Fusing Tumor Evidence Only Where RNA Fails in Cancer of Unknown Primary

When a cancer is found only as a metastasis and doctors cannot tell which organ it started in, the diagnosis is "cancer of unknown primary" (CUP). Treatment in oncology is chosen by organ of origin, so these patients usually get generic chemotherapy and do poorly. Classifiers that read a tumor's gene expression can often recover the organ of origin. On primary tumors they are already very accurate, and adding other data types has not helped. A systematic study of 7,224 primary tumors found that gene expression alone reached 94.63% accuracy, and that combining it with DNA methylation and mutations did not beat it (Evaluating DNA Methylation, Gene Expression, Somatic Mutation, and Their Combinations in Inferring Tumor Tissue-of-Origin). We think that result is correct and also misleading. Averaged over easy primary tumors, a second modality has almost nothing to add. Its value should sit in a small, clinically important tail: metastases in which the expression signal is weak or distorted, either because the biopsy is mostly normal host tissue (liver, lymph node, bone) or because the tumor has lost its lineage program.

We propose a fusion model that trusts the RNA-based prediction by default. It hands the decision to other evidence only when the RNA prediction is unreliable. That other evidence is copy-number and mutation profiles, which normal host cells do not share because they carry the normal complement of chromosomes, and slide morphology. Whether RNA is unreliable is judged from three things: the RNA model's own uncertainty, the estimated fraction of tumor cells in the sample, and the organ that was biopsied. We also propose the evaluation needed to see whether fusion works at all. That means reporting accuracy by RNA-confidence band and by tumor purity × biopsy site, instead of a single averaged accuracy that cannot move.

Core Claim

Our claim is that multimodal gains for tissue-of-origin prediction are real but concentrated. On metastatic samples with low tumor purity and a biopsy organ whose normal expression resembles a candidate primary, a fusion model that arbitrates between modalities will recover a measurable share of RNA's errors. On high-confidence cases it will leave RNA's answer unchanged. Plain concatenation and generic cross-attention fusion will not show this pattern. What would be new is that the fusion mechanism explicitly models why RNA fails. It conditions on purity and host site, and it gives the RNA branch a reference for normal expression in the host organ. The approach does not assume that more modalities are always better.

Research Program

The motivating evidence comes from two papers. The multi-omics profiling of 70 CUP patients in the MASTER program found that transcriptome and methylome predictions "frequently complement each other when one modality yields ambiguous classifications," together giving evidence for a primary in 89% of cases (Comprehensive genomic and epigenomic analysis in cancer of unknown primary guides molecularly-informed therapies despite heterogeneity). That work combined modalities by expert review of separate pipelines, not with a learned model. The Liu et al. comparison supplies the constraint: fusion by concatenation followed by XGBoost tracks RNA without exceeding it. Any design that does not specifically target the ambiguous cases should be expected to fail in the same way.

The model has three parts. The first is an RNA branch: a regularized multinomial classifier in the style of Tempus TO (Validation of a Transcriptome-Based Assay for Classifying Cancers of Unknown Primary Origin). Its input is augmented with a host-tissue term. Given the biopsy organ and an estimated purity $p$, it models the observed expression as a mixture, $x \approx p\, t + (1-p)\, h_{\text{site}}$. Here $t$ is the tumor's own profile and $h_{\text{site}}$ is a reference profile of normal tissue from that organ, estimated from TCGA adjacent-normal samples. The classifier sees both the raw profile and the host-corrected residual.

The second part is a set of secondary branches for evidence that host tissue contaminates less. These are a copy-number and mutation encoder (CNV-based XGBoost origin prediction exists as a standalone method in A Novel XGBoost Method to Identify Cancer Tissue-of-Origin Based on Copy Number Variations), a methylation encoder, and a slide encoder. The slide encoder uses attention-based multiple-instance learning, which scores many small image patches and learns which ones matter for the slide-level label, as in TOAD (AI-based pathology predicts origins for cancers of unknown primary).

The third part is an arbiter. It is a small network whose inputs are the RNA prediction's entropy, the estimated purity, the biopsy site, and which modalities are present. It outputs a mixing weight $g \in [0,1]$, and the final prediction is $g \cdot P_{\text{RNA}} + (1-g) \cdot P_{\text{other}}$. Missing modalities are simply left out of $P_{\text{other}}$. We borrow this from NAIM/MARIA, which treat absence as information instead of imputing it (Next-gen health: from multimodal AI to foundation models), and from the modality dropout used in Multimodal CustOmics (Multimodal CustOmics: A unified and interpretable multi-task deep learning framework for multimodal integrative data analysis in oncology).

The design differs from the heavier fusion systems in the file. CustOmics, CA-MLIF (CA-MLIF: Cross-Attention and Multimodal Low-Rank Interaction Fusion Framework for Tumor Prognostic Prediction) and FOAA (FOAA: Flattened Outer Arithmetic Attention For Multimodal Tumor Classification) learn rich interactions between modalities and were validated on primary tumors or prognosis tasks. Our design deliberately learns very little interaction. The one thing it learns is when to stop trusting the dominant modality.

Labels are the central difficulty. True CUP has no ground truth, so the main training and test data are samples whose primary is known. We use TCGA primaries, the benchmark used by both Liu et al. and Möhrmann et al., together with TCGA's metastatic samples. We add in-silico metastases, made by mixing tumor expression and DNA profiles with adjacent-normal profiles from a chosen host organ at purities from 5% to 60%. Mixing dilutes the allele fractions of mutations and flattens copy-number changes in a controlled way, which is exactly the regime we want to study. Slides cannot be simulated this way, so slide features come only from real samples. We will state openly that the simulation tests the RNA–DNA part of the claim more cleanly than the morphology part.

Baselines

The naive reference is the RNA-only classifier with no host correction, trained on the same data. On TCGA primaries, the file reports 94.63% accuracy for expression-only XGBoost, 87.59% for methylation and 43.33% for mutations (Evaluating DNA Methylation, Gene Expression, Somatic Mutation, and Their Combinations in Inferring Tumor Tissue-of-Origin). We will reproduce these numbers on the same 21-type split before measuring anything else.

The strongest existing RNA system is Tempus TO. It reports 91.1% overall accuracy and 97.5% top-3 accuracy across 68 subtypes on 9,210 internal test samples, and more than 89% accuracy at tumor purities as low as 10% (Validation of a Transcriptome-Based Assay for Classifying Cancers of Unknown Primary Origin). Its training set and panel are proprietary, so we reimplement its recipe: an L2-regularized multinomial logistic regression on a matched gene set, tuned with the same budget as our model. We will cite its reported numbers only as a reference for what the recipe achieves at scale.

The strongest pathology baseline is TOAD. It reports 83.6% top-1 accuracy on its internal test set and 78.5% on an external test set drawn from 202 centers. We retrain it on TCGA slides.

The strongest fusion baselines are concatenation-plus-XGBoost as in Liu et al. and a reimplementation of Multimodal CustOmics with modality dropout. Both receive the same folds and the same hyperparameter budget, measured in GPU-hours and search trials. As far as the file shows, CustOmics has not been evaluated on tissue of origin in metastases, so there is no reported result found on Lacuna for fusion models on metastatic tissue-of-origin.

Two ablations isolate the key ingredients. The first removes the host-tissue correction from the RNA branch. The second replaces the arbiter's purity and site inputs with RNA entropy alone. Beating the baselines means two things. The model must be non-inferior to RNA-only on the high-confidence half of the test set, and it must give a statistically significant gain in top-1 accuracy, with paired bootstrap confidence intervals, in the low-purity, host-confounded stratum.

Why not just…?

Why not just average calibrated unimodal probabilities? Late averaging of well-calibrated classifiers is a real threat, because it already lets a confident DNA model outvote an uncertain RNA model. The weak point is fixed weights. A fixed average also lets the weak mutation branch drag down easy cases, which is plausibly why concatenation stalls. Our experiment includes averaging with weights tuned on validation data, both globally and per biopsy site. If per-site weighted averaging matches the arbiter in every stratum, we will report that the contribution is the stratified evaluation and the host correction, not the learned gate.

Why not just train a better RNA model with biopsy site and purity as extra input features? This is the strongest alternative, and Tempus TO's robustness at 10% purity suggests RNA may already cope with dilution on its own. The ablation that feeds site and purity straight into the RNA classifier, with no other modality, settles it. If that model closes most of the gap in the confounded stratum, the proposal reduces to host-aware RNA classification, and we will say so. Our expectation, which follows from the mixture model, is that a normal liver background pulls predictions toward a hepatic origin in a way that DNA evidence resists and RNA reweighting cannot fully undo.

Why not just use a large cross-attention fusion model? Cross-attention can in principle learn the gating itself. But the evidence in the file shows fusion gains disappearing into an easy majority of cases, and the CUP-relevant tail is small, so a high-capacity model is likely to overfit it. The CustOmics baseline, trained on the same data, tests this directly.

Proposed Experiments

The first experiment reproduces the Liu et al. single-modality and concatenation results on TCGA primaries, to check the setup.

The second is the main test. It evaluates every model on real TCGA metastases and on in-silico metastases, crossing five purity levels with host organs (liver, lymph node, lung, bone). It reports top-1 accuracy, top-3 accuracy and calibration error in each cell, plus accuracy by decile of RNA confidence. The decisive check is the shape of the gain. The claim holds only if gains over RNA-only are near zero in the top confidence deciles, positive in the bottom deciles, and largest where host tissue resembles a candidate primary. A uniform small gain, or no gain, falsifies it.

The third tests leakage from the simulation. We train on one set of host organs and test on held-out host organs, so the arbiter cannot memorize site-specific shortcuts.

The fourth applies the approach to CUP. If a cohort with CUP samples becomes available, we follow the two surrogate checks used in the literature because true labels are absent. The first is concordance with orthogonal DNA-based subtype associations, as in Tempus TO. The second is agreement with the clinical differential diagnosis, as in TOAD. We also report the fraction of cases in which the arbiter overrode RNA and whether DNA evidence supported the override.