MMCTOP: A Multimodal Textualization and Mixture-of-Experts Framework for Clinical Trial Outcome Prediction

TL;DR

MMCTOP introduces a schema-guided, multimodal framework for clinical-trial outcome prediction that unifies molecular, protocol, and disease signals through domain-specific encoders and a drug–disease–conditioned sparse Mixture-of-Experts. The scheme-coupled textualization layer provides auditable, grounded inputs, while top-$k$ routing enables scalable, context-aware specialization across therapeutic areas and trial designs. On TOP and CTOD benchmarks, MMCTOP consistently outperforms unimodal and prior multimodal baselines, with ablations showing the central roles of textualization and conditional routing in performance and stability, especially in late-stage trials. These results support the framework as a governance-friendly, generalizable approach for data-driven clinical development and broader biomedical multimodal AI tasks.

Abstract

Addressing the challenge of multimodal data fusion in high-dimensional biomedical informatics, we propose MMCTOP, a MultiModal Clinical-Trial Outcome Prediction framework that integrates heterogeneous biomedical signals spanning (i) molecular structure representations, (ii) protocol metadata and long-form eligibility narratives, and (iii) disease ontologies. MMCTOP couples schema-guided textualization and input-fidelity validation with modality-aware representation learning, in which domain-specific encoders generate aligned embeddings that are fused by a transformer backbone augmented with a drug-disease-conditioned sparse Mixture-of-Experts (SMoE). This design explicitly supports specialization across therapeutic and design subspaces while maintaining scalable computation through top-k routing. MMCTOP achieves consistent improvements in precision, F1, and AUC over unimodal and multimodal baselines on benchmark datasets, and ablations show that schema-guided textualization and selective expert routing contribute materially to performance and stability. We additionally apply temperature scaling to obtain calibrated probabilities, ensuring reliable risk estimation for downstream decision support. Overall, MMCTOP advances multimodal trial modeling by combining controlled narrative normalization, context-conditioned expert fusion, and operational safeguards aimed at auditability and reproducibility in biomedical informatics.

Figures (3)

• Figure 1: Unified resource$\rightarrow$modality map with typical fields. Central circles denote modalities, solid arrows are primary contributions and dashed arrows indicate contextual linkages (e.g., ontology crosswalks informing protocol text).
• Figure 2: MMCTOP architecture. Heterogeneous clinical-trial signals (disease, drug, enrollment, description, eligibility criteria, summaries, and molecular structure) are first standardized through schema-guided textualization and normalization, producing auditable narrative artifacts and structured slots. Clinical text modalities are encoded using a biomedical language model (ClinicalBERT), while molecular structures are encoded using a chemical language model (ChemBERTa). Modality-specific embeddings are projected into a shared latent space and cached for computational efficiency. A drug--disease--conditioned sparse Mixture-of-Experts (SMoE) performs top-$k$ routing to specialized experts, followed by a final MoE integrator that balances modality contributions and produces the trial outcome prediction.
• Figure 3: Schema-guided textualization. Structured clinical-trial records are linearized into slot-ordered prompts (prefix + key:value pairs) that serve as deterministic inputs to the LLM (GPT-3.5-turbo; $T{=}0$). The model produces two artifacts: a brief summary and a detailed text description, both persisted as JSON (processed) and Pickle (raw) files for audit and reproducibility.