Redesigning Preclinical Oncology: A Framework for Improving Clinical Translatability with AI, RECIST-Inspired Metrics, and Systems-Level Thinking

Despite decades of investment in oncology drug development, the translational pipeline from preclinical models to clinical success remains inefficient and error prone. A staggering number of therapies showing preclinical promise fail during early-phase clinical trials, typically due to unexpected toxicity or insufficient efficacy. These failures are frequently rooted in foundational shortcomings of the preclinical paradigm including inappropriate models, non-clinical endpoints, a disconnect between preclinical results and clinical expectations and single target focus of a complex disease.

This white paper proposes an integrated framework to enhance translational success by aligning preclinical oncology research more closely with human cancer biology. It rests on three synergistic pillars:

1. Clinically aligned efficacy metrics that mirror clinical oncology endpoints (e.g., RECIST-inspired measures such as complete responder, partial responders, tumor free survivors/cures)
2. Artificial intelligence (AI) to integrate multimodal data and model clinical predictiveness
3. Systems biology approaches to reflect tumor complexity and therapeutic interactions at a pathway and disease network level

Together, these innovations form a next-generation translational toolkit that prioritizes clinical relevance, model accuracy, and predictive analytics to transform early-stage oncology research.

Introduction: The Translational Bottleneck in Oncology

In oncology drug development, in vivo animal models are critical for early decision-making. Yet only 5–10% of preclinical oncology agents that demonstrate in vivo efficacy eventually progress to clinical approval. Root causes of this poor translatability include:

• Overdependence on simplistic tumor volume metrics (e.g., %TGI) that fail to capture patient-level response dynamics.
• Non-representative tumor models, such as standard single model subcutaneous xenografts, which inadequately replicate the tumor microenvironment and disease heterogeneity.
• Linear, single-target hypotheses that disregard the networked, adaptive nature of cancer biology.

A paradigm shift is needed, one that rethinks preclinical model selection, data interpretation, and endpoint prioritization through a clinically grounded, systems-aware, and data-integrative lens.

Pillar I: Clinically Aligned Metrics, Integrating CR, PR, and TFS into Preclinical Studies

Current Limitation:

Traditional preclinical endpoints, like percentage tumor growth inhibition (TGI), fail to reflect how oncologists assess response in patients, which is often based on RECIST 1.1 criteria (e.g., complete/partial responses, stable/progressive disease). These standard metrics also mask individual response variation, that is critical for precision oncology.

Proposed Solution:

Alignment of RECIST-like categorical response metrics within preclinical settings, including:

• Complete Response (CR): Disappearance of measurable tumor
• Partial Response (PR): ≥50% reduction in tumor volume
• Stable Disease (SD): Minimal change without progression
• Progressive Disease (PD): ≥20% increase or emergence of new lesions
• Tumor-Free Survival (TFS): Duration of complete remission

Benefits:

• Enhances comparability between preclinical and clinical trial outcomes
• Enables durability assessments, including tumor regrowth and treatment resistance
• Provides granular insight into individual animal responses beyond cohort averages

Implementation Guidelines:

• Define thresholds using volumetric and imaging-based tracking (e.g., MRI, ultrasound, calipers)
• Apply metrics across clinically all models: subcutaneous, orthotopic, patient-derived xenografts (PDX), and humanized mice
• Use longitudinal data capture for patterns of response and relapse

Pillar II: AI-Driven Prediction and Model Assessment

Current Limitation:

Preclinical datasets are often fragmented, high-dimensional, and underutilized. Conventional analysis methods lack the power to detect complex, non-linear relationships between biomarkers, treatment, and outcomes particularly in multimodal datasets.

Proposed Solution:

Use AI and machine learning (ML) tools to:

• Predict translational likelihood by learning from historical compound performance (negative and positive data must be included)
• Evaluate model fidelity by comparing mouse-human molecular profiles (e.g., transcriptomics, metabolomics, proteomics)
• Discover predictive biomarkers from integrated datasets, including imaging, omics, and pharmacokinetics

Key Use Cases:

• Outcome prediction: Train models on preclinical–clinical datasets to simulate patient responses
• Model matching: Use clustering and embedding algorithms to score how closely a model mimics human cancer (e.g., TCGA similarity analysis)
• Digital twin development: Create computational analogs of animal models or patients for in silico testing of treatment sequences and resistance mechanisms

Pillar III: Systems Biology and Network-Informed Targeting

Current Limitation:

Many experimental drugs fail not because the target is unimportant, but because tumors continuously adapt and evolve via redundant or compensatory pathways. Traditional single pathway selective strategies often miss the forest for the trees.

Proposed Solution:

Leverage systems biology and network modeling to:

• Identify key signaling nodes, feedback loops, and resistance pathways
• Prioritize multi-target or synthetic lethal combinations
• Analyze tumor–immune–stromal interactions, which shape response and resistance

Implementation Guidelines:

• Employ graph-based and AI-enhanced pathway mapping to assess pathway crosstalk and redundancy
• Design combinatorial in vivo studies reflecting molecular synergies
• Integrate clinical resistance datasets to anticipate tumor escape mechanisms

A Unified Translational Framework

Recommendations for Implementation

To operationalize this framework in real-world R&D environments:

1. Standardized data capture protocols across studies, including imaging frequency, metadata annotations, and endpoint definitions.
2. Mandate RECIST-inspired metrics as standard for preclinical efficacy reporting.
3. Incorporate multi-omics (RNA-seq, metabolomics, proteomics, IHC) into routine model characterization.
4. Develop preclinical models/assays that reliably mimic clinical screening tools (i.e. preclinical CTNDA).
5. Develop AI workflows capable of cross-platform integration and translational scoring.
6. Reprioritize model selection based on fidelity to patient tumors, not just logistical ease.

Conclusion: AI and Alignment, Using all the tools

The challenge in translational oncology is not simply the volume of preclinical data or the number of models, but the lack of alignment between preclinical systems and clinical reality. We must continue to evolve from traditional research practices to a data-intergraded paradigm, that is biologically informed, and clinically relevant.

By adopting clinically aligned efficacy metrics, deploying AI to model translational risk, and embracing systems-level biology, we can design smarter experiments that better reflect human disease. This will not only reduce attrition but also accelerate the journey of effective therapies from bench to bedside, ultimately benefiting the patients who need them most.