Mount Sinai AI Decodes Non-Coding DNA to Find 20 Alzheimer's Genes

Learning the Regulatory Code of Alzheimer's Genomes, Towfique Raj, Icahn School of Medicine — Ai4

Ai4

Weighting non-coding variants with microglial deep learning predictions uncovers 20 novel Alzheimer's risk genes that standard genome-wide association studies missed.

AI-assisted summary of Learning the Regulatory Code of Alzheimer's Genomes, Towfique Raj, Icahn School of Medicine — original source. Check the original for full context.

Source published: 2026-08-18

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 18, 2026, this presentation features Professor Towfique Raj from the Icahn School of Medicine at Mount Sinai. The talk describes how researchers decode non-coding regulatory signals in Alzheimer's disease by combining human genetics, microglial biology, eQTL analyses, and sequence-to-function deep learning models to uncover novel therapeutic targets.

## TL;DR Towfique Raj explores the challenge of interpreting Alzheimer's disease genetics, where over 96% of identified risk variants lie within non-coding genomic regions. Leveraging rapid-autopsy human brain tissue, his lab demonstrated that disease heritability is heavily concentrated in the enhancers of microglia, the brain's resident immune cells. By fine-tuning sequence-to-function deep learning models like Enformer and DeepSEA on primary human brain data, the team derived single-nucleotide variant functional delta scores that narrowed linked candidate loci down to causal regulatory disruptions. Furthermore, applying these predictive scores as informative priors in Bayesian rare variant burden testing across 100,000 sequenced genomes uncovered 20 novel Alzheimer's risk genes, alongside an autonomous 'AI co-scientist' framework developed to accelerate discovery.

## Atmosphere & style The presentation carries a focused, rigorous academic tone infused with modest humor, as Raj opens by lightheartedly acknowledging his position as an academic speaker within an industry-dominated AI conference. The delivery is methodically paced and grounded, moving from the sobering global burden of neurodegeneration to precise, data-rich computational methodologies. Listening to the lecture feels like attending a masterclass in modern functional genomics, balancing technical clarity in deep learning architectures with grounded realism regarding wet-lab validation and data constraints.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=4s) Neurodegenerative diseases represent the second leading cause of death globally, with Alzheimer's disease affecting approximately 50 million individuals worldwide and lacking disease-reversing treatments. - [00:04](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=4s) The classic neuropathological hallmarks of Alzheimer's disease include extracellular amyloid-beta plaques, intracellular tau neurofibrillary tangles, severe neuroinflammation, and widespread cortical shrinkage. - [01:40](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=100s) Alzheimer's disease carries an estimated heritability of 60% to 80%, with older African-American and Hispanic-American populations facing a two- to four-fold higher risk. - [01:40](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=100s) Over 96% of GWAS-identified Alzheimer's risk variants reside in non-coding regulatory regions of the genome, creating a significant challenge for inferring downstream disease mechanisms. - [05:15](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=315s) The Alzheimer's Disease Sequencing Project (ADSP) is sequencing up to 100,000 patient genomes to provide the massive multi-omic datasets necessary to decipher non-coding variation. - [05:15](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=315s) GWAS heritability analyses show that Alzheimer's genetic risk is overwhelmingly enriched within the regulatory enhancers of microglia rather than in neurons or astrocytes. - [10:06](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=606s) Fine-tuning sequence-to-function deep learning architectures (Enformer, DeepSEA, ChromBPNet) with primary human brain immune cell data significantly boosts the accuracy of predicting single-nucleotide variant effects on chromatin accessibility and splicing. - [14:28](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=868s) Deep learning-guided fine-mapping successfully narrowed a locus of 542 correlated SNPs down to 4 functional variants, highlighting a causal PU.1 transcription factor binding disruption that regulates the RASGEF1C gene. - [14:28](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=868s) Using deep learning delta scores as functional priors in Bayesian rare variant burden testing across 100,000 sequenced genomes uncovered 20 novel Alzheimer's disease candidate genes. - [18:25](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=1105s) An autonomous 'AI co-scientist' agentic framework developed by Mount Sinai and Stanford integrates over 180 computational tools across 60+ multi-modal biological datasets to drive automated hypothesis generation. - [19:08](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=1148s) The deep learning models and genomic resources developed through this initiative are shared openly with the research community via the Alzheimer's Disease Data Initiative (ADDI) portal.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=4s) The Global Burden and Neuropathological Hallmarks of Neurodegenerative Disease

Opening his talk at the Ai4 conference, Towfique Raj introduces his work as a research professor at the Icahn School of Medicine at Mount Sinai in New York. Acknowledging with a touch of humor that he feels like a "minority here in this conference being an academic," he sets the stage by highlighting the immense worldwide burden of neurodegenerative conditions, which collectively represent the second leading cause of death globally. Alzheimer's disease alone affects approximately 50 million individuals across the globe, while Parkinson's disease impacts roughly 10 million, and Amyotrophic Lateral Sclerosis (ALS) accounts for over half a million cases. Despite recent incremental therapeutic progress for Alzheimer's, these debilitating disorders currently lack a cure or treatments capable of reversing disease progression.

Raj details the classical neuropathological hallmarks observed in the Alzheimer's brain, beginning with the accumulation of extracellular amyloid plaques—sticky clumps formed by amyloid-beta ($\text{A}\beta$) protein fragments—and the intracellular buildup of fibrous tau neurofibrillary tangles within nerve cells. As the disease advances, widespread cellular damage causes neurons to lose structural integrity and functionality, culminating in pronounced brain shrinkage and severe neuroinflammation across affected regions.

- **Global prevalence statistics:** ~50 million individuals living with Alzheimer’s disease; ~10 million with Parkinson’s disease; $\ge 500,000$ with ALS. - **Pathological hallmarks:** Extracellular amyloid-beta ($\text{A}\beta$) plaques, intracellular tau fibrous tangles, neuroinflammation, progressive synaptic/neuronal loss, and gross cortical shrinkage.

---

[01:40](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=100s) Genetic Architecture and the Non-Coding Conundrum in Alzheimer's Disease

Turning to disease etiology, Raj categorizes risk into non-modifiable and modifiable factors. Non-modifiable risks include advancing age, sex (with females exhibiting higher susceptibility), and ancestral background (with older African-American and Hispanic-American populations facing two to four times higher risk). As a human geneticist, Raj focuses heavily on the genetic underpinnings, noting that Alzheimer's disease carries an estimated heritability between 60% and 80%. Modifiable lifestyle and environmental risk factors include hypertension, high cholesterol, obesity, and smoking, whereas educational attainment, regular physical activity, and active social engagement serve as protective buffers.

Reflecting on the past decade of genetic mapping, Raj explains how the field evolved from recognizing *APOE* as the primary major genetic determinant to identifying over 75 distinct genomic regions associated with Alzheimer's risk through genome-wide association studies (GWAS), visually depicted via Manhattan plots resembling city skylines. However, this progress revealed a fundamental biological hurdle: over 96% of identified risk variants reside in non-coding regions of the genome, leaving only roughly 4% directly altering protein-coding sequences. Because these variants do not change amino acid sequences or protein structures directly, researchers cannot easily infer target genes, biological pathways, or downstream mechanisms, turning the non-coding genome into an intricate regulatory puzzle.

```mermaid flowchart TD accTitle: Genetic Architecture of Alzheimer's Disease accDescription: Breakdown of non-coding versus coding genetic risk variants and their functional mechanisms.

subgraph RiskBreakdown["Genetic Risk Architecture"] NC["Non-Coding Variants (~96%)"] PC["Protein-Coding Variants (~4%)"] end

subgraph Mechanisms["Downstream Regulatory Impact"] direction TB TF["Transcription Factor Binding"] CA["Chromatin Accessibility"] AS["Alternative Splicing"] GE["Gene Expression (eQTLs)"] end

NC --> Mechanisms PC -->|"Direct Structural Alteration"| TP["Direct Protein Targets"] ```

- **Heritability:** Alzheimer's disease heritability is estimated at 60%–80%. - **Disparities:** Older African-American and Hispanic populations experience a 2- to 4-fold increase in Alzheimer's risk. - **Non-coding dominance:** Over 96% of GWAS-identified risk variants map to non-coding regulatory sequences; only ~4% fall within protein-coding exons.

---

[05:15](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=315s) Large-Scale Sequencing Initiatives and Microglial Regulatory Enrichment

To unravel the functional impact of non-coding variation, researchers require massive molecular datasets connecting genomic variations to intermediate cellular traits such as chromatin accessibility, gene expression, and epigenetic modifications. Raj highlights his involvement with the Alzheimer's Disease Sequencing Project (ADSP) consortium, supported by the National Institute on Aging (NIA), which has generated over 60,000 whole-genome sequences and is expanding toward 100,000 patient genomes. Yet, Raj emphasizes that without mechanistic functional models, raw sequencing data remains a "black box."

To systematically connect sequence variants to cell biology, Raj's laboratory generates multi-omic profiles from rapidly autopsied human brain tissues, collecting immune cells within three hours post-mortem to measure transcriptomes, proteomes, chromatin architectures, and single-nucleus RNA sequencing (snRNA-seq) dynamics alongside peripheral blood and harmonized bulk tissue cohorts. Strikingly, GWAS heritability enrichment analyses reveal that Alzheimer's disease genetic risk is concentrated overwhelmingly within microglia—the resident immune cells of the central nervous system responsible for plaque clearance, debris removal, and synaptic pruning during development—rather than in neurons. Furthermore, these causal risk variants localize specifically within microglial enhancers, the molecular switches dictating gene activation.

- **Sequencing scale:** ADSP consortium dataset currently standing at 60,000 whole genomes with a target of 100,000 Alzheimer's patient genomes. - **Tissue harvesting window:** Rapid autopsy collection of viable human brain immune cells within 3 hours post-mortem. - **Cell-type specificity:** GWAS heritability strongly enriches in microglia enhancers rather than neuronal or astrocytic regulatory domains, driving biopharma attention toward neuroimmune targets.

---

[10:06](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=606s) Applying Sequence-to-Function Deep Learning Models to Brain Regulatory Data

Bridging genomics with artificial intelligence, Raj introduces sequence-to-function deep learning architectures—such as Enformer and DeepSEA—which take long genomic context windows (up to 1 megabase) and employ transformer and convolutional layers to predict multimodal regulatory effects across individual nucleotides. However, baseline sequence models often struggle when trained on generic or non-human data. To overcome this limitation, Raj's lab collaborated with computational biologist David Knowles to fine-tune sequence models using their high-quality, primary human microglial, neuronal, and astrocytic datasets.

By fine-tuning models like DeepSEA, Enformer, and ChromBPNet with human brain cell data, the team generated single-nucleotide variant effect prediction scores ("delta scores") across 50,000 whole genomes. These scores quantify whether a specific mutation alters chromatin accessibility, transcription factor binding, or RNA splicing. The fine-tuned models demonstrated clear accuracy gains (improved ROC-AUC) over baseline architectures, showing strong correlation between predicted delta scores and empirical splicing measures collected from human post-mortem brain subjects.

```mermaid flowchart LR accTitle: Sequence-to-Function Deep Learning Pipeline accDescription: Workflow showing genomic sequences fine-tuned on primary human microglial data to generate functional delta scores.

subgraph InputData["Data Inputs"] DNA["Raw Genomic Sequence (Up to 1 Mb)"] MG["Primary Human Microglia Datasets"] end

subgraph Modeling["Deep Learning Framework"] FT["Fine-Tuned Enformer / DeepSEA / ChromBPNet"] end

subgraph Outputs["Functional Output"] DS["Variant Delta Scores (Chromatin, Splicing, Binding)"] VAL["CRISPR / Massively Parallel Reporter Assays"] end

DNA --> FT MG -->|"Fine-Tuning"| FT FT --> DS DS --> VAL ```

- **Model capabilities:** Sequence-to-function models ingest up to 1 Mb of raw sequence to predict epigenetic, splicing, and expression outcomes. - **Fine-tuning strategy:** Infusing cell-type-specific chromatin accessibility and transcriptome data from primary human microglia, astrocytes, and neurons noticeably boosts predictive performance. - **Empirical validation:** Variant delta scores show measurable concordance with experimentally observed chromatin accessibility shifts and alternative splicing events in patient cohorts.

---

[14:28](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=868s) Fine-Mapping Causal Variants and Bayesian Rare Variant Burden Testing

A major dilemma in human genetics is resolving which specific variant drives disease when dozens of single-nucleotide polymorphisms (SNPs) reside on the same haplotype block in high linkage disequilibrium. Raj illustrates this challenge at a locus containing 542 tightly linked candidate SNPs. Utilizing their deep learning predictive pipeline, the team narrowed the field down to four prioritized functional variants. One of these candidate variants directly disrupts a PU.1 transcription factor binding motif, altering the regulation of the *RASGEF1C* gene. The team subsequently confirmed this computational prediction through experimental validation using Massively Parallel Reporter Assays (MPRA) and CRISPR-Cas9 genome editing in human induced pluripotent stem cell (iPSC)-derived microglia.

Beyond individual variant fine-mapping, Raj's group leverages these delta scores as informative priors in Bayesian rare variant burden testing. Traditional case-control burden tests weight all observed rare variants uniformly. By weighting variants according to their deep learning-predicted functional scores across 50,000 cases and 50,000 controls, their framework substantially elevated statistical power, successfully uncovering 20 novel Alzheimer's disease risk genes that had evaded standard GWAS approaches. Raj cautions, however, that deep learning predictions remain limited in rare, underrepresented cell types where training data is sparse, underscoring that targeted experimental validation remains essential.

- **Fine-mapping resolution:** Reduced a dense locus of 542 correlated SNPs down to 4 putative causal variants, pinpointing a disruption in a PU.1 binding site regulating *RASGEF1C*. - **Validation toolset:** Massively Parallel Reporter Assays (MPRA) and CRISPR-Cas9 editing in iPSC-derived microglia models. - **Bayesian burden testing discovery:** Identified 20 novel Alzheimer's disease candidate genes across 100,000 sequenced genomes (50,000 cases vs. 50,000 controls) by weighting rare non-coding alleles with functional deep learning priors.

---

[18:25](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=1105s) AI Co-Scientist and Autonomous Discovery for Alzheimer's Research

In the final segment of this portion, Raj highlights an institutional collaboration between the Icahn School of Medicine at Mount Sinai and Stanford University (working alongside Kang Wang and collaborators). The multi-disciplinary team developed an autonomous "AI co-scientist" agentic framework tailored specifically for neurodegenerative research. This agentic AI system was recognized as an award recipient in the Alzheimer's Insights AI Prize competition. To drive autonomous hypothesis generation and evidence synthesis, the platform integrates comprehensive multi-modal Alzheimer's datasets—spanning brain imaging, cellular assays, single-cell transcriptomics, and genomic registries—connecting over 180 analytical tools across more than 60 harmonized biological data resources.

- **Collaborating institutions:** Icahn School of Medicine at Mount Sinai and Stanford University. - **Recognition:** Awarded the Alzheimer's Insights AI Prize for agentic scientific reasoning. - **System scope:** Integrates over 180 computational tools across 60+ curated multi-omic and neuroimaging datasets to support automated research hypothesis generation.

[19:08](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=1148s) Benchmarking Against Baseline LLMs, Data Availability, and Concluding Remarks

Towfique Raj concludes his presentation by highlighting the superior performance of their deep learning approach when evaluated on real-world tasks and benchmarked against standard baseline large language models (LLMs). To ensure broad accessibility for the scientific community, the developed models and resources are being made available for download directly through the Alzheimer's Disease Data Initiative (ADDI) portal.

Bringing the talk to a close, Raj transitions to his acknowledgments slide to express gratitude to the members of his research laboratory for their contributions to the project. He thanks the audience for their time, letting attendees know they are welcome to head out while inviting anyone with lingering questions to stay and speak with him directly: "You guys can go, but I'll take questions after my talk if anyone has any. Thank you."

- **Model Performance:** Demonstrated superior capabilities on real-world genomic tasks compared to baseline LLMs. - **Resource Access:** Tools and datasets are hosted on and downloadable via the Alzheimer's Disease Data Initiative portal. - **Acknowledgments:** Recognition extended to the Raj Lab members at the Icahn School of Medicine at Mount Sinai.

## Notable quotes > Most genetic changes linked to Alzheimer’s disease do not alter proteins directly, but instead affect how genes are turned on and off. — Towfique Raj

> In this talk, I will explain how we decode these hidden regulatory signals by combining human genetics with studies of microglia, the brain’s resident immune cells that help protect and maintain brain health. — Towfique Raj

> [00:04](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=4s) minority here in this conference being an academic — Towfique Raj

> I will describe how we use eQTL analyses, which link DNA differences to changes in gene activity, to understand how risk variants influence how genes work inside these cells. — Towfique Raj

> [05:15](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=315s) black box — Towfique Raj

> I will also show how deep learning models can predict how small DNA sequence changes affect gene function and disease risk. — Towfique Raj

> Attendees will gain a clear framework for how genetic risk shapes brain immune cell behavior and contributes to Alzheimer’s disease, and how these insights can guide new therapeutic targets. — Towfique Raj

> [19:08](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=1148s) You guys can go, but I'll take questions after my talk if anyone has any. Thank you. — Towfique Raj

## Mentioned Towfique Raj, David Knowles, Kang Wang, Icahn School of Medicine at Mount Sinai, Stanford University, Ai4, National Institute on Aging (NIA), Alzheimer's Disease Sequencing Project (ADSP), Alzheimer's Disease Data Initiative (ADDI), Alzheimer's Insights AI Prize, Enformer, DeepSEA, ChromBPNet, CRISPR-Cas9, Massively Parallel Reporter Assays (MPRA), APOE, RASGEF1C, PU.1, Amyloid-beta, Tau

## Video description Towfique Raj - Professor, Icahn School of Medicine at Mount Sinai

Most genetic changes linked to Alzheimer’s disease do not alter proteins directly, but instead affect how genes are turned on and off. In this talk, I will explain how we decode these hidden regulatory signals by combining human genetics with studies of microglia, the brain’s resident immune cells that help protect and maintain brain health. I will describe how we use eQTL analyses, which link DNA differences to changes in gene activity, to understand how risk variants influence how genes work inside these cells. I will also show how deep learning models can predict how small DNA sequence changes affect gene function and disease risk. Attendees will gain a clear framework for how genetic risk shapes brain immune cell behavior and contributes to Alzheimer’s disease, and how these insights can guide new therapeutic targets.

[Source](https://www.youtube.com/watch?v=6D3RLbcTQJU)

中文简报

芒特西奈AI解码非编码DNA发现20个阿尔茨海默病风险基因通过小胶质细胞深度学习预测对非编码变异进行加权，发现了常规全基因组关联分析所遗漏的20个阿尔茨海默病新型风险基因。 Towfique Raj探讨了解释阿尔茨海默病遗传学面临的挑战，即超过96%的已知风险变异位于非编码基因组区域。利用快速尸检人脑组织，其实验室证实了该疾病的遗传力高度集中在大脑常驻免疫细胞——小胶质细胞的增强子中。通过在原代人脑数据上微调如Enformer和DeepSEA等序列到功能的深度学习模型，该团队计算出了单核苷酸变异功能差值评分，从而将连锁候选基因座缩小至因果调控紊乱变异。此外，在对10万个测序基因组进行的贝叶斯罕见变异负荷检验中，研究人员将这些预测评分作为信息先验，发现了20个新型阿尔茨海默病风险基因，并构建了自主“AI联合科学家”（AI co-scientist）框架以加速科学发现。神经退行性疾病是全球第二大死因，其中阿尔茨海默病影响了全球约5000万人，且目前仍缺乏逆转病程的治疗方法。 (0:04) 阿尔茨海默病的经典神经病理学特征包括细胞外β-淀粉样蛋白斑块、细胞内tau神经原纤维缠结、严重的神经炎症以及广泛的皮层萎缩。 (0:04) 阿尔茨海默病的估计遗传力为60%至80%，其中非裔和西裔美国老年人群面临高出两到四倍的患病风险。 (1:40) GWAS识别出的阿尔茨海默病风险变异中，超过96%位于基因组的非编码调控区，这给推断下游疾病机制带来了巨大挑战。 (1:40) 阿尔茨海默病测序计划（ADSP）正在对多达10万个患者基因组进行测序，以提供破译非编码变异所需的海量多组学数据集。 (5:15)

Reading summary

Human geneticist Towfique Raj argues that deciphering Alzheimer's disease pathology requires decoding the non-coding genome by coupling rapid-autopsy brain multi-omics with cell-type-specific deep learning models. By demonstrating that genetic risk predominantly concentrates in microglial enhancers, his research establishes computational variant scoring as a viable method to resolve causal variants and uncover novel risk genes. Non-Coding Architecture and Microglial Enhancer Enrichment (1:40) Alzheimer's disease affects approximately 50 million individuals worldwide and carries an estimated heritability between 60% and 80%, with older African-American and Hispanic populations experiencing a two- to fourfold increased risk. Although genome-wide association studies (GWAS) have expanded the number of identified risk loci from APOE alone to over 75 regions, more than 96% of these variants reside in non-coding regions. Because only about 4% alter protein-coding sequences, the majority act indirectly through regulatory mechanisms such as transcription factor binding, chromatin accessibility, and alternative splicing. To translate sequence data into biological mechanisms, Raj's laboratory analyzes multi-omic profiles from post-mortem brain tissues harvested within three hours of death, alongside datasets from the National Institute on Aging's Alzheimer's Disease Sequencing Project (ADSP), which encompasses 60,000 whole genomes and aims for 100,000. Heritability enrichment analyses show that Alzheimer's genetic risk is concentrated overwhelmingly within the enhancers of microglia—the central nervous system's resident immune cells—rather than in neurons or astrocytes. Fine-Tuning Sequence-to-Function Models on Brain Data (10:06) To interpret non-coding sequences across windows up to one megabase, Raj's team, in collaboration with computational biologist David Knowles, adapted sequence-to-function deep learning architectures such as Enformer, DeepSEA, and ChromBPNet. Because baseline models trained on generic genomic datasets exhibit limitations in human brain tissue, the researchers fine-tuned these architectures using primary human microglial, neuronal, and astrocytic datasets. Applying these fine-tuned models to 50,000 whole genomes generated single-nucleotide variant effect prediction scores, termed delta scores. These scores quantify how individual mutations influence chromatin accessibility, transcription factor binding, and RNA splicing. Raj reports that the brain-specific fine-tuning improved model accuracy (measured by ROC-AUC) over baseline architectures, showing measurable concordance with empirical splicing events observed in post-mortem patient brain tissues. Fine-Mapping Causal Variants and Rare Burden Discovery (14:28) Resolving causal alleles within dense haplotype blocks in high linkage disequilibrium remains a major barrier in human genetics. At one genomic locus containing 542 correlated single-nucleotide polymorphisms, Raj's predictive pipeline narrowed the candidate pool down to four putative functional variants. One prioritized variant was predicted to disrupt a PU.1 transcription factor binding site that regulates the RASGEF1C gene—a mechanism the team experimentally validated using Massively Parallel Reporter Assays (MPRA) and CRISPR-Cas9 genome editing in human induced pluripotent stem cell (iPSC)-derived microglia. Beyond single-locus fine-mapping, the team integrated these functional delta scores as Bayesian priors in rare variant burden testing across 100,000 sequenced genomes (50,000 cases and 50,000 controls). Weighting rare variants by predicted functional severity rather than treating them uniformly substantially elevated statistical power, leading to the identification of 20 novel Alzheimer's disease candidate genes that had been missed by standard GWAS approaches. Autonomous AI Co-Scientists and Community Data Access (18:25) Extending computational genomics into automated discovery, Raj highlighted a multi-disciplinary collaboration between the Icahn School of Medicine at Mount Sinai and Stanford University, alongside Kang Wang. The team engineered an agentic "AI co-scientist" framework recognized by the Alzheimer's Insights AI Prize. Designed to generate scientific hypotheses and synthesize biological evidence, the platform links over 180 analytical tools across more than 60 harmonized datasets spanning neuroimaging, single-cell transcriptomics, and genomic registries. Raj reports that their specialized deep learning pipelines demonstrate superior capabilities on real-world genomic tasks compared to baseline large language models. To facilitate community validation and downstream research, the lab's models and functional datasets are being hosted for public access via the Alzheimer's Disease Data Initiative (ADDI) portal. Raj cautions that sequence-to-function deep learning predictions face clear constraints in rare or underrepresented brain cell types where training data remains sparse, maintaining that computational scoring must be paired with experimental validation such as CRISPR editing and reporter assays to reliably confirm causal neuroimmune mechanisms.

精炼阅读摘要

西奈山伊坎医学院的 Towfique Raj 教授指出，阿尔茨海默病高达 60%–80% 的遗传风险集中在非编码调控区，且主要富集于小胶质细胞的增强子网络；通过利用原代人脑细胞多组学数据微调“序列到功能”深度学习模型，研究团队能够破解非编码基因组的调控机制、精细定位因果变异，并挖掘出传统全基因组关联分析无法发现的全新致病基因。遗传架构难题与非编码区调控困境 (1:40) 阿尔茨海默病在全球影响约 5000 万人，是全球第二大死因神经退行性疾病中的主要病理类型，临床伴随细胞外淀粉样蛋白（Aβ）斑块沉积、细胞内 tau 蛋白纤维缠结、神经炎症与严重脑皮质萎缩，且目前尚无能够逆转病程的治愈方案。Raj 强调，该病具有不可控风险（如高龄、女性高发以及非裔和西裔人群面临的 2 至 4 倍更高风险）与可控风险因素，其遗传度估算在 60% 至 80% 之间。尽管过去十余年的全基因组关联研究（GWAS）已从单一聚焦 APOE 扩展到锁定超过 75 个风险基因座，但研究界遭遇了重大生物学障碍：超过 96% 的已知风险变异位点位于基因组的非编码调控区域，仅约 4% 会直接改变蛋白质编码序列。由于绝大多数变异不直接改变氨基酸序列或蛋白质结构，研究人员极难直接推断变异对应的靶标基因、下游生物学通路及分子机制，使非编码基因组成为复杂的调控谜题。小胶质细胞增强子富集与多组学图谱 (5:15) 为了系统阐明非编码变异的功能影响，阿尔茨海默病测序计划（ADSP）已建立超过 6 万个全基因组测序数据集，并正向 10 万个患者基因组的目标推进。然而 Raj 指出，缺乏机制性功能模型的原始基因组测序数据本质上仍是“黑箱”。为此，Raj 实验室通过快速尸检通道，在患者死亡后 3 小时内获取高活性人脑免疫细胞，系统构建涵盖转录组、蛋白质组、染色质可及性及单核 RNA 测序（snRNA-seq）的多组学图谱。遗传度富集分析表明，阿尔茨海默病的遗传风险并非均匀分布，而是压倒性地富集于小胶质细胞（负责斑块清除与突触修剪的中枢神经系统免疫细胞），而非神经元或星形胶质细胞；更关键的是，这些因果变异高度集中在控制基因激活的小胶质细胞增强子区域，直接推动了生物医药领域对神经免疫靶点的关注。深度学习变异预测与因果位点精细定位 (10:06) 针对长达 1 Mb 基因组窗口预测分子调控效应的“序列到功能”深度学习模型（如 Enformer、DeepSEA 和 ChromBPNet），通常因通用或非人源训练数据而在人脑组织预测中受限。Raj 团队与计算生物学家 David Knowles 合作，利用原代人类小胶质细胞、神经元和星形胶质细胞的高质量数据对模型展开微调，在 5 万个全基因组上生成了单核苷酸变异功能效应评分（delta scores）。该评分量化了突变对染色质可及性、转录因子结合及 RNA 剪接的影响，预测结果与人脑尸检队列中测得的剪接事件呈现显著一致性。在因果位点定位方面，针对单倍型块内因高连锁不平衡而难以区分的候选变异，Raj 以一个包含 542 个紧密连锁候选 SNP 的位点为例，说明该深度学习流程将候选范围精准收窄至 4 个功能变异，并确定了其中一个直接破坏 PU.1 转录因子结合基序以调控 RASGEF1C 基因的致病变异。该预测随后在人诱导多能干细胞（iPSC）分化的小胶质细胞模型中，通过大规模并行报告基因测序（MPRA）和 CRISPR-Cas9 基因编辑获得了实验验证。贝叶斯负荷检验新发现与自主智能体体系 (14:28) 除了单个变异解析，Raj 团队还将深度学习预测出的功能先验整合到贝叶斯罕见变异负荷检验中。不同于传统对所有罕见变异赋予均等权重的病例-对照检验方法，该框架在 5 万例患者与 5 万例对照（共 10 万个全基因组）中依据功能评分赋权，大幅提升了统计检验功效，成功发掘出 20 个传统 GWAS 无法识别的新型阿尔茨海默病候选风险基因。在此基础上，西奈山伊坎医学院与斯坦福大学团队（包括 Kang Wang 等合作者）进一步构建了面向神经退行性疾病的自主“AI 科学助手”（AI co-scientist）智能体系统，并荣获 Alzheimer's Insights AI 大奖。该系统连接了超过 180 种分析工具与 60 余个涵盖脑影像、细胞实验、单细胞转录组及基因组的多模态生物数据集，旨在实现科学假设的自主生成与证据综合，且在实际基准测试中展现出优于通用大语言模型（LLM）的性能。 Raj 指出，团队开发的相关模型与数据资源已通过阿尔茨海默病数据计划（ADDI）门户向科学界公开发布。不过他也提示了当前研究的局限与未解问题：深度学习模型的预测能力在训练数据稀疏的罕见或低丰度细胞类型中依然受限，纯计算预测无法完全取代针对性实验验证，利用人源 iPSC 衍生细胞进行功能验证仍是明确致病机制的必要步骤。

Detailed notes — zh

## Background 该视频由频道 Ai4 于 2026 年 8 月 18 日在 YouTube 平台发布，主讲人为西奈山伊坎医学院（Icahn School of Medicine at Mount Sinai）教授 Towfique Raj。页面简介表明，阿尔茨海默病绝大多数遗传风险变异影响的是基因表达调控而非直接改变蛋白质，本次演讲主要探讨如何通过结合人类遗传学、大脑常驻免疫细胞（小胶质细胞）研究、eQTL 分析以及深度学习模型来破译非编码调控密码，进而揭示疾病致病机制并指导开发新药物靶点。 ## TL;DR 阿尔茨海默病具有 60% 至 80% 的高遗传力，但 GWAS 揭示的超过 96% 的关联变异均位于非编码调控区，构成了机制解析的核心瓶颈。研究发现这些遗传风险位点高度富集在大脑常驻免疫细胞（小胶质细胞）特异性的增强子调控元件中。为此，研究团队利用小胶质细胞多组学数据微调了 Enformer 等序列到功能的深度学习模型，有效预测了单核苷酸变异对染色质与表达的调控效应，并成功在连锁单倍型中精确定位出经 CRISPR 和 MPRA 实验验证的因果变异。此外，团队结合深度学习先验权重的贝叶斯负荷测试新发现了 20 个阿尔茨海默病风险基因，并推出了整合 180 多种分析工具的 AI 科研智能体系统以加速神经退行性疾病的研究。 ## Atmosphere & style 整个演讲呈现出严谨扎实且极具前沿探索感的计算生物学学术氛围，主讲人叙述清晰流畅、条理分明，以问题为导向层层拆解非编码基因组的“黑盒”难题。报告将海量全基因组大数据、深度学习 Transformer 架构与干湿实验验证融为一体，展现出 AI 赋能复杂疾病遗传机制破译所带来的突破性进展。尾声阶段对开源数据共享的强调以及开放轻松的问答互动，传递出务实、严谨且充满活力的跨学科科研协作精神。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=4s) 全球神经退行性疾病负担沉重，阿尔茨海默病（AD）遗传力高达 60% 至 80%，但目前尚无根治手段或逆转病程的有效疗法。 - [02:30](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=150s) 全基因组关联分析（GWAS）已发现超过 75 个 AD 风险位点，但超过 96% 的变异落在非编码区，构成了机制解析的核心瓶颈。 - [02:30](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=150s) 由于非编码区机制极其复杂，传统方法难以直接判断风险变异具体调控哪个下游基因以及影响何种细胞内分子过程。 - [05:03](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=303s) ADSP 联盟正在开展多达 10 万例的大规模全基因组测序，为破解非编码变异的生物学功能提供了海量数据基础。 - [05:03](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=303s) 遗传力富集分析表明，阿尔茨海默病的风险变异高度富集在大脑常驻免疫细胞——小胶质细胞特异的增强子等调控区，而非神经元本身。 - [10:06](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=606s) 引入 Transformer 架构的序列到功能深度学习模型，能够基于最高 1 Mb 的原始 DNA 序列预测变异对染色质和表达的多模态影响。 - [10:06](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=606s) 利用人类小胶质细胞与脑细胞真实多组学数据对通用模型进行微调，显著拉开了功能性变异与非功能性变异的预测区分度。 - [14:24](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=864s) 深度学习模型成功破解连锁不平衡的筛选难题，可将包含数百个 SNP 的区域精准收窄至少数几个高置信度因果变异，并通过 MPRA 和 CRISPR 验证。 - [14:24](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=864s) 将深度学习功能分数作为先验权重纳入贝叶斯罕见变异负荷检验，在 10 万人大规模队列中成功新发现了 20 个阿尔茨海默病风险基因。 - [18:25](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=1105s) 研发整合了 180 多种分析工具与 60 多个专业多模态数据集的科研智能体（AI Co-scientist），自主辅助科学家探索复杂疾病机制。 - [19:08](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=1148s) 该研究所构建的工具与模型性能优于基线 LLM，并将通过阿尔茨海默病数据计划门户（ADDI portal）向全球科研共同体开放共享。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=4s) 神经退行性疾病的全球负担与阿尔茨海默病的病理特征神经退行性疾病已成为全球第二大死亡原因。全球约有 5000 万阿尔茨海默病（Alzheimer's disease, AD）患者、1000 万帕金森病（Parkinson's disease）患者以及超过 50 万肌萎缩侧索硬化（ALS）患者。令人遗憾的是，目前这类疾病尚无根治手段，也缺乏能够逆转疾病进程的有效疗法（尽管近期阿尔茨海默病领域取得了一些进展）。从病理特征来看，阿尔茨海默病大脑中存在两大核心标志：一是由 $\beta$-淀粉样蛋白（A$\beta$）聚集形成的黏性斑块，二是神经元内部由 Tau 蛋白纤维化聚集形成的神经原纤维缠结。伴随这些病理改变的是广泛的细胞损伤、大脑萎缩（脑萎缩）以及神经炎症反应。在风险因素方面，不可改变的因素包括年龄（发病率随年龄增长显著上升）、性别（女性发病风险更高）和族裔（高龄非裔与拉美裔美国人的患病风险是其他人群的 2 到 4 倍）。而遗传因素扮演着至关重要的角色，阿尔茨海默病的遗传力（heritability）高达 60% 至 80%。可调节的风险因素则包括高血压、高胆固醇、肥胖、环境暴露和吸烟；受教育程度、体力活动及社交接触则属于保护性因素。 - 全球神经退行性疾病流行现状：AD 约 5000 万人，帕金森病约 1000 万人，ALS 约 50 万人以上。 - 阿尔茨海默病遗传力估算在 60%–80% 之间。 --- ### [02:30](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=150s) 从 GWAS 到非编码区黑盒：遗传学机制解析的困境人类遗传学在过去十年中取得了巨大突破。过去人们主要熟知 *APOE* 这一最为显著的阿尔茨海默病遗传风险因子，而如今全基因组关联分析（GWAS）的曼哈顿图（Manhattan plot）已揭示了基因组中超过 75 个与该病风险相关的位点。然而，这一系列进展也带来了一个巨大的生物学挑战：人群中绝大多数常见的疾病相关单核苷酸多态性（SNP）变异落在**非编码区（Non-coding regions）**。超过 96% 的风险变异属于非编码变异，仅有约 4% 直接映射到蛋白编码区。过去科学界曾认为只要找到编码区的突变，就能明确靶点基因并开发对应药物，但非编码基因组极其复杂，无法直接读懂其下游功能。研究人员难以判断这些非编码变异究竟在调控哪个基因、影响了哪些细胞内分子机制（如染色质可及性、基因表达、蛋白质水平、DNA 甲基化或表观遗传学变化）。 ```mermaid flowchart TD accTitle: GWAS 风险变异的分布与机制解析困境 subgraph GWAS["GWAS 发现 (>75 个基因组风险区域)"] V["疾病关联变异 (SNPs)"] end subgraph NonCoding["非编码变异 (>96%)"] Reg["调控元件 / 增强子 (Enhancers)"] Unknown["下游机制未知: 调控哪个基因？影响哪种细胞？"] end subgraph Coding["蛋白编码变异 (~4%)"] Prot["直接改变氨基酸 / 蛋白质结构 (清晰靶点)"] end V -->|96%| NonCoding V -->|4%| Coding ``` - 阿尔茨海默病全基因组风险区域现已发现超过 75 个。 - 超过 96% 的关联变异位于非编码区，仅约 4% 位于蛋白编码区。 - 非编码变异在细胞层面的功能影响（染色质状态、剪接、表达量等）构成了理解疾病机制的主要瓶颈。 --- ### [05:03](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=303s) 全基因组测序浪潮与小胶质细胞的遗传富集由美国国家老龄化研究所（National Institute on Aging, NIA）主导的阿尔茨海默病测序项目（ADSP）联盟，目前已生成 6 万个全基因组测序数据，并计划扩展到 10 万个全基因组。面对如此海量的数据，仅靠传统的 GWAS 无法破解变异背后的生物学机制。尽管通过患者血液、脑脊液（CSF）或尸检脑组织建立经验型分子关联（如 eQTL）行之有效，但在解析大脑中神经元、星形胶质细胞和小胶质细胞等特定细胞类型时，获取样本和数据的难度极大。遗传学分析发现了一个关键现象：阿尔茨海默病的遗传风险变异高度富集在大脑常驻免疫细胞——**小胶质细胞（Microglia）**中，而非神经元。小胶质细胞负责清除细胞碎片与斑块、参与突触修剪并与神经元相互作用。分析表明，绝大多数 AD 遗传变异集中在小胶质细胞特异性的**增强子（Enhancers）**等调控区域，这一发现促使学术界和制药行业将目光聚焦于神经炎症与小胶质细胞相关的全新药物靶点。 - ADSP 联盟已完成 60,000 个全基因组测序，目标扩展至 100,000 个。 - 尸检脑组织免疫细胞在死亡后 3 小时内采集并分离，用于构建转录组、蛋白质组与染色质图谱。 - 遗传力富集分析证明：AD 风险位点绝大部分高度集中在小胶质细胞特异的调控元件（如增强子）中。 --- ### [10:06](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=606s) 引入序列到功能深度学习模型与微调优化为解决序列功能预测问题，深度学习“序列到功能”（Sequence-to-function）模型（如 DeepMind 的 AlphaGenome、DeepSEA 以及 Enformer）被引入基因组学。这类模型可接收长达 1 Mb 的原始 DNA 序列，通过 Transformer 捕捉长程基因相互作用，预测单个核苷酸变异对染色质可及性、RNA 剪接、转录因子结合及基因表达的多模态影响。但若缺乏疾病相关的特定组织与细胞数据，通用模型的预测能力将严重受限。该实验室与计算生物学家 David Knowles 展开合作，利用从小胶质细胞、神经元和星形胶质细胞中收集的真实人类调控数据集，对 Enformer、DeepSEA 等模型进行了微调（Fine-tuning）。团队计算了 50,000 个全基因组中每个位点的变异效应预测差值（Delta score）。在染色质可及性预测基准测试中（使用 ChromBPNet 与 Enformer 架构），微调后的模型显著拉开了功能性变异与非功能性变异的分布差异，预测的 Delta 分数与经验证的真实染色质改变及 RNA 剪接效应表现出良好的一致性与相关性。 ```mermaid flowchart LR accTitle: 深度学习变异效应预测与微调流程 subgraph Data["输入数据"] DNA["原始 DNA 序列 (最高 1 Mb)"] Microglia["人类小胶质细胞多组学数据 (微调)"] end subgraph Model["深度学习模型"] Arch["Transformer 架构 (Enformer / DeepSEA)"] end subgraph Output["预测输出"] Delta["单核苷酸 Delta Score (剪接 / 染色质 / 表达)"] Downstream["因果变异筛选 & 贝叶斯负荷测试"] end DNA --> Arch Microglia -->|Fine-tune| Arch Arch --> Delta Delta --> Downstream ``` - 模型可评估长达 1 Mb 序列的长程相互作用。 - 针对 50,000 个全基因组中的单核苷酸变异逐一计算 Delta score。 - 融入小胶质细胞、神经元和星形胶质细胞数据集微调后，ROC 曲线下面积（AUC）与预测精度均获得提升。 --- ### [14:24](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=864s) 因果变异优先级排序、实验验证与新型致病基因发现由于连锁不平衡（Linkage Disequilibrium）的存在，GWAS 鉴定出的候选变异往往处于同一单倍型上，统计学上难以区分谁才是真正的因果变异（Causal variant）。深度学习模型能在此发挥关键筛选作用： 1. **因果变异精准定位**：在某个包含 542 个 SNP 的候选关联区域中，模型成功将范围缩小至 4 个具有高功能预测分数的变异。其中一个关键变异位于转录因子 SPI1（PU.1）的结合位点，该变异会破坏转录因子结合并调控下游基因 *RASSF1C*（或相关转录本）的表达。团队随后利用大规模并行报告基因测定（MPRA）以及 iPSC 分化的小胶质细胞进行 CRISPR-Cas9 基因编辑，实验验证了该变异的功能真实性。 2. **贝叶斯稀有变异负荷测试（Bayesian Rare Variant Burden Test）**：传统病例-对照关联分析将所有变异赋予相同权重。团队将深度学习得出的功能预测分数作为先验权重（Priors），在 50,000 例病例与 50,000 例对照的大规模队列中执行罕见变异负荷检验，成功新发现了 20 个此前未被报道的阿尔茨海默病风险基因。 > "we were able to narrow down to four variants, one of which which sits on a transcription factor binding... and show that these are the functional variant and that's a really tremendous progress" - 从 542 个连锁 SNP 中直接筛选出 4 个高置信度因果变异。 - 实验验证路径：计算预测 $\rightarrow$ MPRA / CRISPR-Cas9 编辑 $\rightarrow$ iPSC 衍生小胶质细胞表型验证。 - 引入深度学习先验权重的贝叶斯负荷测试，新鉴定出 20 个阿尔茨海默病风险基因。 --- ### [18:25](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=1105s) AI 科学家智能体协作系统在演讲的最后阶段，主讲人介绍了西奈山团队与斯坦福大学（James Zou / Rex Ying 实验室等合作者）共同开发的项目——AI 科学副驾驶/科研智能体（AI Co-scientist / Agentic AI）。该项目整合了 180 多种分析工具与 60 多个专业数据集（涵盖阿尔茨海默病相关的病理成像、单细胞 RNA 测序、细胞生物学过程等多维度数据），并荣获了 Alzheimer's Insights AI 奖项。该智能体系统旨在通过自主调用工具链与数据挖掘，辅助科学家加速神经退行性疾病复杂机制的探索。 - 整合超过 180 种生物信息学工具与 60 多个阿尔茨海默病多模态数据集。 - 涵盖脑组织病理成像、单细胞/单核 RNA 测序及细胞过程数据。 - 获得 Alzheimer's Insights AI 奖（Agentic AI 赛道）。 ### [19:08](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=1148s) 模型性能评估、数据平台开放与致谢问答在对模型进行综合评估时，该模型展现出了更优异的性能，尤其是在应对真实的现实世界任务（real-world tasks）时，其表现明显优于基线大语言模型（baseline LLMs）。为了促进科学共同体的合作与进一步探索，这一研究成果及相关工具将通过阿尔茨海默病数据计划门户网站（Alzheimer's Disease Data Initiative / ADDI portal）正式对外开放，研究人员届时可以直接访问并下载使用。在演讲的尾声，Towfique Raj 教授展示了致谢幻灯片，向其实验室全体研究团队成员的付出以及现场听众的聆听表达了感谢。随着演讲正式结束，他轻松地对大家说：“你们可以离开了，不过如果有人有任何疑问，我会在演讲结束后继续留下来回答大家的问题（I'll take questions after my talk if anyone has any）。” - **模型优势**：在真实应用任务上的实际表现显著优于基线 LLM 模型。 - **资源共享**：相关工具与数据模型将部署于阿尔茨海默病数据计划门户（Alzheimer's Disease Data Initiative portal），供全球研究人员下载与使用。 - **致谢与互动**：向其实验室团队致谢，并开放现场提问与讨论环节。 ## Notable quotes > Most genetic changes linked to Alzheimer’s disease do not alter proteins directly, but instead affect how genes are turned on and off. — Towfique Raj > In this talk, I will explain how we decode these hidden regulatory signals by combining human genetics with studies of microglia, the brain’s resident immune cells that help protect and maintain brain health. — Towfique Raj > I will describe how we use eQTL analyses, which link DNA differences to changes in gene activity, to understand how risk variants influence how genes work inside these cells. — Towfique Raj > I will also show how deep learning models can predict how small DNA sequence changes affect gene function and disease risk. — Towfique Raj > [14:24](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=864s) we were able to narrow down to four variants, one of which which sits on a transcription factor binding... and show that these are the functional variant and that's a really tremendous progress — Towfique Raj > [19:08](https://www.youtube.com/watch?v=6D3RLbcTQJU&t=1148s) I'll take questions after my talk if anyone has any — Towfique Raj ## Mentioned Towfique Raj, 西奈山伊坎医学院 (Icahn School of Medicine at Mount Sinai), Ai4, 美国国家老龄化研究所 (National Institute on Aging, NIA), 阿尔茨海默病测序项目联盟 (ADSP), 阿尔茨海默病数据计划门户 (ADDI Portal), David Knowles, James Zou, Rex Ying, 斯坦福大学 (Stanford University), DeepMind, AlphaGenome, DeepSEA, Enformer, ChromBPNet, CRISPR-Cas9, SPI1 (PU.1), APOE, RASSF1C ## Video description Towfique Raj - Professor, Icahn School of Medicine at Mount Sinai Most genetic changes linked to Alzheimer’s disease do not alter proteins directly, but instead affect how genes are turned on and off. In this talk, I will explain how we decode these hidden regulatory signals by combining human genetics with studies of microglia, the brain’s resident immune cells that help protect and maintain brain health. I will describe how we use eQTL analyses, which link DNA differences to changes in gene activity, to understand how risk variants influence how genes work inside these cells. I will also show how deep learning models can predict how small DNA sequence changes affect gene function and disease risk. Attendees will gain a clear framework for how genetic risk shapes brain immune cell behavior and contributes to Alzheimer’s disease, and how these insights can guide new therapeutic targets. [Source](https://www.youtube.com/watch?v=6D3RLbcTQJU)

Source quotes

Towfique Raj: Most genetic changes linked to Alzheimer’s disease do not alter proteins directly, but instead affect how genes are turned on and off. Towfique Raj: In this talk, I will explain how we decode these hidden regulatory signals by combining human genetics with studies of microglia, the brain’s resident immune cells that help protect and maintain brain health. Towfique Raj: I will describe how we use eQTL analyses, which link DNA differences to changes in gene activity, to understand how risk variants influence how genes work inside these cells. Towfique Raj: I will also show how deep learning models can predict how small DNA sequence changes affect gene function and disease risk. [14:24] Towfique Raj: we were able to narrow down to four variants, one of which which sits on a transcription factor binding... and show that these are the functional variant and that's a really tremendous progress [19:08] Towfique Raj: I'll take questions after my talk if anyone has any

Source quotes

Towfique Raj: Most genetic changes linked to Alzheimer’s disease do not alter proteins directly, but instead affect how genes are turned on and off. Towfique Raj: In this talk, I will explain how we decode these hidden regulatory signals by combining human genetics with studies of microglia, the brain’s resident immune cells that help protect and maintain brain health. [0:04] Towfique Raj: minority here in this conference being an academic Towfique Raj: I will describe how we use eQTL analyses, which link DNA differences to changes in gene activity, to understand how risk variants influence how genes work inside these cells. [5:15] Towfique Raj: black box Towfique Raj: I will also show how deep learning models can predict how small DNA sequence changes affect gene function and disease risk. Towfique Raj: Attendees will gain a clear framework for how genetic risk shapes brain immune cell behavior and contributes to Alzheimer’s disease, and how these insights can guide new therapeutic targets. [19:08] Towfique Raj: You guys can go, but I'll take questions after my talk if anyone has any. Thank you.

Original transcript

[0:04] I'm delighted to be here and talk a little bit about what my lab does at at Mount Sinai in New [0:08] York. And I'm a I'm a research professor. So I'm an academic. Feel like I'm a minority here [0:13] in this conference being an academic. It's not many, but I'm really excited to talk about some [0:17] of the ongoing work in in my lab. So as we all know, Al neurogenerative disease as a whole is [0:24] the second uh leading cause of death worldwide. Alzheimer's disease uh affects about 50 million [0:31] people worldwide. Parkinson's disease 10 million and ALS about half a million or even more. Um and [0:39] you know unfortunately these diseases uh currently have no cure and there are no effective treatments [0:45] available to reverse their progression though some progress made for Alzheimer's disease [0:50] recently and just to you know give you a little background Alzheimer's disease you know are have a [0:57] pathological hallmarks in the brain you have these amaloid plaques they're sticky clumps of protein [1:03] fragments called a beta amalloid that's forming our brain and then you have another protein [1:08] called it towel or you know fibrous or protein got towel buildups inside our nerve cells as well as [1:14] you know the neurons they lose their ability um they're um then there's widespread damage to the [1:21] cells which causes reduced u shrink shrinkage of the brain as you can see uh here on in this [1:30] plot have reduced um shrinkage of the brain you also have neuroinflammation there other [1:36] uh events that happens in Alzheimer's brains. So what are the risk factors? Uh as we know that [1:42] there are non-modifiable risk factors such as age. As you age, you're more likely to get Alzheimer's [1:48] disease. Sex, um females are more likely. Ancestry, older African-Americans and Hispanic [1:55] Americans are two to four times more likely to get Alzheimer's disease. And the genetics, [1:59] which is uh you know, I'm a human geneticist, so most of my talk will be about genetics. uh [2:05] you know there is a genetic basis to Alzheimer's disease and heritability range from 60 to 80%. Of [2:11] course there are modifiable risk factors such as uh hypertension, cholesterol, obesity, exposure, [2:16] smoking increases uh and there are protective factors such as education, physical activity, [2:22] social contacts. So this figure didn't come up right but um the genetics of Alzheimer's disease [2:30] it's you know I'm I'm a human geneticist. So um there's been a tremendous progress made in the [2:35] last decade in identifying genes that causes Alzheimer's disease. We had one apoE that's [2:42] been most uh common genetic risk factor. Since then there's about over 75 of the regions of our [2:48] genomes linked to risk for Alzheimer's disease. As you can see we call this a Manhattan plot was [2:53] supposed to look like a skyscrapers in Manhattan. Each dot represent a genetic variation. So single [2:59] nucleotide polymorphisms. The y- axis is just the p value. So the remmit we made a lot of [3:05] progress and we've learned a lot about the disease biological pathways you know what are the you know [3:11] targets beyond amaloid and towel right so and one of the things we've learned is that vast majority [3:17] of these variants genetic variants which are common in population right we all some of us [3:22] have them here um they are non-oding they're not protein coding right they don't change the amino [3:29] acid changes the protein structure Because if it was protein coding, we know the gene, [3:35] right? We could target and have drug and that was a promise made ages ago. But it turns out [3:40] biology is hard and vast majority of them are non-coding and non-coding genome is very difficult [3:46] to understand and read, right? We don't know where which gene they're affecting. What are the [3:51] downstream consequences of these variants? And how do these um non-coding variants are um regulate uh [3:59] in within the cells? What are their functions? Uh as you can see 90 over 96% are non-coding. There's [4:06] about 4% of the the genes I was showing you are variants are mapped to a protein coding regions. [4:13] So one of the questions is how do we understand their variance you know within a cell? What are [4:17] they doing? There's genetic variants. They have functional impact. Do they affect chromatin or [4:23] gene expression which are or proteins, right? These are molecules that carry functions of the [4:28] disease. I know we had AI for a meeting, but I just want to give you a little background. So, [4:34] one of the interest in my lab is trying to figure out how can we better understand the [4:38] non-coding genome and its impact, right? So how does a variant affect cellular function [4:45] and molecular function such as gene expression, protein, methylation or epigenetic phenomena? [4:51] So this connection is very important and we you know we are been uh we've been very interested [4:57] in understanding and as we know um there's been immense progress made in generating large [5:03] amounts of data. These are whole genome sequencing data. So we can now sequence our genome and find [5:09] uh exact variation the genetic variations linked to disease. So I'm part of the Alzheimer's disease [5:15] sequencing project consortium run by national institute of aging. We've generated uh we have [5:20] 60,000 we're planning to generate 100,000 whole genomes from Alzheimer's patients and this is a [5:26] massive amount of whole genome sequencing data and we actually have no clue how to analyze [5:32] them data. What are their how do they affect the genome? How do they affect cell and how do they [5:37] affect function? Without it, it's just black box, right? The the genome, we have a variant, but we [5:43] don't know anything about them. And these genome wise associated studies that I was showing you [5:47] are great at finding them, but not understanding their underlying mechanisms, which is what I'm [5:52] interested in my research lab is to ask what are the underlying mechanisms on these variants. And [5:58] there are many ways of doing it. you know we we could take empirical approach where we can collect [6:04] uh data from patients they're living from blood or CS spinal fluid CSF or brain tissues from [6:12] autopsied patients and we can get their genetic background and link their genetic variation with [6:17] cellular function so it's a gene expression or protein these requires a large scale e effort [6:23] uh and we've been very successful in doing that but when it comes to rare cell types in the brains [6:29] their neurons, their astroytes and microglea they're very difficult right this is where AI [6:35] comes in this is where I think the deep learning methods sequence models are very important so one [6:41] of the advances made uh maybe some of you in the room have heard is there are there are sequence to [6:48] uh function models now where people many groups have used deep learning based tools such as alpha [6:54] genome which is popularized by Google deep mind was published like two years ago where you can [6:59] take a DNA sequence and assign its function. So you could predict by training amount large amounts [7:05] of data from cell cells from cell models from human data from animal models and train. So each [7:12] nucleotide genome you can see what they affect the expression or chromatin or RNA splicing or um or [7:20] transcription factor binding. And these models are very useful. Uh however they're limited, [7:26] right? They're limited. We talked a lot about this conference about data. In biology, data is [7:30] so important because if you're training it on data that's not from uh humans or even disease context, [7:37] it's very hard for the model to do anything that's useful. So, one of the things my lab has [7:42] gen we've generated and I'm going to just briefly summarize all these data that we've generated. [7:47] We've generated massive amount of data from human patients. These are human uh immune cells that are [7:54] autopsies. Within three hours we collect them. We generate transcriptton data, protein data, [7:59] chromatin data for each patients that come through our hospital. Um and through our autopsy we get [8:04] their brain tissues and we generate a large scale. We also are starting to generate single nuclei RNA [8:11] sequencing data. For each cell we can measure its expression how the RNA you know turn on and off. [8:18] uh we done peripheral blood and in another project we harmonized publicly available data from bulk [8:24] tissues these are brain tissues uh from mostly Alzheimer's cases uh they're done in autopsy [8:30] obviously there are limitations to these data because they're looking at post-mortem tissues [8:34] they're snapshot they're cross-sectional and that's the limitations that I can you know touch [8:39] on just very quickly one of the findings of the genetic studies was that vast majority of these [8:46] variants are you know are enriched or they were near genes that are highly expressed in these [8:52] immune cells called microglea. Micro glea are the CNS or central nervous system immune cells that [8:58] are found in the brain. Uh and they're important they clear out debris plaques and and interact [9:04] with neurons and they do synaptic pruning early on during the development. It turns out that [9:09] Alzheimer's disease risk really were enriched in these cell types not in neurons as you can [9:15] see from these plots. Microglea were highly enriched for heritability. So vast majority [9:20] of genetic variance risk are within these cell types called microglea. And you can see by two [9:26] plots. One on the left you can see the enrichment heritability compared to other cell types. Uh on [9:31] the right showing that many of the micro uh many of the genetic variants that we discovered through [9:37] genetic studies are are localized in these region regulatory regions called enhancers. [9:43] Enhancers where it switches genome on and off. So many of these variants sits there and this [9:48] is really striking enrichment. It's been a really exciting uh uh progress that we are now focusing [9:54] on you know many pharma and uh biotechs are really focusing on trying to identify novel drug targets [10:01] going after neuroinflammation and these immune cells micro micro so I'm just going to quickly [10:06] now move to the things most of you are here for is machine learning AI and I think this is a figure [10:12] from alpha genome you know alpha genome is a a deep learning tool and it takes raw DNA sequences [10:19] is up to one megabase and and you applies a deep learning architectures which has both informer [10:26] and decoders but also a long range interactions of the each gene using uh transformers and the model [10:33] generates multimodal predictions on a broad range of molecular effects shown here. So it [10:38] could predict over each nucleotide its impact on different functions expression splicing binding [10:46] which is really powerful right you give it a sequence and it can tell you whether each [10:50] nucleate has some impact now of course this is limited because as I was saying data is really [10:55] critical so one of the things we did um this is a our project it's collaboration with David [11:00] Nolles who's an amazing computational biologist we fine-tuned these models and we used a deep sea [11:08] Informer these are other tools I'm happy to talk offline if anyone is interested. uh what we did [11:13] is we fine-tuned right we fine-tuned them with cell type relevant regulatory data sets or data [11:20] sets that I've shown you that we've generated in my lab from micro ga these important cell types [11:24] and we fine-tuned them and uh we applied various different uh deep learning uh sequence to function [11:32] models uh deep sea informer as well as for RNA splicing looking at differences isoforms and [11:39] uh by doing that we're able to precisely at assign these delta score uh whether these variants affect [11:48] you know any of these molecular features and we did it for all 50,000 whole genomes for every [11:52] single nucleotide we had this variant effect prediction score or delta score whatever you [11:57] like to call it and these scores were then taken for further downstream analyses right so we can [12:03] take each score and we can see which of these variants are likely to be functional and important [12:09] so then we can do experimental validation with crisper cast 9. We can go and edit those variants [12:15] or we can use them to undo better studies or better genetic association studies. So I just [12:20] want to very quickly show these plots that our model actually does pretty well in predicting [12:27] chromatin accessibility. So you can see the two distribution in blue and in red and you can see [12:32] some of the variants uh that we found through uh doing empirical associated studies had a higher [12:39] delta score suggesting that these variants that we show by using genetic analysis using data that [12:46] they have effect on chromatin were a we were able to predict them that have shift slightly shifted [12:52] right higher delta score suggesting their impact on chromatin using two different models chrom MVP [12:58] net and former. So these are really a nice um uh kind of test principle that these methods work [13:05] and they do predict the impact of chromatin as which we're very excited about. Uh in addition [13:11] uh we also see by fine-tuning them uh it helps. not surprising because you're using relevant data [13:18] but I think uh the a you can see the AU you see a slight improvement on from a baseline [13:24] model without any CS uh brain related cell type uh data sets uh using uh fine-tuned with our micro GA [13:33] which is the immune cell data that I was referring to as well as neurons and astrotes we see a slight [13:38] improvement in in our accuracy of these models on the right hand side is it's a little um busy slide [13:46] but what I want to show you on the y axis are are kind of our empirical data where we had data sets [13:54] uh from human subjects and we measure their spicing effect on the x-axis are the deep [14:00] learning based delta scores you can see they're somewhat correlated right the deep learning [14:05] models pretty well predict although you can see a lot of a zero the models didn't predict them [14:11] um whereas we had data suggesting this was able to at least do a good job with correlating uh these [14:19] variance to splicing levels in this case. So, you know, they do well, but obviously there's still [14:24] room for improvement. Uh I'm going to show you two other applications very quickly. This is a very [14:29] busy plot and if you're not a geneticist, this is very difficult to understand, but if you are, [14:34] great. On the top are these dots, and you can see each blue dots are single nucleide polymorphism of [14:40] genetic variance. And you can see there's four or five different snips of variants that are [14:46] identified. One of the challenges in the genetics is which one variant are causal are actually [14:51] functional because they're statistically they're correlated. They're on the same hletype. So we [14:56] can't tease apart which one is the causal. So this deep learning model by training them by using [15:02] predictive models we can precisely identify of these variants this is this is what these models [15:07] telling us that it's likely to be functionally a causal variant right I know I'm using causal but [15:13] with gen experimental validation um I go back and forth my colleagues um about what's causal [15:20] u but you can see that it prioritized about four variants um which is shown u that line with deep [15:27] learning um and you You can see there the two lines there, four little dots, four variants. So [15:33] we can out of the 542 snips, we're able to narrow down to four variants, one of which which sits on [15:41] a transcription factor binding. So it affects the binding of these P1 transcription factor which [15:47] regulates the expression of a gene called Ras F1C. I'll skip the the biology behind the gene, [15:53] but just want to show you these deep learning models are capable of prioritizing punitive [15:58] causal variants and we actually have a massively parallel reporter asset and crisper experimental [16:03] validation of this variant which is very very um very cool and exciting because we can take [16:08] computational models identify causal variant and do crisper edits of exactly those DNA in IPSC or [16:15] microglea or stem cell models and show that these are the functional variant and that's a really a [16:20] tremendous progress that we didn't have thanks to these uh machine learning models as well as [16:26] uh you know crisper based technology and the other application just very quickly since we have these [16:32] scores for every single nucleotide we can take these scores as priors right and see which of [16:38] the variants are important and use them for our case control study because in the past what case [16:44] control studies would do they treat every single variant equally but because we have these um deep [16:50] learning based variant effect prediction scores. Uh we can use them as a Beijian rare variant [16:57] burden test. I'm not going to go through the model but there's a paper and preprint if anyone [17:01] is interested to learn more about how we do these burden tests. And you can see that by doing burden [17:06] test it allows us to improve our uh identification of novel genes. So we can to put through the [17:13] 50,000 or 100,000 whole genomes and you know 50,000 cases 50,000 control and each nucleotide [17:21] are are kind of using a different weights based on their functional prediction scores and then [17:26] we can use that as a as a way to identify novel genes using these method that we developed GU [17:32] in collaboration with David Nolles's lab um and uh and then by doing that we were able to [17:38] identify additional 20 novel Alzheimer's disease genes uh that weren't identified before. Right? So [17:46] these models help in prioritizing causal variance, puditive causal variants as well as target genes [17:53] that you know simple models can. So in summary, I think just very quickly these methods are very [17:59] useful uh such as alpha genome, but I think you need to fine-tune them. They help you [18:04] identify causal variants. Uh they identify genes and targets. um they enable and do new discovery. [18:11] However, there are limitations. I think they are predictions poorly represented cell types that [18:15] require experimental validation. Especially in rare cell types where we don't have data to train, [18:21] you have to be very careful. Um this is why experimental validation is going to be key [18:25] uh moving forward. The last minute I'm just going to quickly talk about something different. [18:30] This is an um a project, a collaboration actually and mount si led this effort. [18:36] It's a it's an AI agent, AI co-scientist that I was part of. Very fortunate, very talented [18:42] group Mount Sinai in collaboration with King Wang and um at Stanford. Uh as you can see the [18:48] team members are here. This was selected as the Alzheimer's insights AI prize as a agentic AI. [18:54] there's a lot um of improvement but very quickly again we added a lot of Alzheimer's relevant data [19:01] imaging cellular process RNA sequencing a ton of data over 100 uh 80 tools and 60 data sets [19:08] and it does and has a better performance and it also has a very better uh real world uh task it [19:14] does well in comparing with baseline LLMs so you know and this is going to be available through [19:21] Alzheimer's disease data initiative uh portal and you could uh use it and download it. So in the uh [19:26] this is a thank you acknowledgement slide. This is my research lab and thank you all for your [19:32] time. You guys can go but I'll take questions after my talk if anyone has any. Thank you.