Skip to content

Single-Cell Intelligence Agent -- Learning Guide: Advanced Topics

Version: 1.0.0 Date: 2026-03-22 Author: Adam Jones

1. TME Classification: Deep Dive

1.1 The Four Immunophenotypes

Tumor microenvironment classification is central to immunotherapy patient selection. The field has converged on four canonical phenotypes, each with distinct cellular composition, spatial organization, and therapeutic implications.

Hot-Inflamed TME

Cellular hallmarks: - CD8+ T cells > 15% of total cells - Active cytotoxic gene program (GZMB+, PRF1+, IFNG+) - PD-L1 expression on tumor and immune cells - Tertiary lymphoid structures may be present

Molecular signatures: - Interferon-gamma signaling: STAT1, IRF1, CXCL9, CXCL10, CXCL11 - Cytotoxic effector program: GZMA, GZMB, PRF1, GNLY, NKG7 - Antigen presentation: HLA-A, HLA-B, HLA-C, B2M, TAP1, TAP2

Clinical implication: Strong candidate for checkpoint inhibitor monotherapy. Response rates: 30-50% for anti-PD-1/PD-L1.

Cold-Desert TME

Cellular hallmarks: - Total immune infiltrate < 10% of cells - Minimal T cell presence (< 2% CD8+) - Low neoantigen load (low TMB) - No tertiary lymphoid structures

Molecular signatures: - Absent IFN-gamma signaling - Low MHC class I expression - Active Wnt/beta-catenin signaling (immune exclusion) - PTEN loss (PI3K pathway activation)

Clinical implication: Checkpoint inhibitors alone are ineffective. Requires immune priming: - Oncolytic virus (T-VEC) to induce immunogenic cell death - STING agonist to activate innate immunity - Radiation therapy for abscopal effect - Bispecific T-cell engagers (BiTEs) to bypass recruitment defect

Excluded TME

Cellular hallmarks: - Immune cells present at tumor margin but excluded from core - Dense stromal barrier (fibroblasts, CAFs, myofibroblasts) - Immune cells "stuck" at the invasive front - Angiogenic vasculature without immune migration signals

Molecular signatures: - TGF-beta signaling: TGFB1, TGFB2, SMAD2, SMAD3 - Stromal activation: COL1A1, COL1A2, FN1, POSTN - CXCL12/CXCR4 axis (immune cell trapping) - VEGF-driven angiogenesis without CXCL9/10 chemokines

Clinical implication: Target the stromal barrier: - Anti-TGF-beta (bintrafusp alfa) to reduce fibrosis - Anti-VEGF + anti-PD-L1 combination (atezolizumab + bevacizumab) - FAK inhibitors to disrupt stromal architecture - Anti-CXCR4 to release trapped immune cells

Immunosuppressive TME

Cellular hallmarks: - Immune cells present and infiltrating but functionally suppressed - High regulatory T cell (Treg) fraction (> 10%) - M2-polarized macrophages dominant - Myeloid-derived suppressor cells (MDSCs) present

Molecular signatures: - Immunosuppressive cytokines: IL-10, TGF-beta, IL-35 - Metabolic suppression: IDO1, ARG1, NOS2 (tryptophan/arginine depletion) - Exhaustion markers on T cells: LAG3, TIM3/HAVCR2, TIGIT, TOX - Checkpoint overexpression: CTLA-4, PD-1 on T cells; PD-L1, PD-L2 on myeloid

Clinical implication: Multi-pronged approach: - Dual checkpoint (anti-PD-1 + anti-CTLA-4) - Treg depletion (anti-CCR8, low-dose cyclophosphamide) - Macrophage reprogramming (CSF1R inhibitor) - MDSC differentiation (ATRA, HDAC inhibitor)

1.2 Classification Algorithm

The Single-Cell Intelligence Agent's TMEClassifier implements a hierarchical decision tree:

Step 1: Spatial override
  - "absent" + immune < 0.05  -->  COLD_DESERT
  - "margin" + immune > 0.05  -->  EXCLUDED

Step 2: Hot-inflamed check
  - CD8 >= 15% AND immune >= 25%
    - Suppressive > 0.4  -->  IMMUNOSUPPRESSIVE
    - Otherwise           -->  HOT_INFLAMED

Step 3: Excluded check
  - Immune >= 10% AND stromal > 20%  -->  EXCLUDED

Step 4: Immunosuppressive check
  - Suppressive > 0.3 AND immune >= 10%  -->  IMMUNOSUPPRESSIVE

Step 5: Cold check
  - Immune < 10%  -->  COLD_DESERT

Step 6: PD-L1 rescue
  - PD-L1 high AND CD8 >= 5%  -->  HOT_INFLAMED

Default: COLD_DESERT

The suppressive score is a weighted combination: - 50%: suppressive cell fraction (Treg + MDSC + M2 macrophage) / 0.2 - 50%: suppressive gene score (IDO1, TGFB1, IL10, VEGFA, ARG1, NOS2)

1.3 Evidence Levels for TME Classification

Evidence Available Level Confidence
Spatial context + PD-L1 TPS + scRNA-seq STRONG High
PD-L1 TPS + scRNA-seq (no spatial) MODERATE Medium
scRNA-seq only (no PD-L1, no spatial) LIMITED Low

2. Subclonal Architecture and Clonal Dynamics

2.1 Why Subclones Matter

Cancer is not a monolithic disease. A single tumor contains multiple subclonal populations, each with distinct: - Somatic mutation profiles (driver and passenger) - Copy number aberrations (gains, losses, LOH) - Transcriptomic programs (proliferation, invasion, immune evasion) - Drug sensitivity profiles

Under therapeutic selective pressure, resistant subclones expand:

Before Treatment:
  Clone A (80%): Drug-sensitive, antigen+
  Clone B (15%): Moderate sensitivity, antigen+
  Clone C (5%):  Resistant, antigen-negative

After 8 Weeks of CAR-T:
  Clone A (5%):  Depleted by CAR-T
  Clone B (20%): Partially depleted
  Clone C (75%): Expanded (antigen escape)

2.2 Single-Cell Subclonal Detection

Methods for inferring subclonal architecture from scRNA-seq:

Method Input Output Mechanism
inferCNV scRNA-seq expression Clone-specific CNV profiles Expression deviation from normal reference
CopyKAT scRNA-seq expression Aneuploid/diploid classification Bayesian segmentation
Numbat scRNA-seq + genotype Haplotype-aware CNV + clone tree Allele-specific expression
clonealign scRNA-seq + scDNA-seq Clone-to-transcriptome mapping Statistical alignment

2.3 Escape Risk Scoring

The SubclonalRiskScorer evaluates four risk factors per clone:

Factor Weight Threshold
Antigen-negative (expression < 0.1) +0.4 Binary flag
Clone expanding +0.2 Boolean (serial samples)
High proliferation index up to +0.2 Proportional to MKI67/TOP2A
Resistance genes present +0.05/gene (max +0.2) Count of resistance-associated genes

Overall risk classification: - HIGH: antigen-negative fraction > 10% - MEDIUM: antigen-negative > 3% or any individual clone at HIGH risk - LOW: all clones below thresholds

Timeline estimation: Using exponential growth: t = T_doubling * log2(0.5 / current_fraction)

Example: If antigen-negative fraction is 5% and tumor doubling time is 14 days: - t = 14 * log2(0.5 / 0.05) = 14 * 3.32 = 46.5 days to reach 50% dominance

3. Spatial Transcriptomics

3.1 Technology Landscape

Spatial transcriptomics preserves the physical location of gene expression measurements within tissue:

Visium (10x Genomics)

  • Resolution: 55-micron spots (5-10 cells per spot)
  • Coverage: Whole transcriptome (~20,000 genes)
  • Tissue: Fresh-frozen or FFPE
  • Workflow: Tissue on barcoded slide -> permeabilization -> mRNA capture -> sequencing
  • Analysis: Requires computational deconvolution (cell2location, RCTD) to resolve cell types within spots

MERFISH (Vizgen)

  • Resolution: Subcellular (individual transcripts)
  • Coverage: 100-500 gene panel (custom design)
  • Tissue: Fresh-frozen
  • Workflow: Tissue on slide -> sequential rounds of hybridization + imaging
  • Analysis: Direct cell segmentation and gene assignment

Xenium (10x Genomics)

  • Resolution: Subcellular
  • Coverage: 100-5,000 gene panel (expanding)
  • Tissue: Fresh-frozen or FFPE
  • Workflow: In situ padlock probe hybridization + rolling circle amplification
  • Analysis: Cell segmentation -> direct transcript counting per cell

CODEX (Akoya)

  • Resolution: Single cell
  • Coverage: 40-60 proteins (antibody panel)
  • Tissue: FFPE or fresh-frozen
  • Workflow: Sequential antibody staining + fluorescence imaging
  • Analysis: Protein co-expression -> cell typing

3.2 Spatial Analysis Methods

Analysis Method What It Reveals
Spatial autocorrelation Moran's I Genes with spatially structured expression
Niche identification Cell neighborhood analysis Co-occurring cell type combinations
Cell-cell proximity Pairwise distance analysis Which cell types are physically adjacent
Spatial deconvolution cell2location, RCTD Cell type composition of Visium spots
Tissue segmentation Histological features + expression Tumor vs. stroma vs. necrosis regions
Spatial communication MISTy, SpaTalk Location-aware ligand-receptor analysis

3.3 Spatial Niches in Oncology

Clinically relevant spatial patterns:

Spatial Niche Cell Types Clinical Significance
Tumor-immune interface CD8+ T, tumor, DC Active immune surveillance, checkpoint response
Tertiary lymphoid structure B cell, T cell, FDC Positive prognosis, improved immunotherapy response
Fibrotic barrier CAF, myofibroblast Immune exclusion, anti-TGFb target
Hypoxic core Tumor, few immune Radioresistance, angiogenesis driver
Perivascular niche Endothelial, pericyte, tumor Metastatic dissemination route
Necrotic zone Dead/dying cells Antigen release, DAMP signaling

4. Trajectory Inference

4.1 What Are Cellular Trajectories?

Single-cell snapshots capture cells at different stages of continuous processes (differentiation, activation, exhaustion). Trajectory inference algorithms order cells along these continuous paths in "pseudotime."

4.2 Trajectory Types

Type Start State End State Clinical Relevance
Differentiation Progenitor/stem Mature cell HSC transplant engraftment
Activation Naive T cell Effector T cell Immune response quality
Exhaustion Effector T cell Exhausted T cell (TOX+) Checkpoint inhibitor response
EMT Epithelial Mesenchymal Metastatic potential
Stemness Differentiated tumor Cancer stem cell Treatment resistance
Cell cycle G1 G2/M Proliferation rate, chemo sensitivity

4.3 Trajectory Inference Methods

Method Approach Strengths Key Paper
Monocle3 Principal graph Handles branching, scalable Cao et al., Nature 2019
PAGA Partition-based Robust, preserves topology Wolf et al., Genome Biology 2019
RNA velocity (scVelo) Spliced/unspliced ratios Directionality without time series Bergen et al., Nature Biotech 2020
Palantir Diffusion maps Probabilistic fate assignment Setty et al., Nature Biotech 2019
CytoTRACE Gene counts as proxy Simple, no assumptions Gulati et al., Science 2020

4.4 RNA Velocity

RNA velocity infers the direction and speed of gene expression change by comparing unspliced (nascent) and spliced (mature) mRNA:

  • Positive velocity (unspliced > expected): Gene is being upregulated
  • Negative velocity (unspliced < expected): Gene is being downregulated
  • Zero velocity (equilibrium): Gene is at steady state
import scvelo as scv

# Load data with spliced/unspliced counts
adata = scv.read("sample.h5ad")

# Compute velocity
scv.pp.moments(adata)
scv.tl.velocity(adata, mode='dynamical')
scv.tl.velocity_graph(adata)

# Visualize on UMAP
scv.pl.velocity_embedding_stream(adata)

5. Foundation Models for Single-Cell Biology

5.1 scGPT

Architecture: Transformer-based generative pre-trained model for single-cell data.

Pre-training: 33 million cells from CellxGene, trained on gene expression prediction using masked token modeling.

Capabilities: - Zero-shot cell type annotation - Gene expression imputation - Perturbation response prediction - Multi-batch integration - Gene regulatory network inference

Performance benchmarks (from Cui et al., Nature Methods 2024): - Cell type annotation: 93.5% accuracy (zero-shot on held-out datasets) - Batch integration: superior to scVI on 6/8 benchmarks - Perturbation prediction: R=0.85 correlation with observed perturbation effects

5.2 Geneformer

Architecture: BERT-style transformer trained on gene expression rank order.

Pre-training: 30 million cells from public data, using attention-based gene embeddings.

Key innovation: Represents cells as ordered sequences of genes (ranked by expression), enabling transfer learning across tissues and species.

Capabilities: - Context-aware gene function prediction - Disease state classification - Therapeutic target nomination - Dosage sensitivity prediction

Performance (from Theodoris et al., Nature 2023): - Transfer learning accuracy: 85-95% across tissue types - Network biology prediction: improved over expression-based methods - Chromatin dynamics prediction: validated experimentally

5.3 scFoundation

Architecture: Large-scale pre-trained model (100M+ parameters) for cell representation learning.

Pre-training: 50 million+ cells from diverse tissues and species.

Capabilities: - Universal cell embeddings for cross-dataset integration - Drug response prediction - Cell fate prediction

5.4 Integration with the Single-Cell Intelligence Agent

Foundation models can serve as: 1. Embedding backbone: Replace BGE-small-en-v1.5 with scGPT cell embeddings for cell-level vector search 2. Annotation engine: Zero-shot cell type prediction via scGPT 3. Perturbation simulator: Predict drug response at single-cell resolution 4. Integration layer: Cross-dataset harmonization via Geneformer embeddings

The agent's knowledge base documents these models and their capabilities. NIM endpoint integration is planned for v2.0.

6. GPU Benchmarks for Single-Cell Analysis

6.1 RAPIDS vs. CPU Benchmarks

Operation Dataset Size CPU (seconds) GPU (seconds) Speedup
PCA (50 comps) 50K cells 45 0.9 50x
PCA (50 comps) 500K cells 480 4.2 114x
UMAP 50K cells 120 2.4 50x
UMAP 500K cells 1,800 12 150x
kNN (k=30) 50K cells 90 0.8 112x
kNN (k=30) 500K cells 960 3.5 274x
Leiden (res=0.5) 50K cells 30 1.0 30x
Leiden (res=0.5) 500K cells 350 5.0 70x
Full pipeline 50K cells 345 7.5 46x
Full pipeline 500K cells 3,590 24.7 145x

Benchmarks on NVIDIA A100 80GB. CPU benchmarks on AMD EPYC 7742 64-core.

6.2 Memory Requirements

Dataset Size CPU RAM GPU VRAM
10K cells 2 GB 1 GB
50K cells 8 GB 4 GB
100K cells 16 GB 8 GB
500K cells 64 GB 32 GB
1M cells 128 GB 64 GB

6.3 rapids-singlecell

The rapids-singlecell package provides GPU-accelerated Scanpy-compatible functions:

import rapids_singlecell as rsc

# GPU-accelerated preprocessing
rsc.pp.normalize_total(adata)
rsc.pp.log1p(adata)
rsc.pp.highly_variable_genes(adata)
rsc.pp.pca(adata)

# GPU-accelerated analysis
rsc.pp.neighbors(adata)
rsc.tl.leiden(adata)
rsc.tl.umap(adata)

# Results are identical to Scanpy, 50-150x faster

7. Cell-Cell Communication Analysis

7.1 Ligand-Receptor Databases

Database Interactions Source
CellPhoneDB 2,500+ Curated from literature
CellTalkDB 3,000+ Curated + predicted
NicheNet 6,000+ Ligand-target predicted
CellChatDB 2,000+ Curated with pathway context
LIANA Meta-database Consensus of multiple databases

7.2 Analysis Methods

Method Approach Output
CellPhoneDB Permutation test on L-R co-expression P-values per L-R pair per cell type pair
CellChat Quantitative mass-action model Interaction strength, pathway activity
NicheNet Ligand activity prediction from target genes Ligand prioritization by downstream effect
LIANA Consensus of multiple methods Aggregated interaction scores

7.3 The Single-Cell Intelligence Agent's L-R Knowledge

The agent curates 25 ligand-receptor pairs across clinically actionable pathways:

Pathway Ligand Receptor Clinical Relevance
Checkpoint CD274 (PD-L1) PDCD1 (PD-1) Checkpoint inhibitor target
Checkpoint CD80 CTLA4 Ipilimumab target
Chemokine CXCL12 CXCR4 Immune cell migration/trapping
Chemokine CCL2 CCR2 Monocyte/macrophage recruitment
Growth factor EGF EGFR TKI target (erlotinib, osimertinib)
Growth factor HGF MET MET inhibitor target
Notch DLL1 NOTCH1 Cancer stem cell maintenance
Wnt WNT5A FZD5 Immune exclusion, beta-catenin
Hedgehog SHH PTCH1 Stromal activation
Angiogenesis VEGFA KDR (VEGFR2) Anti-VEGF target (bevacizumab)

8. Biomarker Discovery at Single-Cell Resolution

8.1 Advantages Over Bulk Discovery

Feature Bulk Discovery Single-Cell Discovery
Specificity Tissue-level Cell-type-specific (AUROC > 0.9)
Confounders Cell composition changes confound DE Direct cell-type DE
Sensitivity Rare cell markers diluted Detectable at 0.1% frequency
Actionability Unknown cellular source Known cell type enables targeted therapy

8.2 Discovery Workflow

scRNA-seq data (disease vs. control)
         |
         v
Cell type annotation (57 cell types)
         |
         v
Per-cell-type differential expression
         |
    +----+----+----+
    |    |    |    |
    v    v    v    v
  CD8  Treg  Mac  ...
  DE   DE    DE
    |
    v
Specificity scoring (AUROC per gene per cell type)
    |
    v
Surface protein filter (is_surface = True)
    |
    v
Clinical validation check (existing assay, clinical trial)
    |
    v
BiomarkerCandidate output

8.3 Biomarker Types

Type Definition Example
Diagnostic Distinguishes disease from normal CD19 for B-ALL detection
Prognostic Predicts outcome regardless of treatment TOX+ exhausted CD8 fraction predicts poor OS
Predictive Predicts treatment response PD-L1 on tumor cells predicts anti-PD-1 response
Pharmacodynamic Measures treatment effect CD8/Treg ratio change under immunotherapy

9. CAR-T Target Validation

9.1 The Ideal CAR-T Target

Property Ideal Acceptable Unacceptable
On-tumor coverage > 95% > 70% < 50%
Off-tumor vital organs 0 hits Low-level (< 0.5 TPM) High in heart, brain, lung
Therapeutic index > 10 > 3 < 3
Heterogeneity Low (uniform expression) Moderate High (bimodal)
Escape risk Low (essential gene) Medium High (dispensable antigen)

9.2 The Agent's Target Validation Pipeline

Target Gene (e.g., CD19, MSLN, HER2)
         |
    +----+----+
    |         |
    v         v
On-Tumor    Off-Tumor
Analysis    Safety Check
    |         |
    v         v
Coverage    8 vital organs:
percentage  brain, heart, lung,
Mean expr.  liver, kidney, pancreas,
            bone_marrow, intestine
    |         |
    +----+----+
         |
         v
Therapeutic Index = mean_on_tumor / (max_off_tumor + 0.01)
         |
         v
    +----+----+----+
    |         |    |
    v         v    v
FAVORABLE  COND.  UNFAVORABLE
(safe +    (risk   (safety or
 effective) mitig.) efficacy fail)

9.3 Safety Switch Integration

For CONDITIONAL targets, the agent recommends: - iCasp9 (inducible caspase 9): Dimerizer-activated suicide switch - EGFRt: Truncated EGFR enabling cetuximab-mediated depletion - Affinity-tuned CAR: Reduced scFv affinity discriminates high-expression tumor from low-expression normal tissue

10. Advanced Study Resources

10.1 Key Papers

Year Paper Impact
2017 Zheng et al., "Massively parallel digital transcriptional profiling" 10x Chromium technology paper
2018 Wolf et al., "SCANPY: large-scale single-cell gene expression data analysis" Standard Python toolkit
2019 Stuart et al., "Comprehensive Integration of Single-Cell Data" Seurat v3, integration methods
2020 Bergen et al., "Generalizing RNA velocity" RNA velocity dynamical model
2021 Stahl et al., "Visualization and analysis of gene expression in tissue sections by spatial transcriptomics" Visium technology
2023 Theodoris et al., "Transfer learning enables predictions in network biology" Geneformer foundation model
2024 Cui et al., "scGPT: toward building a foundation model for single-cell multi-omics" scGPT foundation model

10.2 Online Courses

  • Single Cell Genomics (Wellcome Sanger Institute) -- Comprehensive bioinformatics training
  • Analysis of Single Cell RNA-seq Data (Cambridge University) -- Scanpy/Seurat tutorials
  • NVIDIA RAPIDS for Single-Cell -- GPU acceleration training

10.3 Practice Datasets

Dataset Cells Tissue Modality Access
PBMC 3K 2,700 Blood scRNA-seq 10x Genomics
PBMC 68K 68,000 Blood scRNA-seq 10x Genomics
Tabula Sapiens 500,000 Multi-tissue scRNA-seq CellxGene
Human Lung Cell Atlas 580,000 Lung scRNA-seq + spatial CellxGene
TCGA Pan-Cancer scRNA 1M+ Multi-cancer scRNA-seq TISCH2

HCLS AI Factory -- Single-Cell Intelligence Agent Learning Guide: Advanced Topics v1.0.0