Model | SYPHU-CHINA

Platform Demo

Tip: the host may sleep when idle. If you see a wake/activate button, click it; first load can take a moment.

Tip: All figures in this chapter are click-to-zoom. Click any image to view it in full size; click the dark background to exit.

Integrated Modeling of ATRA-Driven Core Network Reprogramming to Reverse Malignant States in HCC

Abstract

Background: Hepatocellular carcinoma (HCC) remains a leading cause of cancer mortality. All-trans retinoic acid (ATRA) shows promise for forcing malignant cells toward differentiated, less aggressive states.
Methods: Multi-omics integration (bulk & single-cell), computational modeling, and targeted experimental validation.
Findings: Identification of key ATRA targets and regulatory circuits underpinning state reversal and selective vulnerability.
Innovation: A mechanism-to-therapy framework spanning molecular logic, delivery design, and combination strategies.

~800k

Cells integrated (multi-cohort)

>2,000

Robust HVGs for network inference

Inferred regulatory modules (core GRN)

Reviewer note: We prioritize interactive intuition over dense math. Equations are used sparingly; wherever possible, we expose assumptions via toggles, tabs, and dynamic annotations to make the logic transparent.

1.1 Current Status & Challenges in Liver Cancer Therapy

Globally, HCC is among the most prevalent and lethal malignancies. Late diagnosis, high heterogeneity, and limited durability of current treatments (resection/ablation, TACE, TKIs, immunotherapy) leave substantial unmet needs. Therapeutic reprogramming aims to shift cell states toward less aggressive, more differentiated phenotypes.

HCC burden clusters in regions with chronic hepatitis prevalence and metabolic risk. The integrated atlas allows us to compare tumor vs. normal hepatocytes across donors and cell lines, improving generalization.

Regions are illustrative to demonstrate the dashboard layout; exact epidemiology plots are provided in Chapter 2.

Surgery and locoregional therapies suffer recurrence; TKIs and immunotherapy face resistance, toxicity, and patient stratification challenges. The outcome is underpinned by intratumoral heterogeneity.

Modality	Strength	Limitation	Implication
Resection/Ablation	Local control	High recurrence	Microresidual disease
TACE	Bridging/palliative	Hypoxia-driven escape	Adaptive stress responses
TKIs	Pathway blockade	On-target/off-target AEs	Limited durability
Immunotherapy	Long tails	Variable response	Biomarker need

The reprogramming paradigm aims to stabilize malignant ecosystems by nudging cells toward differentiated fates, thereby reducing proliferation, invasion, and immune evasion pressures—ATRA is a prime candidate to achieve this.

Why reprogramming may generalize better than cytotoxicity

Cytotoxic approaches select for resistant clones. Reprogramming reduces selective pressure by redirecting the state rather than eliminating cells—potentially improving durability and synergy with standard of care.

Interactive “What-if” · Recurrence Reduction with ATRA-guided Reprogramming

Use the slider to hypothesize an average 20% recurrence reduction among high-risk subclones. The narrative updates to illustrate expected benefits and where combination therapy could help.

A 20% reduction could shift early recurrence curves downward, easing pressure on salvage therapy and enabling longer windows for immune engagement. Clusters with high cell-cycle scores would benefit most.

Hypothesized reduction

Illustrative slider; formal modeling is presented in later chapters.

1.2 Biological Basis of ATRA

Retinoic acid (RA) signaling through RAR/RXR modulates chromatin accessibility and transcriptional programs governing differentiation, apoptosis, and cell-cycle control. In acute promyelocytic leukemia (APL), ATRA induces lineage maturation and durable remission—proof that transcriptional rewiring can be therapeutic. In solid tumors, context-dependent efficacy implies that core regulatory networks (GRNs) determine responsiveness, motivating our network-first analysis.

Minimal math, maximum intuition: Instead of long derivations, we build a core-network map and test whether ATRA-sensitive modules (e.g. differentiation, stress, apoptosis) are controllable under realistic perturbations and delivery constraints.

1.3 Objectives & Scientific Questions

Core hypothesis: ATRA can reverse malignant phenotypes in HCC by reprogramming core regulatory networks that govern differentiation and survival, yielding tractable design rules for combinations and delivery.

Question	Operationalization	Readout
Q1: Who are the actionable targets?	Infer TF→target circuits, score ATRA receptors & downstream modules	Ranked targets / modules; testable hypotheses
Q2: Can we predict responsiveness?	Integrate single-cell states with graph-aware predictors	Response scores; state-transition likelihoods
Q3: How to design therapy?	Combine with SoC; optimize delivery (e.g., Lipo-ATRA)	In-silico prioritization; validation plan

Model · Technical Roadmap (PDF)

Open in new tab Download PDF

Tip: If the PDF does not render in your browser, use “Open in new tab” or “Download PDF”.

Tip: All figures in this chapter are click-to-zoom. Click any image to view it full size; click the dark backdrop to exit.

Chapter 2 · Materials & Methods

2.1.1 Data Sources & Characteristics

We assembled a comprehensive liver single-cell compendium by integrating authoritative resources: LaminDB ARC virtual atlas (multi-omics reference), hepatoma cell lines (HepG2, Huh7, PLC/PRF/5, Hep3B), normal liver references (THLE-2 and HHSteC) and the Tabula Sapiens liver atlas. This mixture provides tumor vs. normal contrasts and population diversity necessary for robust modeling.

Dataset	Cells	Genes	Batches	Source
LaminDB ARC	~800,000	~20,000	Multiple	Integrated database
Hepatoma cell lines	~400,000	~18,000	4 lines	Lab-curated
Tabula Sapiens (liver)	~300,000	~19,500	Multi-donor	Public

2.1.2 Quality Control Models

To ensure reliable downstream analysis, we applied distribution-aware QC. Genes were kept if expressed in at least:

\[ \max\left\{100,\;0.001\times N_{\text{total}}\right\} \tag{2.1} \]

The mitochondrial fraction identifies stressed/low-quality cells:

\[ \mathrm{MT}_{\text{ratio}} \;=\; \frac{\sum_{i=1}^{n_{\text{mt}}} C_{\text{mt},i}}{\sum_{j=1}^{n_{\text{total}}} C_{\text{total},j}} \times 100\% \tag{2.2} \]

QC (a) Gene count histogram

Histogram of detected genes per cell before/after filters. The right-shift after QC indicates removal of low-complexity droplets while preserving informative cells.

QC (c) Library size distribution

We normalize library depth and stabilize variance with the following:

\[ \hat{C}_{ij}=\frac{C_{ij}}{\sum_{k=1}^{m} C_{ik}}\times 10^{4} \tag{2.3} \] \[ X_{ij}=\ln\!\left(\hat{C}_{ij}+1\right) \tag{2.4} \]

Raw counts (pre-normalization) vary by orders of magnitude across cells; scaling plus log1p reduces this bias.

2.1.4 Highly Variable Gene (HVG) Selection

Per-gene mean and dispersion are estimated and filtered as:

\[ \sigma_g^2 = \frac{1}{n}\sum_{i=1}^{n} \big(x_{g i}-\mu_g\big)^2 \quad\text{with}\quad 0.0125\le \mu_g\le 3\;\; \text{and}\;\; \sigma_g^2>0.5 \tag{2.5} \]

HVG (a) Mean–dispersion scatter

Points above the fitted trend (red) mark HVGs driving downstream clustering and state inference.

HVG (b) HVG count across datasets — filmstrip (28 panels)

Browse per-dataset HVG barplots. Consistent HVG yield across cohorts supports cross-study integration.

2.1.5 Batch-Effect Correction

We constructed batch-balanced KNN graphs in PCA space (BBKNN) and also used Harmony for MAP-based embedding realignment. Together they remove technical offsets while preserving biological structure.

Before / After UMAP (colored by batch)

Left: strong batch segregation. Right: batches mix well after correction while clusters remain distinct, indicating successful removal of technical variance.

Batch mixing score

Quantitative improvement in neighbor-mixing validates integration across cohorts and platforms.

2.1.6 Cell-Cycle Scoring

Using curated S and G2/M gene sets, each cell receives standardized stage scores:

\[ S_{\text{phase}} \;=\; \frac{1}{|G_s|}\sum_{g\in G_s}\frac{E_g-\mu_g}{\sigma_g} \tag{2.6} \]

Cell-cycle stages on embeddings

Spatial localization of S/G2M-rich clusters clarifies proliferative niches within hepatoma populations.

S and G2/M score distributions

Violin plots show stage-specific heterogeneity and identify outlier proliferative subclones.

Cell-type specific cycle features

Cycle activity differs between tumor and normal lineages, guiding downstream adjustment in differential analysis.

2.1.8 Output Validation

Each preprocessing step produces QC reports and figures for auditability; processed objects (e.g., analyzed_data.h5ad) are persisted for reproducibility.

Preprocessing overview

End-to-end flow: input → QC → normalization → batch correction → validation exports.

2.2.2 ATRA Receptor Expression Landscape

Composite receptor score per cell:

\[ E_{r,c}=\frac{1}{n_r}\sum_{g\in\{RARA,RARB,RARG,RXRA,RXRB,RXRG\}} \frac{X_{g,c}-\mu_g}{\sigma_g} \tag{2.7} \]

ATRA receptor expression profiles

Violin and heat subpanels summarize receptor distributions and co-expression, highlighting responsive niches.

2.2.3 Differential Expression & Pathway Enrichment

Ranking and significance:

\[ \log_2\!\left(\frac{\mu_{g,\text{hepatoma}}+\epsilon}{\mu_{g,\text{normal}}+\epsilon}\right), \qquad p_{\text{adj}}<0.05 \tag{2.8} \]

Volcano gallery — filmstrip (30 panels)

Each panel corresponds to a specific contrast/statistical pipeline. Red/blue lobes highlight up/down-regulated genes.

Top KEGG pathways

Bar chart of top enriched KEGG terms among significant DEGs; liver/cancer-related axes dominate.

2.2.4 Random-Forest Targeted Delivery Model

Classifier and aggregation:

\[ X\in\mathbb{R}^{n\times p},\quad y\in\{\text{hepatoma},\text{normal}\},\qquad \hat y = \mathrm{mode}\left\{T_b(X)\right\}_{b=1}^{B},\; B=50 \tag{2.9} \]

Gini-based importance ranks candidate targets for Lipo-ATRA.

Feature importance ranking

Top genes with largest impurity reduction; these guide selective knock-down in hepatoma.

Model performance

Stratified train/test with cross-validation demonstrates stable accuracy and balanced precision/recall.

ROC curve

AUROC summarizes separability between hepatoma and normal compartments.

2.2.5 Lipo-ATRA Treatment Simulation

Target scoring and tumor-selective downregulation:

\[ S_g = w_1\,\mathrm{Importance}_g + w_2\,E_{g,\text{hepatoma}} - w_3\,E_{g,\text{normal}} \tag{2.10} \] \[ X^{\text{post}}_{g,c} = \begin{cases} 0.5\,X^{\text{pre}}_{g,c}, & c \in \text{hepatoma}\ \land\ g=g_{\text{target}} \\ X^{\text{pre}}_{g,c}, & \text{otherwise} \end{cases} \tag{2.11} \]

Per-gene expression change — filmstrip (28 panels)

Each slide shows a target’s pre/post distribution across cell types, demonstrating selective tumor suppression.

Box/Violin summaries — filmstrip (28 panels)

Distributional summaries quantify heterogeneity and the fraction of responsive cells across compartments.

Response statistics

Proportion of cells meeting response thresholds per lineage/condition guides dosage and co-therapy design.

2.2.6 Mechanism-of-Action Network

Post-treatment DEG thresholds and effect size:

\[ p_{\text{adj}}<0.01,\qquad \left|\log_2\mathrm{FC}\right|>1.5 \tag{2.12} \]

Mapped to liver/cancer-relevant pathways to generate a hypothesis network for ATRA-mediated reprogramming.

(a) Heatmap of responsive genes

Hierarchical clustering reveals coordinated modules consistent with differentiation and survival rewiring.

(b) Interaction networks

Multi-view networks (TF–target, pathway crosstalk, PPI) converge on core hubs potentially mediating ATRA response.

Proposed cascade: receptor activation → chromatin remodeling → lineage re-differentiation and growth restraint.

2.4.6 Integrated Validation Metrics

Integrated validation dashboard

Composite scores summarize sequence/structure consistency, docking reliability, intervention predictability, biological interpretability, and reproducibility.

2.4.7 Reproducibility & Robustness

Robustness analyses

Sensitivity heatmaps and cross-validation agreement indicate stable conclusions across parameter sweeps.

2.5.1 Quantifying Cell Behavior

Trajectory definition (centroid linkage):

\[ \mathbf{p}_i(t)=(x_i(t),y_i(t)) \tag{2.13} \]

We compare constant-velocity, quadratic, and AR(1) extrapolations for smoothing and short-horizon prediction.

Trajectory & extrapolation

Overlays show raw detections (green), smoothed trajectories (colored) and extrapolated segments (blue), confirming continuity.

MSD curves & anomalous diffusion

Log–log fits estimate the diffusion exponent \( \alpha \) in \( \mathrm{MSD}(\tau)=C\tau^\alpha \); deviations from \(\alpha=1\) indicate directed or confined motion.

Behavior clusters

PCA layouts colored by DBSCAN clusters reveal distinct modes (directed migration, random walk, restricted motion).

Cluster trajectory exemplars

Representative tracks for each cluster type; contrasts highlight persistence vs. tortuosity differences.

Parameter correlations & network

Spearman matrices and the derived network expose strong couplings (e.g., straightness vs. MSD slope).

Population velocity fields & density maps

Arrows depict interpolated velocity from Delaunay triangulation; co-located density hot-spots suggest collective flow.

2.6 Engineering Design

Explore Engineering

ATRA biosynthetic pathway

GGPP→Phytoene→Lycopene→β-Carotene→ATRA with key enzymes (crtE, crtB, crtI, crtY); ODE/kinetics support flux allocation.

Pareto front for plasmid optimization

Multi-objective trade-off among GC deviation, CAI, and secondary-structure stability.

Dose–response system curves

Hill-type response surfaces with EC50 annotations guide induction and promoter/RBS choices.

Restriction site density map

Candidate MCS segments balance cut-site richness and compact length to facilitate cloning strategies.

Protein physicochemical properties

Distributions of molecular weight, pI and instability index help screen stable constructs for expression.

System parameter sensitivity

Global sensitivity map identifies parameters with largest influence on pathway output (priority for tuning).

Chapter 3 · Results

All figures are zoomable — click any image to view it in a lightbox.

3.1 Heterogeneity Atlas of Liver Cancer Cells

Faceted visualizations summarize cellular diversity, malignant program activation, and cycling states.

3.1.1 Cell-State Diversity

We identify major subpopulations using graph-based clustering on integrated scRNA-seq embeddings. Subcluster-specific marker genes (e.g., AKR1B10, EPCAM, ALB) delineate malignant, progenitor-like, and hepatocyte-like states. Heterogeneity is quantified via intra-/inter-cluster dispersion and silhouette indices.

Subcluster A Subcluster B Subcluster C Subcluster D

Figure 3.1A (demo): UMAP summarizing cell-state diversity. Toggle chips to show/hide subclusters and inspect separations.

Figure 3.1B: Tabula Sapiens liver UMAP anchors major hepatic lineages and supports annotation transfer.

**Figure 3.1C:** t-SNE by drug exposure; integration limits residual drug-driven separation.

t-SNE by phase — **Figure 3.1D:** t-SNE by cell-cycle phase; proliferative bands outline S / G2M-rich clusters.

t-SNE by sample — **Figure 3.1E:** t-SNE by donor/sample; good post-correction mixing across cohorts.

**Figure 3.1F:** t-SNE by species confirms human-only analysis footprints.

Figure 3.1G: Malignant-only UMAP showing distinct proliferation, stress and differentiation signatures.

3.1.2 Malignant-State Features

Transcriptome contrasts between HCC and normal hepatocytes highlight up-regulated cell-cycle/DNA-repair programs and reduced hepatocyte metabolism. Cell-cycle staging reveals proliferative bias in specific subclusters.

Figure 3.1H (demo): DEG categories reflect proliferative and repair programs in HCC.

Figure 3.1I (demo): G1/S/G2M distribution; malignant clusters bias toward S/G2M.

3.2 ATRA Target Expression

Retinoic acid receptor isoforms (RAR/RXR) show enriched co-expression in malignant subclusters with high cycling activity, motivating receptor-guided delivery and rational combinations.

Figure 3.2A: Composite visualization of receptor expression across states and compartments.

3.3 Core Regulatory Network Identification

Using feature selection and graph metrics over the integrated atlas, we prioritize hepatoma-specific regulators and modules.

Figure 3.3A (demo): Importance ranking of the 10 most informative genes.

Figure 3.3B (demo): Cross-validated performance separating malignant vs. normal states.

3.4 Mechanistic Validation

3.4.1 Molecular Docking (narrative)

\[ \mathrm{Affinity}_{\mathrm{ATRA}\rightarrow r}\;\propto\;\sum_{k\in\mathrm{contacts}} w_k\,\Delta G_k \] (3.4)

Prior literature supports plausible ATRA binding within RAR/RXR pockets (conserved residues, pocket geometry). Without a provided docking figure in the current assets, conclusions here remain text-only and consistent with receptor expression maps in Section 3.2.

3.4.2 Virtual Intervention (Lipo-ATRA)

\[ X^{\mathrm{post}}_{g,c} \;=\; (1-\alpha)\, X^{\mathrm{pre}}_{g,c},\qquad c \in \mathrm{hepatoma},\; g \in \mathcal{T} \] (3.5)

We simulate receptor-informed down-regulation and pathway inactivation; higher receptor scores predict stronger reprogramming toward less aggressive states.

Lipo-ATRA virtual intervention — **Figure 3.4A:** Predicted suppression of cycle programs and gain of hepatocyte markers.

**Figure 3.4B:** Simulation-derived panel showing module-level shifts under intervention.

3.5 Experimental Validation

Behavior assays reveal reduced migration persistence and altered trajectory patterns after ATRA exposure.

Trajectory snapshot 1 — **Figure 3.5A:** Representative trajectories show shorter persistent runs post-treatment.

Trajectory snapshot 2 — **Figure 3.5B:** Turning-angle distributions broaden, indicating more diffusive motion.

3.6 Therapeutic System Design

Plasmid and delivery engineering focus on maximizing on-target reprogramming in receptor-high malignant niches while sparing receptor-low hepatocytes.

Engineering design panel 1 — **Figure 3.6A:** Design trade-offs among CAI, GC window, and payload stability.

Engineering design panel 2 — **Figure 3.6B:** Assembly plan and component compatibility map.

Figure 3.6C: End-to-end pipeline schematic with QA gates and release criteria.

3.6 Plasmid ORF / Protein Analysis — Visual Summary

Interactive exploration of predicted ORFs and protein properties derived from the plasmid file. Hover points to inspect details; click table headers to sort. Filters update all charts.

—

Total ORFs

—

Stable ORFs (%)

—

Median pI

—

Mean CAI

Stability: Length (aa) ≥ 0

MW vs pI (dot size = length aa)

Color: Stable Unstable

Larger circles indicate longer ORFs. Clusters in mid-pI range with moderate MW often correspond to more stable predictions.

Length Distribution (aa)

Binned ORF lengths. Use the slider above to focus on longer CDS candidates.

Top ORFs by Codon Adaptation Index (CAI)

High CAI suggests host-preferred codon usage (putative expression advantage).

ORF Table

Download CSV

ORF	Start	End	Strand	Length (aa)	MW (Da)	pI	Instability	Stability	CAI

Click headers to sort. Filters above apply here as well.

Chapter 4 · Discussion

4.1 Integrated View of Major Findings

We synthesize evidence from expression mapping, network modeling, virtual intervention, and cell-behavior validation to provide a unified interpretation of how ATRA drives malignant-to-differentiated transitions in hepatocellular carcinoma.

\[ \mathcal{W}_{\text{ATRA}} \;=\; \sum_{\ell \in \{\text{scRNA, net, ML, sim, beh}\}} w_\ell \, z_\ell \quad\text{with}\quad \sum_\ell w_\ell = 1 \tag{4.1} \]

Interactive evidence filter: toggle facets to highlight how conclusions are supported across layers.

scRNA-seq Regulatory network ML feature ranking Simulation Cell behavior

Conclusion	scRNA-seq	Network	ML	Simulation	Behavior
ATRA-responsive subclusters align with high RAR/RXR expression.	✓	✓	✓	✓	—
Core circuits controlling cycle-to-differentiation shifts can be perturbed.	—	✓	✓	✓	—
Virtual knockdowns predict reduced proliferation, matching behavior assays.	—	—	—	✓	✓
Hepatocyte-like programs re-emerge under receptor-guided delivery.	✓	✓	—	✓	—

Figure 4.1: Visual summary of agreement strengths across layers (illustrative). Dark bars indicate overlapping support from multiple analyses.

Compared to traditional views of ATRA as a differentiation agent primarily in hematologic malignancies, our integrated evidence suggests that solid-tumor contexts like HCC can also exhibit receptor-guided responsiveness—provided that the delivery and combination design respects the tumor’s heterogeneous microenvironments.

4.2 Biological Significance

The reprogramming effect arises when receptor activation suppresses proliferative circuits and relieves brakes on hepatocyte-like functions. Microenvironmental cues (hypoxia, stromal interactions, cytokines) modulate response heterogeneity.

\[ \text{Shift}_{\text{state}} \;=\; \beta_0 \;-\; \beta_1 \,\Pi_{\text{prolif}}(p) \;+\; \beta_2 \,\Phi_{\text{hep}}(p) \;-\; \beta_3 \,\Sigma_{\text{stress}}(p) \tag{4.2} \]

Interactive microenvironment model: drag the slider to vary microenvironmental pressure and preview expected shifts.

Pressure: 35%

Proliferation

Hepatocyte functions

Stress/Inflammation

With moderate pressure (35%), proliferation remains dominant but starts to decline as receptor signaling competes with cytokine-driven growth.

4.3 Methodological Innovations

Multi-omics integration strategy

We align bulk and single-cell layers, unify gene identifiers, and infer cross-modal modules to expose circuits otherwise obscured by batch and platform differences. This provides a stable substrate for downstream causal modeling.

Closed-loop validation between computation and experiment

Predictions (target sets, directionality of change) are iteratively tested in behavior assays; discrepancies feed back into model refinement (feature selection, priors, constraints), tightening inference over time.

Systematic engineering design

Delivery and combination choices are mapped to receptor landscapes and safety constraints, yielding actionable, modular design rules.

Figure 4.3: Closed-loop workflow: integration → inference → simulation/design → behavior validation → model refinement.

4.4 Limitations and Outlook

Our analyses are bounded by data scale/coverage and by the depth of mechanistic probing achievable in current assays. Translational gaps remain between simulated reprogramming and durable clinical responses. We outline risks and mitigation below.

\[ \mathcal{R}_k \;=\; \text{Impact}_k \times \text{Likelihood}_k \quad\text{and}\quad \text{Priority} \;=\; \operatorname*{arg\,max}_k \mathcal{R}_k \tag{4.3} \]

Figure 4.4: High-priority risks cluster around delivery specificity, toxicity control, and generalizability, guiding the next stage of experiments and safety design.

Pick your near-term priority (updates the action list):

Delivery specificity Toxicity profiling Generalizability

Profile receptor landscapes at single-cell resolution in additional HCC cohorts.
Benchmark off-target uptake in hepatocyte-like neighborhoods with low receptor scores.
Stress-test dose scheduling to maximize on-target reprogramming.

Chapter 5 · Conclusions & Outlook

5.1 Main Conclusions

Reprogramming model established: a structured framework that links receptor landscapes to state transitions in HCC.
Key regulatory nodes identified: features and modules prioritized by network-aware ML strategies.
Effects validated from molecules to cells: simulations and behavior assays converge on cycle↓ / differentiation↑.
Feasible delivery system designed: receptor-guided Lipo-ATRA with rules for dosing and combinations.

\[ \text{Readiness} \;=\; w_{\text{cov}}\cdot C \;+\; w_{\text{rob}}\cdot R \;+\; w_{\text{trl}}\cdot T, \qquad w_{\text{cov}}+w_{\text{rob}}+w_{\text{trl}}=1 \]

Interactive confidence dashboard (adjust the sliders; the overall readiness updates):

Data coverage

Current: 60%

Model robustness

Current: 65%

Translational readiness

Current: 50%

Figure 5.1: Dashboard aggregates coverage, robustness, and translational readiness into an overall score (illustrative).

5.2 Scientific Contributions

Contribution	Highlights	Utility
Theory	Mechanism of state conversion via receptor-guided reprogramming	Clarifies how differentiation signals counteract malignant circuits
Method	Integration → network inference → simulation → closed-loop validation	Reusable blueprint for solid-tumor reprogramming studies
Application	Lipo-ATRA design rules; receptor-aware dosing/combination	Translational path for HCC with safety-aware targeting

5.3 Future Directions

We prioritize preclinical validation and personalization. Early animal studies define efficacy/toxicity envelopes, while patient-stratification rules refine delivery and combinations to match receptor landscapes and microenvironmental pressures.

\[ \max_{\;\text{schedule}}\; \mathbb{E}[\Delta\text{Diff}] \;-\; \lambda_1 \mathbb{E}[\text{Toxicity}] \;-\; \lambda_2 \mathbb{E}[\text{Off\text{-}target}] \]

Personalization sandbox: toggle receptor landscape and microenvironment pressure; the recommended strategy updates.

Receptor landscape Microenvironment pressure

Receptor-guided Lipo-ATRA dosing; monitor off-target uptake.
Combine with cycle dampeners if proliferative niches persist.
Schedule pulses to consolidate differentiation gains.

Mixed receptors & moderate pressure: balance efficacy with safety; consider adaptive scheduling.

Appendix · Data & Model Gallery

Data DOI: 10.5281/zenodo.17259332

All gallery images are click-to-zoom. Use ←/→ on desktop; swipe on mobile.

A. Data Source & Contents

Item	Content	Notes	Source
Original model images	UMAP / t-SNE / QC / mechanism & engineering schematics	Base visuals used across Chapters 2–3	DOI: 10.5281/zenodo.17259332
Analysis artifacts	HVG, volcano sets, simulated interventions, behavior panels	Intermediate products (non-personal)	Same as above
Plasmid dataset	Fusion ATRA plasmid key elements & constraints	See Chapter 3, Section 3.6	Same as above