ALyzer3D.AI: a more generalizable deep learning predictor of light chain amyloidogenicity powered by structural and...

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in medicine

Background:

• Predicting amyloidogenic risk of immunoglobulin light chains is challenging in amyloid light-chain (AL) amyloidosis.
• Existing computational models often fail to generalize to new patient data, limiting their clinical utility.

Purpose of the Study:

• To develop an advanced computational tool for predicting the amyloidogenic risk of immunoglobulin light chains.
• To improve the generalizability and interpretability of models for AL amyloidosis.

Main Methods:

• Development of ALyzer3D.AI, a multi-modal deep learning architecture.
• Integration of evolutionary features (ESM-2 Protein Language Model), structural metrics (ColabFold), and biophysical features.
• Evaluation of generalizability on diverse datasets and model interpretability using SHAP analysis.

Main Results:

• ALyzer3D.AI demonstrated superior performance, maintaining accuracy on new data where other models faltered.
• The final model achieved an AUC of 0.86 and accuracy of 0.84 on independent test data.
• Synergistic integration of protein language model, structural, and scalar features drove model performance.

Conclusions:

• ALyzer3D.AI offers a robust and interpretable solution for predicting light-chain amyloidogenicity.
• The tool exhibits better generalization to independent datasets compared to existing methods.
• ALyzer3D.AI is publicly available for broader research and clinical application.