Antibody-Drug Conjugate (ADC) Clinical Trial Pre-screening

Data Characteristics for ADC Clinical Trial Pre-screening

Data for Antibody-Drug Conjugate (ADC) clinical trial pre-screening comes from diverse sources. These include public clinical trial registries (e.g., ClinicalTrials.gov), pharmaceutical company internal R&D databases, academic papers, patent literature, and related biomedical knowledge graphs. This data often combines unstructured text (e.g., study protocols, reports, patient medical descriptions) and structured data (e.g., genomics, proteomics, imaging, pathology data, drug target information, toxicity data, efficacy indicators). Update frequency varies: public clinical trial data typically updates in real-time as trials progress, while internal R&D data iterates based on project cycles and data generation speed. Document structures are complex, containing numerous medical terms, abbreviations, and specific formats. Fields include target expression levels, linker types, payload drug molecular structures, dosages, dosing regimens, adverse event grading (e.g., CTCAE grades), and efficacy evaluation criteria (e.g., RECIST criteria). Units cover molar concentrations, dosages (mg/kg), time (days, weeks, months), and biomarker concentrations (ng/mL), often with specific reference ranges.

Constraints from ADC Data Characteristics on Model Integration and Configuration

The unique complexity and diversity of ADC data impose specific requirements on model integration and configuration. The mixed structured and unstructured nature of the data necessitates a knowledge base construction approach that balances text segmentation with structured field extraction. The abundance of medical terminology and abbreviations requires using domain-specific embedding models to improve semantic understanding accuracy. The real-time nature of data updates means the knowledge base must support incremental updates and effectively handle version conflicts. Differences in data formats from various sources increase data preprocessing complexity, requiring flexible parsers and normalization strategies. For example, the presence of specific evaluation systems like CTCAE grades and RECIST criteria requires the model to identify and correctly interpret their significance in clinical trial results. Furthermore, chemical information like drug molecular structures may require specialized cheminformatics tools for preprocessing to generate feature vectors usable by the model. All these factors directly influence data ingestion, segmentation strategies, embedding model selection, and retrieval model configuration.

Recommended Configuration

Configuration Item	Suggested Value	Rationale
`Chunk size` (Chunk Size)	`500–800 characters` (characters)	Clinical trial documents often have long paragraphs containing multiple medical concepts. This length balances contextual completeness with retrieval efficiency.
`Overlap Size`	`50–100 characters` (characters)	Ensures semantic continuity between adjacent chunks, preventing critical information from being split.
`Embedding Model`	`bge-large-zh-v1.5`	Possesses strong Chinese semantic understanding capabilities and good generalization for medical terminology.
`Recall count` (Recall Count)	`8–12 entries` (items)	Provides sufficient coverage while reducing the computational burden on subsequent reranking models, balancing relevance and efficiency.
`Similarity threshold` (Similarity Threshold)	`0.75–0.85`	A higher threshold is set to filter out less relevant document segments, addressing the precision requirements of ADC clinical pre-screening.
`Rerank result count` (Reranked Return Count)	`3–5 entries` (items)	Focuses on the most relevant core information, reducing the engineer's reading burden and improving decision-making efficiency.

Common Pitfalls

The text extraction component in the workflow fails to correctly parse unstructured text from certain clinical trial reports, leading to missing key information. This occurs because reports may contain complex tables, figures, or non-standardized medical expressions that the current model lacks the ability to recognize and extract.
After deploying the reranking model, it performs normally during testing, but the reranker_score field is consistently empty or returns false during actual retrieval. This might be due to incompatible GPU driver versions in the computing environment with the model inference framework, or network communication issues between the reranker service and FastGPT, causing model inference to fail.
Mcp plugin configuration and testing show no abnormalities, but only XML or JSON code blocks are output during the response generation process, without rendering charts. This could be because the chart data format in the generated content does not meet the Mcp plugin's expected input standards, or there are version conflicts or missing dependencies in the rendering library used by the Mcp plugin.

Verification Steps

Select multiple ADC clinical trial documents from different sources (e.g., ClinicalTrials.gov, internal reports), upload them, and perform knowledge base chunking. Check if the document chunks in the knowledge base are complete and semantically coherent, without obvious truncation or incorrect merging.
For a specific ADC drug or target, input several complex queries related to clinical trial pre-screening (e.g., queries involving specific gene mutations, adverse events, or dosing regimens). Check if the score and reranker_score of the recalled results are reasonable and if the returned document segments indeed contain the core medical concepts from the query.
Use FastGPT's testing feature to input simulated clinical pre-screening questions. Observe if the model's generated answers accurately cite ADC-related information from the knowledge base and evaluate its understanding and application of specific terms like CTCAE grades and RECIST criteria.

Note: The values provided are common starting points and should be measured against your own samples.

Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-21.