Patient Stratification

Scope. A mechanistic, honestly-caveated overview of all major current approaches to patient stratification — sorting an apparently-homogeneous disease population into subgroups that differ in prognosis or treatment response — across oncology, immunology, and common disease, not limited to any one modality. Organized by approach family: (Part A) established molecular subtypes & unsupervised class discovery; (Part B) multi-omics, single-cell/spatial, and signature-scoring stratifiers; (Part C) clinical/genomic biomarkers, PRS, imaging, and the regulatory frame; (Part D) ML/AI, causal HTE, adaptive trials, EHR phenotyping, and immunotherapy stratification. For each: what it is → how it defines the subgroup (mechanism) → key tools → clinical use → strengths → limitations, closing with an honest read on what is actually adopted / winning now.

Two distinctions that unify the whole field.

1. Predictive vs prognostic. A prognostic marker informs likely outcome regardless of treatment; a predictive marker informs differential benefit from a specific treatment and requires a significant treatment-by-biomarker interaction (Ballman, JCO 2015, 10.1200/JCO.2015.63.3651). Plain ML captures the prognostic signal; causal/HTE methods and enrichment designs target the predictive signal. Conflating them drives wrong decisions.
2. Supervised vs unsupervised. Discover subgroups de novo (clustering) vs assign a new sample to a pre-defined subtype (nearest-centroid/classifier). The systems that reach the clinic pair a robust single-sample classifier with a specific, mechanism-linked decision.

The recurring, generalizable caution: bulk-expression subtypes are routinely contaminated by non-tumor/stromal/immune content (always ask what cells produced the signal), stability statistics measure reproducibility of a procedure, not biological reality, and a valid endotype biomarker does not guarantee a successful drug or a perfect responder-predictor.

Part A — Established Molecular Subtypes & Unsupervised Class Discovery

A1. Transcriptomic molecular subtypes (the deployed workhorse)

Partition a disease by genome-wide mRNA using either unsupervised clustering (discover) or a supervised nearest-centroid classifier (assign), then tie subtypes to biology and outcome.

• PAM50 breast cancer (Luminal A/B, HER2-enriched, Basal-like, Normal-like). Derived from the Perou/Sørlie intrinsic-subtype concept; Parker et al. 2009 (JCO, 10.1200/JCO.2008.18.1370) minimized to 50 genes and a nearest-centroid (Spearman) classifier + a continuous Risk-of-Recurrence (ROR) score. Commercialized as Prosigna on NanoString nCounter; FDA 510(k)-cleared Sept 2013 as a prognostic assay (10-yr distant-recurrence risk in postmenopausal HR+ early breast cancer). (Note: Oncotype DX/TAILORx is the assay with level-1 predictive chemo-benefit evidence; Prosigna's cleared claim is prognostic.) Limit: subtype instability near centroid boundaries (LumA/LumB flip); Normal-like is largely a low-cellularity artifact/QC flag.
• Consensus Molecular Subtypes (CMS) of colorectal cancer (CMS1 MSI-immune / CMS2 canonical / CMS3 metabolic / CMS4 mesenchymal). Guinney et al. 2015 (Nat Med, 10.1038/nm.3967) for the CRC Subtyping Consortium reconciled six competing classifiers via network-based consensus on ~4,000 transcriptomes → subtypes stable to method choice. ~13% are mixed/indeterminate (honest "unclassified" bin). Frequencies ~CMS1 14% / CMS2 37% / CMS3 13% / CMS4 23%. Classifiers: CMSclassifier, CMScaller (Eide 2017, 10.1038/s41598-017-16747-x). CMS1 rationalizes checkpoint blockade (aligns with dMMR/MSI-H pembrolizumab); CMS4 (worst prognosis) is stroma/TGF-β-driven. Limit: CMS4 signal is stroma-driven (cellularity-sensitive); single-sample calls less stable than the pooled consensus.
• TCGA expression subtypes. GBM (Proneural/Neural/Classical/Mesenchymal; Verhaak 2010, 10.1016/j.ccr.2009.12.020) — later refined to three subtypes after Wang 2017 (10.1016/j.ccell.2017.06.003) removed Neural as a normal-brain contamination artifact (the canonical bulk-contamination cautionary tale). Bladder (6-class consensus, Kamoun 2020, 10.1016/j.eururo.2019.09.006); Gastric (EBV/MSI/GS/CIN, TCGA 2014, 10.1038/nature13480) — EBV/MSI → checkpoint immunotherapy.
• SLE / autoimmune interferon-signature subtypes (IFN-high vs IFN-low) — a pathway-activation subtype from coordinated ISG over-expression (Baechler 2003, PNAS 10.1073/pnas.0337679100); the Chaussabel modular framework (Immunity 2008, 10.1016/j.immuni.2008.05.012) and longitudinal Banchereau 2016 (Cell, 10.1016/j.cell.2016.03.008) added plasmablast/neutrophil modules tracking flares/nephritis. Trial-enrichment closed the loop: anifrolumab (anti-IFNAR1) had larger effect in IFN-high patients (TULIP-2, Morand 2020, NEJM 10.1056/NEJMoa1912196); FDA-approved Aug 2, 2021. Limit: the IFN signature is dynamic (flares/infection/treatment); IFN-high enriches but doesn't guarantee response.
• Asthma T2-high/T2-low — a 3-gene bronchial-epithelial signature POSTN/CLCA1/SERPINB2 (Woodruff 2009, AJRCCM 10.1164/rccm.200903-0392OC — not PNAS, a common miscitation). Underpins eosinophil/FeNO-guided biologics (anti-IL-5/IL-4Rα). Cautionary tale: serum periostin as a predictive biomarker for anti-IL-13 failed in Phase 3 (LAVOLTA, Hanania 2016, 10.1016/S2213-2600(16)30265-X) — a valid endotype biomarker did not yield a successful drug.

A2. Unsupervised clustering / class discovery (how subtypes are found de novo)

The hard problem is not finding clusters (algorithms always return a partition) but proving they are real, reproducible, biologically meaningful.

• Hierarchical clustering (Eisen heatmap/dendrogram, Eisen 1998, PNAS 10.1073/pnas.95.25.14863) — foundational; discovered breast intrinsic subtypes (Perou 2000) and DLBCL GCB-vs-ABC cell-of-origin (Alizadeh 2000, 10.1038/35000501). Limit: greedy/irreversible; metric+linkage sensitive; cut-height arbitrary; no significance test.
• Consensus clustering / ConsensusClusterPlus — resampling meta-method quantifying stability and selecting k via consensus matrix / CDF / Δ(k) (Monti 2003, Mach Learn [10.1023/A:1023949509487]; package Wilkerson & Hayes 2010, 10.1093/bioinformatics/btq170). The TCGA-subtyping workhorse.
• NMF / metagenes — parts-based, interpretable factorization with cophenetic-correlation rank selection (Brunet 2004, PNAS 10.1073/pnas.0308531101); conceptual basis of mutational-signature extraction.
• The reproducibility caveat (load-bearing). Şenbabaoğlu 2014 (Sci Rep 10.1038/srep06207) showed consensus clustering produces crisp, stable-looking heatmaps on structureless null data; Δ(k) is often uninformative; proposes PAC (Proportion of Ambiguously Clustered pairs) and a reference/null model. Guardrails: compare against a null (M3C/SigClust/gap), report PAC, require external-cohort validation, rule out batch/technical covariates, and confirm biological+clinical coherence before claiming a subtype exists.

Part B — Multi-Omics, Single-Cell/Spatial & Signature-Scoring Stratifiers

B1. Multi-omics integrative clustering

Given N omics on the same patients, find one subtype set reflecting joint structure — better than any single layer.

• iCluster / iClusterPlus / iClusterBayes — joint latent-variable models (shared Z drives every layer). iClusterPlus adds mixed-data GLM likelihoods + lasso; iClusterBayes is fully Bayesian (MCMC, spike-and-slab, posterior uncertainty). Bioconductor iClusterPlus; TCGA workhorse. Shen 2009 (Bioinformatics 25:2906); Mo 2018 (Biostatistics [10.1093/biostatistics/kxx017]). Limit: compute-heavy grid search; linear.
• MOFA / MOFA+ — unsupervised Bayesian PCA-generalization → interpretable factors with per-modality variance decomposition; handles missing assays and mixed likelihoods; MOFA+ scales to single-cell. Argelaguet 2018 (Mol Syst Biol e8124), 2020 (Genome Biol 10.1186/s13059-020-02015-1). Canonical CLL application recovered IGHV/trisomy-12 + an oxidative-stress factor predictive of survival. Limit: linear; clustering is a downstream choice.
• Similarity Network Fusion (SNF) — build per-omic patient-similarity networks, fuse via iterative cross-diffusion, spectral-cluster the fused network. Nonlinear, noise-robust, a top benchmark performer. Wang 2014 (Nat Methods 10.1038/nmeth.2810); GBM/kidney examples. Limit: black-box (no feature drivers); needs complete data; hyperparameter-sensitive.
• MOVICS — ensemble meta-framework running 10 algorithms + consensus + full cancer-subtyping downstream (Lu 2020, Bioinformatics [10.1093/bioinformatics/btaa1018]). DIABLO / mixOmics — the supervised counterpart: multi-block sparse PLS-DA toward a known outcome → compact cross-omics biomarker panel (Singh 2019, 10.1093/bioinformatics/bty1054).
• Benchmarks (no universal winner). Rappoport & Shamir 2018 (NAR 10.1093/nar/gky889) — SNF/similarity methods among the strongest across 10 TCGA types; performance varies by cancer and metric. Cantini 2021 (Nat Commun 10.1038/s41467-020-20430-7) — intNMF best for clustering, MCIA most consistent, MOFA strongest for interpretability. Practical rule: match method to data type + clinical question; validate on survival/external cohorts.

B2. Single-cell & spatial stratification

Convert high-dimensional expression into patient-level features — a composition vector, a latent multicellular program, a whole-TME archetype, or a spatial neighborhood.

• Cell-composition / cell-state signatures — stratify by relative abundance of cell types/states; because proportions are compositional, use scCODA (Bayesian Dirichlet-Multinomial) or Milo (kNN-neighborhood NB-GLM, captures continuous states). Verified disease examples: COVID-19 severity (Stephenson 2021, Nat Med [10.1038/s41591-021-01329-2]), SLE (Perez 2022, Science [10.1126/science.abf1970] — ISG-high monocytes + GZMH+ CD8 split patients into two molecular subtypes), UC anti-TNF resistance via IL13RA2+/IL11+ inflammatory fibroblasts (Smillie 2019, Cell [10.1016/j.cell.2019.06.029], PMID 31348891). Limit: compositionality trap; annotation non-portability; dissociation bias; discards space.
• Ecotypes / EcoTyper — deconvolve bulk (CIBERSORTx) → NMF cell states → ecotypes (co-occurring multicellular communities) that stratify prognostically; unlocks single-cell-like structure from large legacy bulk cohorts. Luca 2021 (Cell [10.1016/j.cell.2021.09.014], 10 conserved carcinoma ecotypes) and Steen 2021 (Cancer Cell [10.1016/j.ccell.2021.08.011], DLBCL ecotypes). (Flag: exact "CE1–CE10"/"LE1–LE10" labels are high-confidence but not line-verified.) Limit: deconvolution reference-bound; co-occurrence ≠ physical co-localization.
• TME archetypes — coarse, deployable pan-cancer taxonomy. Bagaev 2021 (Cancer Cell [10.1016/j.ccell.2021.04.014]) — 29 functional signatures → 4 conserved TME subtypes (IE / IE-F / F / D) that predicted anti-PD-1/PD-L1/CTLA-4 response across melanoma/bladder/gastric (immune-favorable respond best). Combes 2022 (Cell [10.1016/j.cell.2021.12.004], PMID 34963056 — corrected from a common mislabel) — 12 dominant immune archetypes recurring across cancers. Limit: coarse-graining; batch/cohort-sensitive; correlative, not a qualified CDx.
• Spatial neighborhoods — the only family that measures who is next to whom. Schürch 2020 (Cell [10.1016/j.cell.2020.10.021]) CODEX CRC → 9 conserved cellular neighborhoods; CN coupling/fragmentation predicted outcome (organization, not just abundance, stratifies — why some "infiltrated" tumors still fail immunotherapy). Tools: SpaGCN, BayesSpace, squidpy. Limit: low throughput; small cohorts; immature standardization.

B3. Signature scoring & deconvolution (expression → a per-patient score)

• Enrichment/module scoring — ssGSEA (rank-based per-sample enrichment; Barbie 2009, Nature [10.1038/nature08460]) and GSVA (Hänzelmann 2013, BMC Bioinformatics [10.1186/1471-2105-14-7]); Seurat AddModuleScore (background-corrected module activity; Tirosh 2016, Science [10.1126/science.aad0501]). Continuous scores (IFN, proliferation, exhaustion) thresholded into strata. Limit: scores are cohort-relative (GSVA), gene-set-dependent, arbitrary thresholds, conflate cell source.
• Deconvolution — CIBERSORT/CIBERSORTx (ν-SVR against the LM22 22-cell-type matrix + batch correction + cell-type-specific expression imputation; Newman 2015 Nat Methods [10.1038/nmeth.3337], 2019 Nat Biotechnol [10.1038/s41587-019-0114-2]); plus xCell, MCP-counter, EPIC (absolute fractions), quanTIseq (RNA-seq TIL10). Immune fractions (CD8, M1/M2) stratify prognosis/ICB response. Limit: relative-vs-absolute confusion, reference-matrix dependence, spillover between correlated subsets, platform/batch effects.
• Companion-diagnostic-adjacent: the 18-gene T-cell-inflamed GEP / TIS — Ayers 2017 (JCI 10.1172/JCI91190); LASSO-derived across 9 tumor types on pembrolizumab, fixed weights + fixed cutoff (−1.540), AUC ~0.75 (vs PD-L1 IHC ~0.65). The closest of these to a CDx role (still decision-support, not an FDA-approved CDx replacing PD-L1 IHC).

Part C — Clinical/Genomic Biomarkers, PRS, Imaging & the Regulatory Frame

C1. Single-analyte & genomic biomarkers (the mature, adopted layer)

• Driver mutations (EGFR, BRAF, ALK/ROS1/RET/NTRK fusions, KRAS G12C) → matched TKIs, via tissue NGS or plasma cfDNA. FoundationOne CDx (tissue) and Guardant360 CDx (liquid) are FDA companion diagnostics. Limit: acquired resistance; a positive test ≠ response (BRAF V600E colorectal responds poorly to BRAF-inhibitor monotherapy — context matters).
• HER2 (IHC → reflex ISH) → trastuzumab/T-DXd; HercepTest was the first protein-based FDA CDx. The emerging "HER2-low" category exposes the limits of the original binary cut-point.
• ER/PR (IHC) → endocrine therapy — oldest predictive biomarker; predicts candidacy, not magnitude (why gene-expression assays were layered on top).
• MSI-H / dMMR → the basis of the first-ever tissue-agnostic FDA approval (pembrolizumab, 23 May 2017; Le 2017, Science 10.1126/science.aan6733, ORR 53%).
• TMB ≥10 mut/Mb → tissue-agnostic pembrolizumab (16 Jun 2020). Key nuance: no universal pan-cancer threshold — the predictive cut-point varies by histology (Samstein 2019, Nat Genet 10.1038/s41588-018-0312-8).
• PD-L1 IHC — companion or complementary depending on drug/indication; the "harmonization problem" (non-interchangeable clones 22C3/28-8/SP263/SP142; TPS/CPS/IC scoring; interobserver variability; Blueprint discordance, Hirsch 2017).
• Multigene expression assays — Oncotype DX (21-gene RS; the only assay validated for chemo-benefit prediction, TAILORx/RxPONDER) and MammaPrint (70-gene, MINDACT). Routine, reimbursed.

C2. Polygenic risk scores (PRS)

Aggregate hundreds-to-millions of common variants (GWAS-weighted) into one liability score that shifts an individual relative to a clinical baseline (a risk enhancer, not a binary call). CAD/ASCVD example (AHA 2025, Genomics plc): adding PRS to the AHA PREVENT tool gave ~6% NRI, reclassified ~8%, and high-PRS near-threshold individuals were ~2× likelier to develop ASCVD (vendor press release — directionally reliable, not peer-reviewed). ESC 2025 formally endorsed cautious PRS use. Central limitation — ancestry portability: European-dominated GWAS training degrades PRS performance in non-European populations (the PRIMED consortium exists to fix this). Adoption: mostly LDT/research; harder to validate than single-analyte assays.

C3. Protein/serum panels & imaging/pathology

• High-plex proteomics — Olink (PEA, ~21–5,000 proteins from 1–2 µL) and SomaScan (aptamer, ~1,300→11,000 proteins) → learned multi-protein signatures that beat single markers for CV risk and immune phenotyping. Mostly research/translational, few FDA-cleared CDx; platform-specific units (NPX vs RFU) not comparable; overlapping inflammatory pathways make disease-specific signatures hard.
• Radiomics — quantitative CT/MRI/PET features (increasingly DL) for prognosis/response/subtype inference; predominantly research-grade, growing use for trial enrichment.
• Digital-pathology foundation models (2024 frontier) — self-supervised on millions of WSI tiles → transferable embeddings predicting subtype/biomarker/mutation/prognosis directly from cheap H&E. UNI (Chen 2024, Nat Med [10.1038/s41591-024-02857-3]; >100M images/100k WSIs/20 tissues), Virchow (Vorontsov 2024, Nat Med [10.1038/s41591-024-03141-0]; 1.5M WSIs, 0.95 specimen-level AUROC incl. rare cancers), Prov-GigaPath (Xu 2024, Nature [10.1038/s41586-024-07441-w]). Strong benchmarks, mostly not yet standalone FDA-cleared (task-specific products like Paige Prostate are the cleared exceptions).

C4. Regulatory & trial framing

• Companion (CDx) = information essential for safe/effective use (must run before prescribing; co-approved with the drug). Complementary = aids the benefit/risk decision but not required (e.g. some PD-L1 tests).
• FDA Biomarker Qualification Program — three-stage pathway (21st Century Cures Act) qualifying a biomarker for a defined "context of use" reusable across programs.
• Enrichment (FDA guidance, final Mar 2019) — decrease heterogeneity; prognostic enrichment (likelier to have the event); predictive enrichment (likelier to respond by mechanism — the biomarker-driven, CDx-co-development case).

Part D — ML/AI, Causal HTE, Adaptive Trials, EHR Phenotyping & Immune Stratification

D1. Supervised ML for outcome/response prediction (the prognostic workhorse)

Penalized regression (glmnet, elastic-net signatures) → deep survival (DeepSurv, Cox-nnet) → multimodal fusion (attention-MIL PORPOISE, Cancer Cell 2022, PMID 35944502 — WSI + RNA-seq/CNV/mutation → per-patient hazard; high-attention regions containing TILs corroborate favorable prognosis in 12/14 cancers). The dominant story is failure-by-methodology, not algorithm: Roberts 2021 (Nat Mach Intell [10.1038/s42256-021-00307-0]) — of 2,212 COVID imaging-ML studies, none of 62 fully-reviewed were clinically usable (leakage, no external validation/calibration); Kapoor & Narayanan 2023 (Patterns [10.1016/j.patter.2023.100804]) — leakage in 294 papers across 17 fields. External validation + calibration + leakage control decide whether a stratifier is real.

D2. Causal inference / heterogeneous treatment effects (the predictive signal)

Estimate the CATE τ(x)=E[Y(1)−Y(0)|X] under unconfoundedness + overlap, then stratify by who benefits — distinct from D1 (which predicts outcome level, not treatment effect).

• Meta-learners — X-learner (Künzel 2019, PNAS [10.1073/pnas.1804597116]; robust under imbalanced assignment), R-learner (Nie & Wager 2021, Biometrika; quasi-oracle, Neyman-orthogonal).
• Forests — Causal Forests (Wager & Athey 2018, JASA; honesty → valid CIs) and Generalized Random Forests (grf; IV forests under unobserved confounding).
• Bayesian trees — BART (Hill 2011) and Bayesian Causal Forests (Hahn 2020; prognostic+treatment reparameterization removes regularization-induced confounding, fewer spurious subgroups).
• Neural — DragonNet (Shi 2019; targeted regularization → doubly-robust).
• Subgroup ID + honest intervals — Virtual Twins, SIDES/SIDEScreen (splits on a differential-effect criterion → recovers predictive biomarkers), conformal ITE intervals (Lei & Candès 2021; distribution-free coverage). Limit: all assume unconfoundedness + overlap (silently violated in RWD); subgroup methods risk multiplicity-driven false discovery.

D3. Adaptive & enrichment trial designs (the deployed oncology winner)

Governed by FDA Enrichment (2019), Adaptive Designs (2019), and Complex Innovative Trial Designs (2020) guidances (prospectively-planned adaptations, study-wide Type-I control, simulation reports). Methods: Adaptive Signature/Enrichment Designs (Freidlin & Simon; Simon & Simon 2013). Master protocols (Woodcock & LaVange 2017, NEJM [10.1056/NEJMra1510062]): basket (one therapy across diseases), umbrella (one disease, biomarker-matched arms), platform (perpetual). Landmark trials: VE-BASKET (Hyman 2015, NEJM — BRAF V600 across histologies; taught the falsifiable lesson biomarker ≠ histology-agnostic guarantee: colorectal did not respond), Lung-MAP, BATTLE (response-adaptive randomization), I-SPY 2 (Bayesian RAR within 8 biomarker signatures, arms "graduate"), NCI-MATCH (Flaherty 2020, JCO — 5,954 profiled, 37.6% actionable, ~26% assignable — the real-world yield of genomic matching at scale). Limit: predictive enrichment narrows the label + assumes a locked assay; Bayesian borrowing trades power for Type-I risk in discordant baskets.

D4. EHR / real-world phenotyping

Rule-based portable algorithms (eMERGE, Kho 2012 — replicated TCF7L2 across 5 EMRs; shared via PheKB) → PheWAS (Denny 2010) → unsupervised subtyping (TDA T2D 3 subtypes, Li 2015, Sci Transl Med; Deep Patient, Miotto 2016) → weakly-supervised (Anchor & Learn; APHRODITE/OHDSI) → federated learning (EXAM, Dayan 2021, Nat Med — 20 institutions, +16% AUC, +38% generalizability without pooling data) → EHR foundation models (Med-BERT, BEHRT, ETHOS — zero-shot health-timeline forecasting; ETHOS is npj Digital Medicine, not Nat Med). Limit: billing codes are noisy phenotype proxies; unsupervised subtypes need clinical validation; FMs inherit coding/site bias.

D5. Immune / immunotherapy stratification (converging on combinations, not single markers)

• Expression signatures — TIDE (Jiang 2018, Nat Med [10.1038/s41591-018-0136-1] — models T-cell dysfunction vs exclusion via CTL-interaction genes), GEP (Ayers, §B3), IMPRES (Auslander 2018 — 15 checkpoint-gene pairwise inequalities; ⚠️ contested: Carter 2019 matters-arising [10.1038/s41591-019-0671-4] showed it doesn't robustly replicate).
• Mutational/DNA-repair — TMB (Samstein, no universal threshold) and MSI-H/dMMR (Le, tissue-agnostic approval).
• irAE risk (an open problem) — Jing 2020 (Nat Commun) multi-omic prediction; ⚠️ no locked clinical-grade irAE biomarker exists (HLA, autoantibodies, IL-6/CRP, microbiome, TCR clonality all candidate-only).
• The incumbent & its limits — PD-L1 IHC is predictive on average but unreliable (SITC taskforce, Kluger 2020; Blueprint discordance). Independent benchmark (Litchfield 2021, Cell [10.1016/j.cell.2020.11.041]): clonal TMB + GEP + CXCL9 carry independent predictive information — combinations outperform singletons.

D6. Emerging (frontier with a credibility gap)

Digital twins (Unlearn PROCOVA prognostic covariate → ~9–26% sample-size reduction in a peer-reviewed AD trial; ⚠️ the circulated "35% smaller control arms" is vendor marketing), generative EHR models (ETHOS, Foresight), LLMs over records (⚠️ LLM phenotyping still needs expert refinement; Med-PaLM is QA-benchmarked, not real cohort stratification), and single-cell FMs (scGPT/Geneformer; ⚠️ independent benchmarks report zero-shot sometimes underperforms simple baselines — do not overstate).

What's Actually Adopted / Winning Now (the honest split)

Three throughlines:

1. The predictive-vs-prognostic discipline decides what licenses a treatment claim — effect-modification methods (causal forests, R-/X-learners, BCF, SIDES, adaptive enrichment), not plain outcome prediction.
2. The mature, adopted layer is single-analyte/genomic oncology biomarkers with FDA companion diagnostics + validated multigene assays, deployed inside adaptive master protocols.
3. The dominant failure mode across the ML/omics frontier is methodological, not algorithmic — leakage, no external validation, no calibration, arbitrary thresholds, bulk contamination, and stability mistaken for biology. The field is moving toward multimodal, causally-grounded, uncertainty-aware stratification embedded in adaptive trials — bottlenecked less by model capacity than by validation rigor and locked, assay-ready biomarkers.