Table 1 Key AI/ML methodologies and their specific applications in ADC target identification and validation
AI/ML methodology | Data sources utilized | ADC-specific challenge addressed | Specific application in target validation | Example reference/Tool |
|---|---|---|---|---|
Deep learning (CNNs, GNNs, Autoencoders) | Genomics, transcriptomics, proteomics, digital pathology images, molecular structures | Overcoming target heterogeneity and ensuring functional relevance (e.g., internalization) | Assessing antigen density and homogeneity from imaging; Predicting internalization efficiency; Correlating spatial expression with outcomes | |
Natural language processing (NLP) | Scientific literature, patents, clinical trial databases, EHRs | Aggregating fragmented evidence for less-studied or novel targets | Aggregating evidence for target function and expression patterns; Identifying reported associations with resistance or sensitivity | |
Bayesian networks | Multi-omics data, clinical data, pathway databases | Identifying driver targets with causal relationships, beyond simple correlation | Modeling impact of target modulation on cellular pathways; Assessing likelihood of on-target, off-tumor toxicity | Causal network inference platforms (e.g., packages in R/Python) |
Support vector machines (SVMs) | Gene expression data, protein feature data | Classifying tumor vs. normal tissues with high precision for safety assessment | Predicting whether a protein is membrane-bound and accessible; Classifying targets based on predicted immunogenicity | Scikit-learn, LIBSVM, various custom models |
Random forests/Gradient boosting (e.g., XGBoost) | Multi-omics data, chemical databases, preclinical data | Prioritizing targets based on a weighted combination of multiple complex features | Predicting target-drug interactions; Correlating target expression with preclinical drug response; Ranking targets based on weighted criteria | XGBoost, LightGBM |
Multimodal data integration platforms | Genomics, proteomics, imaging, clinical outcomes, RWD | Building a holistic view of the target by integrating disparate data types | Validating targets by converging evidence from disparate sources; Stratifying patient populations based on multi-modal target signatures | MOFA+, Owkin (MOSAIC), TCGA pan-cancer atlas |
