MIST: Multimodal Survival Prediction with Genomic-Guided Histology Attention
Muhammet Sami Yavuz, Sabri Mustafa Kahya, Richard R. Chen, Jana Lipkova, Benedikt Wiestler
cs.AI · cs.CV
Quick verdict
Claims genomic-guided attention beats fusion baselines on 4 external survival cohorts; credible, but gains lack significance tests and are modest vs WSI-only.
Summary
MIST is a multimodal survival-prediction model in which genomic feature tokens act as queries in cross-attention over compact histology context tokens projected from TITAN pathology foundation-model slide embeddings, before a shared transformer outputs discrete-time (K=4) hazards. Training combines this fusion with variable-rate genomic masking, WSI dropout (p=0.3), and a symmetric InfoNCE alignment between the survival representation and the matched slide embedding, extending the authors' SHIFT genomics-only backbone. Models are trained on TCGA cohorts (colon, renal, lung squamous, glioblastoma) and evaluated externally on CPTAC COAD (n=101), CPTAC KIRC (n=94), US LUSC (n=102), and German GBM (n=132). Across all four primary comparisons, MIST attains the highest fold-wise external C-index and time-dependent AUC against Concat, Bilinear, and Co-Attention baselines and their mask-aware variants (e.g., 0.669 vs 0.612 on COAD-CPTAC). Key caveats: the headline fold-wise comparisons carry no significance tests; on a fully-paired GBM control MIST (0.491) underperforms Co-Attention (0.555); the genomics-only baseline sits at chance everywhere, so gains over the WSI-only model are modest outside COAD and GBM; and external cohorts are small with heavy censoring (COAD-CPTAC has roughly seven events). Code is publicly released.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Letting observed genomic feature tokens cross-attend to compact TITAN-derived histology context tokens yields higher external C-index and tAUC than late fusion, bilinear fusion, and symmetric co-attention in all four primary external evaluations (Table 2).
- Training-time components are load-bearing: removing InfoNCE alignment costs 10.5% (KIRC-CPTAC) and 13.5% (LUSC-US) relative C-index, and removing histology attention costs 15.3% and 7.2% (Table 4).
- MIST's advantage is tied to real-world modality incompleteness: on a fully paired GBM control the gain over co-attention disappears (Table 2, last column).
- MIST risk scores stratify external patients prognostically, with statistically significant median-split separation only in KIRC-CPTAC (log-rank p=0.0098, Table 3).
Strengths
Controlled fusion-operator comparison
All fusion baselines (Concat, Bilinear, Co-Attn, plus mask-aware variants) share MIST's genomic backbone, TITAN embeddings, discrete NLL loss, VRM and WSI dropout, differing only in the fusion operator (Sec 3.2), which isolates the contribution of the fusion mechanism in Table 2.
Genuinely external, heterogeneous evaluation
Training on TCGA with testing on CPTAC (COAD n=101, KIRC n=94) plus two institutional cohorts (US LUSC n=102, German GBM n=132) across four assay designs (mutation panels, CNV log2 ratios, a 36-feature GBM panel) rather than a single held-out TCGA split (Sec 3.1, Table 1).
Transparent reporting of unflattering results
Component ablations on the informative cohorts
Missingness-aware design matched to real deployment
Weaknesses
No significance testing for the headline comparisons
The matched-pair control reverses the ranking
Gains over the strongest unimodal baseline are thin
Selective definition of 'best baseline' on LUSC-US
Fragile statistical substrate
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
MIST (arXiv 2609.21811) is a method paper on multimodal survival prediction from whole-slide images (WSIs) and genomic profiles, formatted for a medical-imaging venue. The central claim is architectural: instead of late fusion or symmetric co-attention, genomic feature tokens should act as queries over four compact histology context tokens projected from TITAN slide embeddings (Sec 2, Eq. 1). Secondary claims: training-time components — variable-rate genomic masking (VRM, f=0.5), WSI dropout (p=0.3), and symmetric InfoNCE alignment (tau=0.07) between the survival CLS token and the paired slide embedding — are load-bearing (Secs 2.1-2.2, Table 4). Evaluation spans four TCGA-trained cohorts tested externally on CPTAC (COAD n=101, KIRC n=94), a US LUSC cohort (n=102), and a German GBM cohort (n=132). The headline claim — highest external C-index and tAUC in all four primary comparisons (Table 2) — is accurate as stated for the tested fusion baselines, but two framing choices deserve scrutiny.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is MIST and what problem does it solve?
MIST is a deep learning model that predicts cancer patient survival by fusing whole-slide histopathology images (via TITAN foundation-model embeddings) with genomic profiles such as mutations and copy-number variations. Its distinctive step is letting each genomic feature query a small set of histology context tokens through cross-attention instead of merging the two modalities only at the end. It is built to keep working when patient data are incomplete or measured differently across hospitals.
How much better is MIST than previous fusion methods?
Across four external test cohorts (colon, kidney, lung, glioblastoma; 94-132 patients each), MIST reported the highest C-index among the tested fusion baselines, for example 0.669 vs 0.612 for the strongest comparator in colon cancer (CPTAC) and 0.646 vs 0.602 for the best fusion baseline in kidney cancer. However, gains over an images-only model were smaller (about 0.02-0.03 C-index in kidney and lung), and no statistical significance tests were reported for these fold-level comparisons.
Is the MIST code and data available?
The code is at github.com/samiyavuuz/MIST. Training uses public TCGA and CPTAC data with precomputed TITAN slide features, so the COAD and KIRC experiments are in principle replicable. Two of the four external test cohorts (a US lung cohort and a German glioblastoma cohort) are institutional, and the paper does not state whether they can be shared.
Does MIST work when a patient has only images or only genomics?
Yes — that is its core design goal. Missing genomic features stay as NaNs and are excluded from attention, and when no slide exists the histology attention step is skipped and the model proceeds on genomics alone. In the reported experiments all external patients had both modalities, and notably in a fully paired control cohort MIST did not beat a simpler co-attention baseline.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Table 1 dataset statistics are misaligned in the extracted text (e.g., the GBM training row appears shifted, and KIRC shows the feature count 920 in the patients column), so training-cohort counts quoted here rely on partial reconstruction; external cohort sizes (101/132/94/102) are unambiguous in prose and Table 3.
- Figures 1-2 are available only as embedded text and axis labels; the architecture diagram and Kaplan-Meier curves could not be visually inspected.
- The precise definition of tAUC and the censored term of the discrete-time NLL are referenced to prior work ([15], [16]) but not fully spelled out in the extracted text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.