State-Aware Interaction MIL for Rare Joint Molecular Phenotype Prediction in Colorectal Cancer and Lung Adenocarcinoma

1 Introduction

• Joint BRAF/MSI and EGFR/TP53 status has clinical relevance through molecular classification, prognostic patterns, and treatment response.BRAF/MSI assessment contributes to molecular classification and hereditary-risk evaluation; EGFR/TP53 co-occurrence has been associated with shorter survival during EGFR-targeted therapy.Source: §1 · pp. 1–2
• Computational pathology studies often predict biomarkers separately, and rare joint-positive phenotypes have received limited attention.MIL offers weakly supervised slide-level prediction when molecular labels are available at patient or slide level rather than for individual tissue regions.Source: §1 · p. 2
• State-Aware Interaction MIL preserves biomarker-specific histological representations in separate MIL branches and models interactions between their slide representations.The method is proposed as a weakly supervised approach for rare joint molecular phenotype prediction from whole-slide histology.Source: §1 · p. 2
• The method supervises all four molecular configurations instead of grouping the double-negative and two single-positive states into one negative class.The four-state formulation distinguishes the double-negative, two single-positive, and joint-positive configurations.Source: §1 · p. 2

2.1 State-Aware Interaction MIL

• Separate gated-attention MIL branches produce biomarker-specific slide representations from tile embeddings for the two molecular targets.Each branch has its own tile projection and attention distribution.Source: §2.1 · p. 3
• The interaction vector retains both biomarker-specific representations and augments them with their elementwise product and absolute difference.An interaction MLP maps this vector to a scalar interaction logit.Source: §2.1 · p. 4
• The model predicts four states—double-negative, two single-positive configurations, and joint-positive—using biomarker-specific logits and a learned interaction term.The double-negative state is the reference; the joint-positive logit combines both biomarker-specific logits with the interaction term.Source: §2.1 · pp. 3–4
• When the interaction term is zero, the joint-positive logit is additive; a nonzero term captures departures from that additive combination.The four-state logits are parameterized as [0, b, a, a + b + i].Source: §2.1 · p. 4
• Class-weighted four-state cross-entropy is combined with weighted binary cross-entropy losses that directly supervise each biomarker-specific branch.The overall objective uses the auxiliary-loss weight λaux.Source: §2.1 · pp. 4–5

3.1 Dataset and Labels

• The evaluation includes 443 TCGA-COAD+READ patients with BRAF and MSI labels and 428 TCGA-LUAD patients with EGFR and TP53 labels.The joint-positive endpoints are BRAF+/MSI+ for colorectal cancer and EGFR+/TP53+ for lung adenocarcinoma; patients require valid labels for both biomarkers and available representations.Source: §3.1 · p. 5

3.2 Foundation Models and Baselines

• All models use frozen UNI2-h and CONCH tile embeddings; the baselines include direct joint prediction, independent biomarker models, multitask learning, post-hoc logistic regression, and four-state prediction.UNI2-h and CONCH produce 1,536- and 512-dimensional embeddings, respectively.Source: §3.2 · pp. 5–6

3.3 Metrics and Training Parameters

• Joint-positive performance is measured by AUROC and average precision, reported as mean ± standard deviation across five patient-level outer folds stratified by four-state label.Test folds contain 5–7 joint-positive cases in colorectal cancer and 7–8 in lung adenocarcinoma.Source: §3.3 · p. 6

4.1 Joint-Positive Prediction Performance

• 0.5566 vs 0.5161 for UNI2-h AP in TCGA-COAD+READ and 0.2784 vs 0.2525 in TCGA-LUAD; State-Aware improved AP across both representations.The UNI2-h comparisons are against NaiveMTL in COAD+READ and IndependentPair in LUAD.Source: Table 2 · p. 6, §4.1 · pp. 6–7

4.2 Comparison with Four-State ABMIL

• State-Aware exceeded FourState-ABMIL in AP in all four cohort–representation comparisons, while AUROC was also higher in each comparison.AP was 0.5566 versus 0.4965 with UNI2-h and 0.4410 versus 0.3859 with CONCH in COAD+READ; in LUAD, it was 0.2784 versus 0.2418 and 0.1659 versus 0.1529, respectively.Source: §4.2 · p. 8, Table 3 · p. 8

4.3 Four-State Molecular Separation

• Prediction scores differed across four states in both cohorts, but the joint-positive group separated from all alternatives only in COAD+READ.In LUAD, joint-positive scores were higher than the double-negative and TP53-only groups, but not significantly higher than the EGFR-only group.Source: §4.3 · pp. 8–9, Table 4 · p. 8

4.4 Component Ablation

• Binary supervision reduced AUROC and AP across cohorts and representations; removing interaction or auxiliary supervision usually lowered AP, with a CONCH-LUAD interaction exception.The no-interaction variant achieved higher AP with CONCH in LUAD, while removing auxiliary supervision changed CONCH-LUAD AP little.Source: §4.4 · p. 9, Table 5 · p. 9

4.5 Spatial Attention Analysis

• In held-out cases, State-Aware attention was more localized than DirectJoint attention; the visualizations use geometric-mean biomarker attention weights.Both models focused on overlapping tumor-containing regions in COAD+READ; in LUAD, DirectJoint attention covered a wider slide area. The localized regions may contain patterns associated with the joint phenotypes.Source: §4.5 · pp. 9–10, Figure 3 · p. 10, Figure 4 · p. 10

• AP improvements were consistent across both representations, while AUROC patterns depended on representation; joint-positive prediction was lower in LUAD across methods.Source: §5 · p. 10
• BRAF+/MSI+ scores differed from all alternative states in COAD+READ, whereas EGFR+/TP53+ scores overlapped with the EGFR-only state in LUAD.Source: §5 · p. 10
• Retaining four molecular states improved joint-positive prediction versus binary supervision, while interaction and auxiliary-supervision effects varied by cohort and representation.Source: §5 · pp. 10–11
• Prediction formulation can affect spatial attention as well as the final slide-level prediction.Source: §5 · p. 11
• Evaluation is limited to retrospective TCGA, two cancer–biomarker pairs, and two frozen representations; few joint-positive cases per test fold contribute to performance variability.Source: §5 · p. 11

• State-Aware improved AP with both representations across TCGA-COAD+READ and TCGA-LUAD, while AUROC improvement was observed with CONCH.Source: §6 · p. 11
• The method preserves biomarker-specific representations, models their interaction, and supervises the complete four-state molecular configuration.Source: §6 · p. 11
• Ablations found that retaining the complete molecular configuration improved joint-positive prediction, and prediction scores differed across molecular states.Source: §6 · p. 11
• The results support structured molecular-state supervision for predicting rare co-occurring biomarker phenotypes from pathology foundation-model representations.Source: §6 · p. 11