies (e.g., from Han et al. to Qu et al.).
  • Confidence Filtering: Predictions were retained only if the posterior probability was $\ge$ 0.5; otherwise, cells were labeled "undefined."

• Validation & Use Cases:

  • Validated annotation transfer by comparing predicted labels against original author annotations and checking for tissue-specific marker concordance.
  • Applied the atlas to three translational use cases: concordant toxicity targets, ocular toxicity in Antibody-Drug Conjugates (ADCs), and species-specific immune profiling (CD28).

3. Key Contributions

• The Largest Cynomolgus Atlas: Created a unified atlas comprising 2.5 million cells from 30 studies, representing the most extensive single-cell resource for M. fascicularis to date.
• Foundation Model-Based Harmonization: Demonstrated the efficacy of UCE for cross-species annotation, achieving high concordance (median posterior probability ~0.95) between human and NHP cell identities without species-specific fine-tuning.
• Modular Framework: Established a reproducible pipeline for integrating fragmented NHP data, enabling future expansion to other species or experimental conditions.
• Open Resource: All data, scripts, and the final annotated object are made publicly available to accelerate translational research.

4. Key Results

• Atlas Composition: The final dataset includes 2.59 million cells across 43 organs and 14 systems. Demographics show representation across age groups (juvenile to aged) and sexes.
• Annotation Quality:
  • High agreement was observed between transferred labels and original annotations (e.g., in the bladder and skin).
  • Some discrepancies were noted in organs with lower prediction probabilities (e.g., trachea, uterus) or where domain shifts existed, highlighting the need for expert review in specific contexts.
• Translational Use Cases:
  1. Concordant Toxicity: Analyzed targets (CD44, CD47, EGFR, F3, NECTIN4, VEGFR2) associated with known toxicities. Found that while organ-level expression often correlated, single-cell resolution was crucial. For example, EGFR and NECTIN4 showed low organ-level ranking but high enrichment in specific skin subtypes (keratinocytes, sebocytes), explaining skin toxicity mechanisms.
  2. Ocular Toxicity (ADCs): Investigated ADC targets in eye tissues. Found strong cross-species concordance for targets like NECTIN4 and ERBB2 in conjunctival/corneal epithelial cells, explaining dry eye/keratitis. Conversely, FOLR1 and MET showed divergent expression patterns between human and NHP eyes, suggesting potential translational gaps for specific ADCs.
  3. Immune Profiling (CD28): Reproduced the known species-specific difference in CD28 expression. The atlas confirmed that CD28 is highly expressed in human CD4+ effector memory T-cells but not in cynomolgus counterparts, providing a mechanistic explanation for the failure of the TGN1412 (Theralizumab) trial in humans despite safe NHP studies.

5. Significance

• Advancing 3Rs: This resource directly supports the "Reduction" and "Refinement" of NHP use by enabling in silico target qualification and toxicity prediction. It helps researchers determine when an NHP study is necessary versus when human-relevant in silico models suffice.
• Mechanistic Insight: By resolving expression at the cell-subtype level, the atlas moves beyond bulk tissue analysis, uncovering mechanisms of toxicity (e.g., specific cell types driving skin or eye adverse events) that were previously obscured.
• Regulatory Alignment: The framework aligns with emerging FDA and EFPIA guidance on New Approach Methodologies (NAMs), providing a data-driven basis for evidence-based decisions on the relevance of NHP models in biologics development.
• Future-Proofing: The use of foundation models (UCE) creates a scalable architecture that can easily incorporate new datasets or species, fostering a continuous improvement cycle for preclinical safety assessment.