The scarcity of labeled data limits development of deep learning biomarkers in oncology. We develop a two-stage AI model predicting pathological complete response (pCR) to neoadjuvant therapy in breast cancer. The first stage learns the transcriptome from histopathology using 8,742 patients across 32 cancer types, corroborated by pathologist review and spatial agreement with measured expression. This simplifies the second stage to predicting pCR from inferred expression and clinical variables. Developed using 1,080 patients (five cohorts) and evaluated in 1,412 patients (nine cohorts), the model achieves a pooled AUROC of 0.79 (95% CI, 0.73-0.85), discriminating responders within molecular subtypes. It outperforms histopathological biomarkers, remaining stable across intratumoral sampling and with minimal biopsy tissue. Ablations show transcriptome-wide inference improves discrimination over clinical variables alone or one-stage pathology models, and robustness by avoiding genomic assays' gene selection constraints. These results indicate that biologically informed compression may generalize to data-sparse applications in precision oncology.
Quantifying labeled-data scarcity in oncology biomarker development
Deep learning biomarkers require large annotated datasets, yet labeled medical data remain expensive to acquire. In neoadjuvant breast cancer therapy, pathological complete response (pCR) serves as the primary efficacy endpoint, but biopsy cohorts across institutions are typically small. The paucity of cases hampers model generalizability and limits external validation. This paper addresses the constraint by transferring transcriptome information from histopathology, thereby reducing dependence on genome-scale assays for each new prediction task.
Two-stage AI architecture: transcriptome inference from histopathology
The proposed model operates in two sequential stages. The first stage maps a histopathology slide to a genome-wide expression profile. Training data comprise 8,742 patients spanning 32 cancer types; expression labels are established through bulk RNA sequencing, and the learned mapping is validated by pathologist review as well as spatial correspondence with measured expression patterns. The second stage takes the inferred expression profile together with routinely collected clinical variables and predicts pCR status. This decoupling allows the computationally intensive transcriptome-learning phase to be performed once, after which pCR prediction can be executed on smaller, task-specific cohorts.
Pooled AUROC of 0.79 and subtype-discriminative power
Across five development cohorts (1,080 patients) and nine evaluation cohorts (1,412 patients), the model achieves a pooled area under the receiver operating characteristic curve (AUROC) of 0.79, with a 95% confidence interval of 0.73–0.85. The discriminative ability is maintained within molecular subtypes, indicating that the model captures biology relevant to tumor-intrinsic pathways rather than generic prognostic signals. This performance exceeds that of established histopathological biomarkers such as tumor grade and mitotic count, which typically report lower AUROCs in comparable tasks.
Stability across intratumoral sampling and minimal biopsy tissue
Ablation and cross-validation experiments demonstrate that the model's performance is stable across intratumoral sampling variations. Even when only a small core biopsy specimen is available, the inferred transcriptome retains sufficient signal for pCR prediction. This robustness is critical for clinical deployment, where tissue procurement may be limited by patient tolerance or institutional protocols. The model does not require whole-slide imaging at high magnification, further reducing computational and storage burdens.
Ablation evidence for transcriptome-wide inference benefits
Separate ablations isolate the contribution of the transcriptome inference step. Removing the transcriptome stage and predicting pCR directly from histopathology features yields a lower AUROC, confirming that the expression bridge adds predictive value. Likewise, using clinical variables alone produces inferior discrimination. The full model, which combines inferred expression with clinical data, consistently outperforms both unimodal alternatives. These results underscore the importance of the biologically informed compression step.
Robustness gains from avoiding genomic assay gene-selection constraints
Traditional approaches to pCR prediction often rely on genomic assays that measure a pre-specified gene panel. Such panels introduce selection bias and may omit genes with predictive relevance outside the curated set. The transcriptome-informed model avoids this constraint by learning a mapping from histopathology that can, in principle, capture any transcript expressed in the tissue. The ablation results show that this broader transcriptional coverage contributes to the model's improved and more stable performance across diverse patient populations.
Biologically informed compression for data-sparse precision oncology
The authors argue that the two-stage compression—from histopathology slide to transcriptome profile to pCR prediction—represents a generalizable strategy for precision oncology applications where labeled data are sparse. By transferring knowledge from a large, multi-cancer training set, the model can be adapted to new cancer types or clinical questions with minimal additional labeling. The reported results, while specific to breast neoadjuvant therapy, demonstrate a principle that may extend to other tumor types, treatment modalities, and outcome endpoints.