This video presents the second session of a series dedicated to Artificial Intelligence (AI) and deep learning (DL) applications in bioinformatics, specifically focusing on gene expression data analysis, particularly single-cell and spatial transcriptomics.
Dr. Alex Pils from the Wellcome Sanger Institute outlines the significance of single-cell and spatial RNA sequencing for understanding cell-type-specific changes and niche definition, detailing the evolution of related technologies like 10x Genomics and Xenium. He discusses mainstream deep learning applications such as SCVI for batch correction and CellBender for background RNA ("soup") removal, emphasizing their effectiveness for large datasets. Pils also introduces the emerging concepts of foundational models and virtual cells but stresses the critical need for abundant, high-quality data and reliably established "ground truth" for successful deep learning applications, warning against models that overfit or lack proper validation.
Dr. Alexis Agasachev from Washington University in St. Louis provides a complementary perspective, starting with general machine learning principles like interpolation, extrapolation, and the importance of ground truth. He expresses skepticism regarding the current state of foundational models for gene expression, arguing that reported "cell numbers" often represent redundant data rather than independent observations, thus falling short of the data scale that enabled breakthroughs like GPT-3. He also notes the ongoing challenges in fully automating cell type annotation due to a lack of consensus in biological ground truth. Conversely, Agasachev is optimistic about the utility of Large Language Models (LLMs) for text-based bioinformatics tasks, including metadata curation, normalizing diverse biological terminology, and assisting in interpreting gene lists for novel insights. Both speakers conclude that while AI is a powerful tool for prediction and exploration, it still requires human expertise for diverse questioning, validation, and achieving true biological understanding, especially given the complexities and ground truth limitations inherent in biological data.
Summarizing AI in Gene Expression Analysis
0:03 Session Introduction: This is the second meeting dedicated to artificial intelligence and its application in bioinformatics, focusing on understanding the scope and future impact of these methods on daily bioinformatics routines.
0:49 Today's Topic & Speakers: The discussion is specifically dedicated to modern deep learning and AI for analyzing gene expression data. Speakers are Dr. Alex Pils from the Wellcome Sanger Institute and Dr. Alexis Agasachev from Washington University in St. Louis.
2:32 Dr. Alex Pils's Background: A chemist by training, transitioned to bioinformatics 12-15 years ago. Now a principal bioinformatician at the Sanger Institute, dealing with vast amounts of single-cell genomics data and supporting Human Cell Atlas projects.
3:45 Why Single-Cell RNA-seq Matters: Single-cell RNA sequencing (scRNA-seq) helps differentiate gene expression changes within individual cell types from changes in cell population proportions (e.g., in inflammation). This detailed view is crucial as averaging bulk samples can be uninformative.
6:49 Importance of Data Integration: ScRNA-seq data enables logical integration across different labs and protocols by assuming consistency within cell types (e.g., all macrophages should look similar). This enhances reproducibility and validation of biological insights.
8:03 Evolution of Single-Cell Technologies: Technologies evolved significantly after 2015 with droplet-based methods like 10x Genomics, which quickly became dominant due to its reliability and commercial success.
10:00 Diverse Single-Cell Assays: Beyond standard RNA-seq, methods include single-nucleus RNA-seq (for hard-to-dissociate tissues), single-cell ATAC-seq (profiling open chromatin), multi-ome (combined RNA-seq and ATAC-seq), CITE-seq (protein measurement via antibody tags), and VDJ-seq (immune repertoire analysis).
13:08 Rise of Spatial Transcriptomics: Spatial methods, such as 10x Genomics' Visium and Xenium, provide crucial context by mapping gene expression back to tissue location.
14:49 Xenium's Sub-cellular Resolution: Xenium, an imaging-based method, offers subcellular resolution, allowing researchers to localize RNA to specific cellular compartments (nucleus, cytosol, organelles) and segment cells using machine learning. It can profile hundreds of thousands of cells for 5,000 genes per sample.
16:32 Spatial Context Defines Niches: The spatial environment significantly influences cellular gene expression, sometimes as much as cell lineage, defining specific "niches" or "micro-environments" crucial for understanding processes like tumor heterogeneity.
19:11 Broad Applications of ML in Single-Cell/Spatial: ML is used for many tasks, including batch correction, automated cell type annotation, cell segmentation, inferring cell composition, identifying niches, discovering regulatory networks, optimizing organoid differentiation, and developing "virtual cells" or "foundational models."
23:31 Mainstream Deep Learning: Data Integration (SCVI): SCVI uses variational autoencoders for non-linear batch correction, effectively integrating large single-cell datasets from different sources, and dramatically outperforms other methods for large-scale data, requiring GPUs.
27:02 Mainstream Deep Learning: Soup Correction (CellBender): CellBender, another deep learning method (variational autoencoder), corrects for "soup" β background RNA leaked from broken cells in droplet-based and single-nucleus assays, which can lead to misinterpretations.
30:01 LLM Defined Cell Type Definitions: Large Language Models (LLMs) are being used with biological ontologies (controlled vocabularies) to mine literature and assist in cell type annotation, leading to more standardized metadata.
33:12 When Deep Learning Works Well: DL is most effective when there's abundant, good quality data and a reliably established ground truth for training and validation (e.g., variant calling in genomics, AlphaFold for protein structures). Overfitting is a major risk without independent validation.
35:40 Caution on "Hype" Models (Foundational/Virtual Cells): While ambitious, "foundational models" aiming to integrate all multi-modal biological data and "virtual cells" aiming to predict phenotypes often struggle to outperform simple linear models and lack robust validation due to the absence of clear ground truth and independent benchmarks.
38:31 Alex Pils on AI Agents: AI agents won't fully replace human bioinformaticians due to the diverse nature of biological questions. However, with overwhelming data volumes, agent-based analysis will become common, especially for tasks where "good enough" results are acceptable (like UMAP visualizations).
43:02 Dr. Alexis Agasachev's Background: Computer science background, focusing on developing algorithms for genomic data processing and analysis, along with collaborative applications.
44:12 Classical Machine Learning: Requires fixed tasks (classification, regression) and ground truth for training. Interpolation (predicting within known data) is easier than extrapolation (predicting outside known data). K-nearest neighbors is a strong baseline model, relying on meaningful distance metrics.
50:09 Skepticism on Gene Expression Foundational Models: Agasachev argues that reported data sizes (e.g., 100 million cells) for gene expression foundation models are misleading; many cells are replicates, not independent data points. The actual number of unique experimental conditions or data points is orders of magnitude lower than the text data used for GPT-3's breakthrough.
57:55 Criticism of Current Foundation Models: These models often don't outperform simpler linear models, perform poorly in zero-shot learning scenarios, and face challenges with appropriate metrics. Agasachev is skeptical of them achieving a GPT-3 level breakthrough soon.
59:16 Skepticism on Fully Automated Cell Type Annotation: While classification methods exist, full automation is challenging because there's no universal consensus on "true cell types," cell states, or their definitions, making it hard to establish consistent ground truth. Platforms like "Solation" aim to foster community consensus for better training data.
1:02:48 Optimism on LLMs for Textual Data: LLMs excel at text processing and are highly promising for bioinformatics tasks involving textual data, such as:
1:03:05 Data Set Annotation: Curating and processing metadata (e.g., RUMBLE search engine, data.ai startup).
1:05:07 Normalization of Biological Terms: Handling spelling variations and synonyms for biological terms (e.g., mouse strains) with surprising accuracy, even by general LLMs like ChatGPT.
1:08:41 Interpreting Gene Lists: Helping researchers interpret gene lists from experiments by identifying interesting candidates, filtering irrelevant results, and providing context, saving significant human effort.
1:11:09 Prompt-Based Bioinformatics: Allows users to input prompts for data analysis, potentially generating simple scripts (e.g., ggplot parameters). While it can save time, it requires user expertise to validate results and avoid poor quality outputs.
Science prioritizes understanding over mere prediction; AI is a tool for prediction, not a replacement for understanding.
Simple methods often work as well as complex AI.
The lack of reliable ground truth in biology significantly limits AI applications.
No "AlphaFold moment" (a definitive breakthrough) for gene expression is imminent due to these challenges.
LLMs are promising for textual tasks, but they are not fully reliable (require cross-checking), can be costly to scale, and open-source models may lag commercial ones.
LLMs are valuable for exploration, drafting, and general thinking, offering a "low-risk, high-reward" situation for generating unexpected insights.
1:16:51 Q&A: LLMs for Gene Set Enrichment Analysis: LLMs can provide good starting points for interpreting differentially expressed genes, identifying candidates, and providing context. However, they are less suitable for generating gene sets or interpreting gene ontologies (GO) directly, as GO describes functional interactions beyond expression changes.
1:23:04 Q&A: AI Spotting Patterns Better Than Humans: Agasachev is skeptical. AI models, in their current state, either require extensive human-designed mechanistic models or primarily rely on a k-nearest neighbors approach (identifying what's already seen in the data). They are not yet capable of generating genuinely novel biological insights or spotting patterns with greater efficiency than human researchers beyond interpolation.
1:25:51 Q&A: LLMs for PCA Interpretation: LLMs and other automated methods can be helpful starting points for interpreting principal components or other dimensionality reduction results in single-cell data, but still require domain expert biologists for final validation and meaningful annotation. Full automation is not yet achievable.
1:28:22 Q&A: AI for Therapeutic Discovery from Gene Lists: Directly finding a therapeutic route from a gene list using AI alone is currently not feasible for common, well-studied diseases, as straightforward biology has often been solved. AI can serve as a tool to uncover mechanistic biological processes that could lead to drug targets. It is more promising for drug screening or repurposing with known targets, especially in large-scale pharmaceutical applications, but not for de novo cure discovery for new diseases without further scientific steps.