UNI2-h-DSS: Pan-Cancer Prognostic Dataset

• UNI2-h-DSS is a comprehensive pan-cancer dataset curated from TCGA whole-slide images that benchmarks zero-shot prognostic model transfer across 13 cancer types.
• It leverages the UNI2-h foundation model for patch-level feature extraction, ensuring high-quality curation and sufficient event rates for survival analysis.
• The dataset underpins evaluations of intra- and inter-tumor prognostic signal transfer using ABMIL-based models and the advanced MoE-PKT approach.

UNI2-h-DSS is a large-scale, pan-cancer dataset curated explicitly for the systematic paper of prognostic knowledge transfer in computational pathology. Constructed from TCGA whole-slide images (WSIs) and leveraging the foundation model UNI2-h to extract patch-level features, the dataset integrates 8,818 WSIs from 7,268 patients, representing 13 cancer types, after stringent curation to ensure data quality and sufficient event rates. UNI2-h-DSS serves as a comprehensive benchmark for evaluating and quantifying the zero-shot transferability of survival models across disparate cancer types, enabling detailed investigation into the intra- and inter-tumor factors that modulate prognostic information generalizability and the development of adaptive multi-cancer predictive frameworks.

1. Dataset Curation and Composition

UNI2-h-DSS was curated from a large initial pool of TCGA WSIs, comprising 11,646 slides from 9,547 patients. To ensure suitability for prognostic modeling and generalizable conclusions, the following steps were undertaken:

• Extraction: Patch-level features for all WSIs were generated using the state-of-the-art UNI2-h foundation model.
• Filtering: Patients with unknown disease-specific survival (DSS) were excluded, yielding 11,188 WSIs from 9,190 patients.
• Cohort Requirements: Cancers with fewer than 200 patients or an event (death) occurrence rate below 5% were removed, enabling robust model training and evaluation.
• Final Composition: The resulting dataset contains 8,818 WSIs from 7,268 patients, covering 13 cancer types. This diverse representation supports the exploration of both common patterns and tissue-specific prognostic signals.

The curation process was explicitly designed to enable reliable cross-cancer benchmarking of prognostic models and ensure adequate data coverage for rare and common cancer types alike.

2. Benchmarking Prognostic Knowledge Transfer

UNI2-h-DSS is utilized as the primary testbed for systematic assessment of transferability in WSI-based prognostic models. The methodology follows a rigorous experimental protocol:

• For each cancer $c_i$ in the set $\mathcal{C} = \{c_1, c_2, \ldots, c_{13}\}$, an ABMIL-based multiple instance learning (MIL) model is trained using 5-fold cross-validation to yield cancer-specific predictors.
• For knowledge transfer, a model trained on a source cancer $\mathcal{S}$ is evaluated directly on a target cancer $\mathcal{T}$ in a zero-shot manner, i.e., without fine-tuning on the target data.
• Transfer performance $P_{(\mathcal{S} \rightarrow \mathcal{T})}$ is quantified via the concordance index (C-Index), representing the degree of transferable prognostic signal.
• A transfer is considered positive if $P_{(\mathcal{S} \rightarrow \mathcal{T})} > 0.5$, exceeding the random guess baseline.

This cascade supports systematic measurement of shared prognostic information, yielding a matrix of transferability scores across all cancer pairs.

3. Experimental Design and Transfer Assessment

The experimental framework for evaluating transferability consists of two stages:

1. Cancer-Specific Model Training: Each cohort is used to fit a separate ABMIL-based model, optimized and validated internally using its own data.
2. Zero-Shot Model Transfer: Models are deployed (fixed weights) to predict survival outcomes in the remaining 12 cancer cohorts, with no adaptation or additional supervision.

This structure isolates the transferrable prognostic information encoded by each source model, free from confounding target-specific tuning. The use of zero-shot evaluation establishes a lower bound for transferability and highlights intrinsic cross-cancer prognostic commonalities or divergences.

4. Determinants of Prognostic Knowledge Transfer

Empirical analyses on UNI2-h-DSS reveal that transfer performance between cancer pairs is modulated by both intra-task and inter-task variables:

• Intra-task Factors:
  • $P_{(\mathcal{S})}$: Source model quality. Higher C-Index on the source cancer correlates with superior transfer outcomes.
  • $P_{(\mathcal{T})}$: Target task difficulty. Prognostically complex target cancers (lower inherent predictability) receive less transferable gain even from strong source models.
• Inter-task Factors:
  • $$C_{(\mathcal{S} \rightarrow \mathcal{T})}^{\text{(RMST)}}$$: Closeness in tumor invasiveness, quantified by restricted mean survival time (RMST) within 10 years.
  • $$C_{(\mathcal{S} \rightarrow \mathcal{T})}^{\text{(Dist.)}}$$: Closeness in data distribution, captured via feature projections such as t-SNE, indicating histopathological similarity.

Ordinary least squares (OLS) regression demonstrates that these (especially the intra-task) factors significantly explain observed variation in transferability. This suggests optimizing source model accuracy and considering source-target biological similarity are pivotal to effective transfer.

5. The MoE-PKT Approach

Recognizing the limitations of naive cross-cancer transfer, the MoE-PKT (Mixture-of-Experts based Prognostic Knowledge Transfer) method was introduced to optimally combine prognostic signals from multiple cancers:

• Framework: Each cancer-specific model, including the target’s own, operates as an "expert." For any WSI represented as $X = \{x_i\}_{i=1}^M \in \mathbb{R}^{M \times d}$, a router module (ABMIL + MLP) computes scores $\{w_t\}_{t=1}^{N}$—one per expert.
• Expert Selection: The router selects the top $K$ experts (typically $K=5$). Each expert consists of a frozen MIL encoder ($E_t$) and a trainable MLP adapter.
• Aggregation and Prediction: The chosen experts’ outputs are aggregated, yielding a fused representation used by a final fully connected layer to predict survival.

The end-to-end formalism is $$X \rightarrow \text{Router} \rightarrow \{w_1, \ldots, w_N\} \rightarrow \text{Select top-}K \text{experts} \rightarrow \text{Aggregate} \{E_t(X)\}_{\text{top-K}} \rightarrow \text{Prediction}$$. Empirically, MoE-PKT achieves higher overall C-Indexes than single-cancer models or unweighted transfer, demonstrating effective leveraging of complementary and non-overlapping knowledge from multiple sources.

6. Impact and Applications

The establishment of UNI2-h-DSS enables, for the first time, systematic, quantitative investigation into prognostic knowledge transfer across a broad spectrum of cancers using WSIs. This foundation supports:

• Identification of both universally prognostic and cancer-specific signals.
• Benchmarking of transfer and adaptation methods, critical for rare tumor prognosis where data is inherently scarce.
• Empirical discovery of factors that govern effective model transfer and inform the rational design of future transferable pathology frameworks.

A plausible implication is that datasets with wide tissue diversity, clear event annotation, and sufficient case coverage—as embodied in UNI2-h-DSS—are essential resources for research advancing computational pathology beyond isolated, cancer-specific paradigms. This work lays groundwork for further exploration of adaptive, multi-cancer prognostic modeling strategies.