Deep Batch Active Learning for Protein Structure Modeling.

Area of Science:

• Structural biology and bioinformatics
• Computational drug discovery and development
• Machine learning applications in molecular modeling

Background:

• Accurate molecular structure prediction is crucial for pharmaceutical research and understanding protein function.
• Current deep learning models, while advanced, exhibit limitations in predicting structures for underrepresented molecules such as VHH antibodies.
• Experimental structure determination is time-consuming and costly, making large-scale data collection for model training impractical.

Purpose of the Study:

• To develop a strategic data selection method, DEWDROP, to improve the performance of molecular structure prediction models through iterative fine-tuning.
• To address the challenge of underrepresented molecular domains in existing training datasets by optimizing the curation of new experimental data.
• To enable superior model performance with reduced iterations and costs by maximizing the information content of selected structures.

Main Methods:

• Proposed DEWDROP, an active learning selection method utilizing Monte Carlo dropout to generate prediction ensembles for optimal data selection.
• Employed a structured prediction model, Equifold, based on coarse-grain molecular representations, independent of multiple sequence alignments.
• Conducted retrospective iterative fine-tuning experiments and batch selection analysis on VHH antibodies (SAbDab-nano) and *Mycobacterium leprae* proteins (AlphaFold Protein Database).

Main Results:

• DEWDROP significantly improved model training efficiency through optimized batch selection, outperforming baseline methods in iterative fine-tuning.
• The method successfully identified and selected structurally informative data with high information content, crucial for enhancing prediction accuracy.
• Demonstrated DEWDROP's effectiveness and broader applicability across different molecular domains beyond VHH antibodies.

Conclusions:

• DEWDROP offers a model-agnostic approach for strategic data selection in structural biology, particularly beneficial for underrepresented molecular families.
• The active learning strategy enhances the efficiency and cost-effectiveness of improving deep learning models for molecular structure prediction.
• This method facilitates superior model performance by maximizing the value of newly acquired experimental structural data.