Open Notebook

Monday, March 24, 2025

Asian Immune Diversity Atlas (AIDA) CZI Cell Science

https://cellxgene.cziscience.com/collections/ced320a1-29f3-47c1-a735-513c7084d508

Asian Immune Diversity Atlas (AIDA)

CZI Cell Science

The relationships of human diversity with biomedical phenotypes are pervasive, yet remain understudied, particularly in a single-cell genomics context. Here we present the Asian Immune Diversity Atlas (AIDA), a multi-national single-cell RNA-sequencing (scRNA-seq) healthy reference atlas of human immune cells. AIDA comprises 1,265,624 circulating immune cells from 619 donors, spanning 7 population groups across 5 Asian countries, and 6 controls. Though population groups are frequently compared at the continental level, we found that sub-continental diversity, age, and sex pervasively impacted cellular and molecular properties of immune cells. These included differential abundance of cell neighbourhoods, as well as cell populations and genes relevant to disease risk, pathogenesis, and diagnostics. We discovered functional genetic variants influencing cell type-specific gene expression which were under-represented in non-Asian populations, and helped contextualise disease-associated variants. AIDA enables analyses of multi-ancestry disease datasets and facilitates the development of precision medicine efforts in Asia and beyond. Please note that the AIDA Phase 1 Data Freeze v1 object comprises 1,058,909 peripheral blood mononuclear cells from 503 healthy donors from Japan, Singapore, and South Korea alongside common controls. This first AIDA data freeze was released to the research community pre-publication, and was also part of the first CZ CELLxGENE Census assembled in May 2023. Going from Data Freeze v1 to Data Freeze v2, we added additional healthy Asian donor samples and control samples - please see our publications for further details.

Group fairness in AI

Group fairness in AI refers to ensuring that machine learning models treat different demographic groups equitably by achieving parity in statistical measures across groups defined by sensitive attributes like race, gender, or age. This approach evaluates fairness at the population level rather than the individual level16.

Key Principles

Protected Groups: Groups are defined by sensitive features (e.g., race, gender), which may or may not have privacy implications16.
Statistical Parity: Requires outcomes to be independent of sensitive attributes. For example, demographic parity mandates equal acceptance rates for job applicants across groups2 37.

Equality of Metrics: Common fairness metrics include:

Metric	Definition
Equalized Odds	Equal true positive and false positive rates across groups4 7
Equality of Opportunity	Equal true positive rates across groups7
Predictive Parity	Similar precision (positive predictive value) for all groups7

Implementation Challenges

Trade-offs with Individual Fairness: Group fairness may conflict with individual merit. For instance, enforcing demographic parity in hiring could prioritize underrepresented candidates over more qualified ones to meet statistical targets23.
Defining Sensitive Features: While some attributes (e.g., race) are obvious, others (e.g., language proficiency) require context-specific analysis16.
Mathematical Constraints: Fairness is often framed as parity in expectations over the distribution of data, such as $P(\hat{Y}=1 \mid A=a) = P(\hat{Y}=1 \mid A=b)$ for demographic parity, where $\hat{Y}$ is the prediction and $A$ is the sensitive attribute56.

Evaluation

Confusion Matrix Analysis: Metrics like recall, precision, and accuracy are compared across subgroups4.
Toolkits: Frameworks like Fairlearn and AI Fairness 360 operationalize group fairness through parity constraints, enabling developers to audit and mitigate biases16.

While group fairness is widely adopted, it has limitations—it cannot address individual-level disparities and may require balancing trade-offs with model accuracy2 35.

Citations:

Answer from Perplexity: pplx.ai/share

Sunday, March 23, 2025

Human Brain scRNA-seq Data related to AD and cognitive decline

Aging/AD Cortical Cell Atlas
- Profiles 24 human frontal cortex samples with clinicopathologic traits (including cognitive decline metrics) using single-nucleus RNA-seq12.
- Identified cell subtypes (e.g., SST+ inhibitory neurons, oligodendroglial states) and multicellular communities altered in Alzheimer’s disease (AD). Associations with tau pathology and cognitive decline were statistically validated12.
SCAD-Brain Database
- Aggregates 17 AD/mild cognitive impairment (MCI) projects from public repositories (GEO, Synapse):
  - 21 datasets | 359 samples | 10 brain regions | 1.56 million cells36.
  - Supports analysis of cell communication, trajectory, and differential gene expression in AD/MCI vs. controls36.
- Access: SCAD-Brain.

TauAge

TauAge model and its significance in understanding regional tau pathology in aging, Alzheimer's disease (AD), and Primary Age-Related Tauopathy (PART):

🧬 What is TauAge?

TauAge is a DNA methylation (DNAm)-based epigenetic "clock" created by the authors to estimate the severity of phosphorylated tau (p-tau) pathology in specific brain regions.

Unlike traditional epigenetic clocks (e.g., Horvath clock) that estimate biological age, TauAge is trained to predict the pathological burden of p-tau, adjusted for chronological age. It provides a molecular readout of tau accumulation independent of aging.

⚙️ How TauAge Was Built

Input Data:
- DNA methylation profiles (DNAm) from the frontal cortex (Illumina arrays).
- Corresponding histological quantification of p-tau in either the hippocampus or frontal cortex (depending on model).
- Two cohorts: PWG (PART-only) and ROSMAP (AD and PART).
Modeling Approach:
- The authors used elastic net regression (a regularized machine learning model) to predict age-adjusted residuals of p-tau.
- This means they removed the effect of chronological age to focus on DNAm markers that predict tau pathology beyond what you'd expect for someone’s age.
Output:
- A TauAge score per sample, reflecting how much tau burden is present in a region (hippocampus or frontal cortex), based purely on DNAm.
- Each model uses hundreds of CpG sites as features.

🧠 Why Separate Models for Hippocampus and Frontal Cortex?

Tau pathology originates in the hippocampus (early) and spreads to the frontal cortex (later) in AD.
PART, by contrast, typically remains confined to the hippocampus and does not spread to cortex.

Thus, the biology driving p-tau accumulation is likely region-specific. The authors found:

Only 8 CpGs overlapped between the hippocampal and frontal TauAge models.
The models were not interchangeable — e.g., using hippocampal CpGs to predict frontal tau gave poor results.

➡️ Conclusion: p-tau pathology is regulated by distinct molecular programs in different brain regions.

🧠 What Does TauAge Reveal About AD vs. PART?

1. Hippocampal TauAge (PART & AD)

Biological associations:
- Genes linked to TauAge CpGs are enriched in:
  - Synaptic transmission
  - Ion transport
  - Mitochondrial function
- These functions are shared in both PART and AD.
Interpretation:
- Tau accumulation in the hippocampus is driven by synaptic dysfunction, regardless of whether it's PART or AD.
- Suggests that PART and early AD may share early tau-related vulnerability via similar epigenetic pathways.

2. Frontal Cortex TauAge (Only in AD)

Biological associations:
- CpG-correlated genes are enriched in:
  - Inflammatory signaling
  - Cytokine production
  - Immune cell recruitment (e.g., T-cells, macrophages)
Interpretation:
- Tau pathology in the frontal cortex (seen in later AD) is associated with inflammation, which is absent in PART.
- This provides molecular evidence that neuroinflammation is a distinguishing feature of AD, not PART.

🧪 Validation and Predictive Power

TauAge outperformed traditional clocks like the Horvath pan-tissue clock or the brain-specific Cortical clock in predicting tau burden.
It accurately tracked regional pathology and even distinguished:
- Mild Cognitive Impairment (MCI) vs. Dementia.
- Cognitive resilience in Predicted-PART vs. Predicted-AD groups.

💡 Why TauAge Matters

Region-specific insight: First model to quantify tau pathology in specific brain areas using methylation.
Separates aging from disease: Measures pathology beyond what’s expected from age alone.
Molecular divergence of PART vs. AD: Synaptic dysfunction is shared; inflammation is unique to AD.
Biomarker potential: Could be developed into a diagnostic or risk-stratifying tool, especially if adapted to blood or CSF methylation data.
Framework for further modeling: Could be extended with scRNA-seq, single-cell methylomics, or used in AI-based diagnostic pipelines.

Would you like a visual summary or help implementing a TauAge-like model on your own data?

Saturday, March 22, 2025

attack strategies targeting reward functions in reinforcement learning (RL) AI models,

To systematically explore attack strategies targeting reward functions in reinforcement learning (RL) AI models, the following research strategy integrates theoretical analysis, empirical validation, and mitigation development:

1. Problem Definition & Vulnerability Analysis

Objective: Identify how reward function design influences exploitability.
Approach:
- Classify reward functions by type (extrinsic/intrinsic/shaped1), structure (sparse/dense1 7), and optimization goals (cost minimization vs. reward maximization1).
- Analyze attack surfaces:
  - Reward hacking: Exploiting unintended loopholes (e.g., infinite reward loops4).
  - Reward poisoning: Manipulating training data or reward signals to induce malicious policies56.
- Prioritize vulnerabilities in adaptive vs. non-adaptive attacks, noting that adaptive methods can achieve goals faster6.

2. Attack Methodology Development

Objective: Design and benchmark attack vectors.
Approach:
- Poisoning Attacks:
  - Training-time: Inject adversarial perturbations into reward signals (e.g., δₜ in rt + δₜ6).
  - Preference-based: Manipulate human feedback datasets (e.g., label-flipping in RLHF5).
- Exploratory Attacks:
  - Test reward-shaping vulnerabilities (e.g., sparse reward exploitation7).
  - Develop adaptive attacks leveraging agent policy updates during training6.
- Metrics: Measure attack success via policy divergence, time-to-compromise, and stealthiness (e.g., detectability of manipulated rewards5).

3. Empirical Validation

Objective: Evaluate attacks across RL environments.
Testbeds:
- Simulations: Gridworlds, MuJoCo, Atari games.
- Real-world tasks: Autonomous navigation, recommendation systems.
Variables:
- Compare attack efficacy under different reward structures (e.g., dense vs. sparse1 7).
- Test robustness of model-based vs. model-free RL algorithms.

4. Impact Analysis

Objective: Quantify consequences of compromised reward functions.
Key Questions:
- How do poisoned rewards degrade policy performance or safety?
- Can attacks induce catastrophic forgetting or goal hijacking?
- What is the minimum perturbation (δₜ) required for successful attacks6?

5. Defense Mechanisms

Objective: Propose countermeasures.
Strategies:
- Robust Reward Design:
  - Uncertainty-aware reward inference (e.g., MURAL7).
  - Constrained optimization to limit reward manipulation3.
- Detection Systems:
  - Anomaly detection in reward distributions.
  - Adversarial training with perturbed rewards6.
- Formal Guarantees: Certify safety thresholds for reward perturbations6.

6. Ethical & Practical Considerations

Objective: Ensure responsible research practices.
Guidelines:
- Restrict testing to controlled environments.
- Collaborate with AI safety communities to preempt real-world misuse.
- Disclose vulnerabilities to affected frameworks (e.g., OpenAI, DeepMind).

Implementation Timeline

Phase	Duration	Deliverables
1–2	6 months	Taxonomy of vulnerabilities; attack prototypes
3–4	9 months	Benchmark results; impact analysis framework
5–6	12 months	Defense toolkit; ethical guidelines

This strategy balances offensive exploration (to identify risks) and defensive innovation (to mitigate harms), advancing both RL security and robustness.

Citations:

Answer from Perplexity: pplx.ai/share