Two revolutions are happening in parallel in cancer biology. One extracts molecular signals from a vial of blood. The other maps gene expression across tissue architecture at single-cell resolution. Until recently, these fields evolved independently — different platforms, different computational toolkits, different clinical entry points. AI is beginning to connect them, and the convergence has consequences for how we detect, monitor, and ultimately defeat cancer.
Figure 1. AI as the integration layer between liquid biopsy (ctDNA, fragmentomics, methylation profiling, CH filtering) and spatial transcriptomics (domain identification, deconvolution, cell-cell communication, drug resistance mapping), converging toward a validated clinical decision pipeline.
The liquid biopsy revolution: hearing cancer in the blood
Circulating tumor DNA (ctDNA) is shed by tumors into the bloodstream. The core challenge has always been signal-to-noise: tumor-derived fragments may represent less than 0.1% of all cell-free DNA in plasma. Early approaches relied on targeted panels — you needed to know which mutations to look for before you looked.[8]
That constraint is dissolving. A new generation of AI-driven multimodal approaches reads multiple ctDNA features simultaneously: nucleosome positioning patterns (fragmentomics), methylation state, copy number variation, and end-motif composition.[2] SPOT-MAS integrates all four modalities through a machine learning framework and achieves multi-cancer early detection from a single blood draw.[3] MRD-EDGE uses deep learning on whole-genome sequencing to push sensitivity into the parts-per-million range for minimal residual disease monitoring after surgery.[1]
MRD-EDGE[1]
Deep learning WGS-based SNV + CNV detection. Pushes MRD sensitivity to parts-per-million allele frequency.
SPOT-MAS[3]
Multimodal ML integrating methylomics, fragmentomics, CNV, and end motifs for multi-cancer early detection.
PRIME[6]
Multi-algorithm ML combining ctDNA-MRD with clinical features to predict NSCLC progression risk pre-imaging.
MetaCH[5]
ML framework classifying clonal hematopoiesis variants from tumor-derived ctDNA — reducing false positives in plasma.
The critical clinical application is treatment monitoring. The PRIME model — a multi-algorithm ML system combining ctDNA-MRD with clinical features — predicts progression risk in non-small cell lung cancer before imaging changes become visible.[6] In the GOZILA study, ctDNA-guided targeted therapy in advanced gastrointestinal cancers demonstrated that liquid biopsy can actively route treatment decisions, not just inform them.[7] A companion challenge is clonal hematopoiesis: aging immune cells accumulate somatic mutations that mimic tumor signals in plasma. MetaCH, an ML framework, now classifies CH variants from tumor variants in cfDNA, making ctDNA interpretation substantially more trustworthy.[5]
What liquid biopsy tells you is systemic and dynamic: what the tumor is shedding, how its genomic landscape is shifting, and whether treatment is working. What it cannot tell you is where in the tumor drug resistance originates.
The spatial revolution: mapping the tumor in its landscape
Spatial transcriptomics (ST) measures gene expression while preserving the physical coordinates of each measurement within tissue. Rather than averaging across dissociated cells, ST maintains the architecture: which immune cells border the tumor nest, how drug-resistant subclones cluster near vasculature, which ligand-receptor interactions are active between fibroblasts and malignant cells.[13]
The computational toolkit has matured rapidly. STAGATE and GraphST identify spatial domains — contiguous tissue regions with coherent expression programs — using graph neural networks that weight spot similarity by both gene expression and physical proximity.[9][10] Cell2location and RCTD deconvolve bulk spatial spots into constituent cell types, allowing fine-grained mapping of immune infiltration and stromal composition across the tumor microenvironment.[11]
STAGATE[9] / GraphST[10]
Graph attention auto-encoders for spatial domain identification, integrating gene expression and spatial proximity.
Cell2location[11]
Hierarchical Bayesian model for fine-grained cell-type mapping. Top performer across benchmarks for rare cell types.
CellChat v2[12]
Incorporates spatial coordinates for proximal cell-cell communication inference; expanded CellChatDB with 1,000+ interactions.
SpaRx[14]
Maps spatially heterogeneous drug resistance at single-spot resolution via dynamic adversarial domain adaptation.
Cell-cell communication inference has become a discipline unto itself. CellChat v2 incorporates direct spatial location to infer proximal interactions — not just which ligand-receptor pairs are co-expressed, but which physically adjacent cells are actively signaling.[12] This has direct clinical consequence: the spatial pattern of immune exclusion at the tumor margin is one of the strongest predictors of immunotherapy response.[13]
Most directly translatable is SpaRx, which maps spatially heterogeneous drug resistance at single-spot resolution before treatment is initiated — answering the question that liquid biopsy cannot: not just whether resistance will emerge, but where it already lives.[14]
Figure 2. Tumor biopsy spatial zones (drug-resistant core, immune margin, stromal niche) shed ctDNA into plasma. AI tools (MRD-EDGE, SPOT-MAS) extract signal from plasma; spatial context from tissue is fed simultaneously. A multimodal AI fusion layer integrates both streams into a unified clinical decision.
The convergence: AI as the bridge
The missing piece in the emerging precision oncology architecture is the integration of these two data streams. The logic is compelling: ctDNA reports the tumor's systemic genomic state and responds to treatment in near real-time; spatial transcriptomics provides a high-resolution map of intra-tumoral heterogeneity and resistance topology. Together, they are complementary channels of the same underlying biology.[4]
Foundation models are beginning to operate at this intersection. HEIST is pretrained on 22.3 million cells from 124 tissues across 15 organs, modeling spatial interactions with gene-cell cross-modal attention.[16] SToFM integrates gene expression and spatial coordinates directly into pretraining, outperforming single-modality approaches on downstream spatial tasks.[17] Nicheformer, pretrained on 110 million cells from the SpatialCorpus dataset, enables spatial composition prediction and label transfer across experiments and platforms.[15]
A clinical-grade integration pipeline connecting ctDNA monitoring to spatial transcriptomics profiling must address: harmonized variant calling across platforms, reproducible spatial domain identification across tissue sections and institutions, a validated multimodal fusion method, and a regulatory framework for AI-derived clinical recommendations. None of these are insurmountable — but all require deliberate bioinformatics infrastructure, not just model development.
What clinical deployment requires
Enthusiasm for multimodal AI in oncology must be tempered by a clear-eyed view of the pipeline requirements. A 2025 benchmark of histology-to-spatial gene expression prediction methods concluded that current approaches show limited clinical accuracy at the individual patient level, even as group-level performance metrics look compelling.[18] The gap between benchmark accuracy and clinical utility is not unique to spatial transcriptomics — it is a recurring theme across computational pathology.
For ctDNA, the primary challenges remain preanalytical (blood collection, cfDNA extraction, library preparation) and analytical (variant calling at ultra-low allele frequencies, clonal hematopoiesis filtering). Both have validated solutions in clinical genomics laboratories, but those solutions were built for targeted panels, not multimodal whole-genome approaches.[8][2]
The road ahead
Liquid biopsy answers the question of what: what mutations are circulating, what the MRD signal looks like, what the epigenomic landscape of shed DNA reveals.[4] Spatial transcriptomics answers the question of where: where resistance lives in tissue, where immune exclusion is occurring, where the most dangerous cell subpopulation is positioned relative to vasculature and stroma.[13]
AI is the grammar that will eventually let these two vocabularies form a sentence. The tools are maturing rapidly — foundation models pretrained on tens of millions of spatially resolved cells,[15][16] deep learning frameworks that extract signal from plasma at allele frequencies below 1 in 10,000,[1] and multimodal fusion architectures that are beginning to span data modalities.[4] The bottleneck has shifted from algorithm to infrastructure: robust, validated, and auditable pipelines capable of operating within the regulatory perimeter of clinical genomics.
That infrastructure challenge is where the next decade of precision oncology will be decided.
Zetobit builds compliant bioinformatics pipelines for multi-omics and liquid biopsy applications. If you're navigating the transition from research-grade tools to clinical-grade infrastructure, we can help.
Get in touch →