FLIGHTED: Inferring Fitness Landscapes from Noisy High-Throughput Experimental Data.

Area of Science:

• Computational biology
• Machine learning
• Protein engineering

Background:

• Machine learning (ML) for protein design relies on high-throughput experimental data.
• Existing ML models often ignore experimental noise, negatively impacting performance and benchmarking.
• Uncertainty quantification is crucial for robust protein fitness modeling.

Purpose of the Study:

• To introduce FLIGHTED, a Bayesian method for modeling protein fitness landscapes from noisy experimental data.
• To improve the performance and reliability of ML models in protein design by accounting for experimental uncertainty.
• To re-evaluate factors limiting ML model performance in protein fitness prediction.

Main Methods:

• Developed FLIGHTED, a Bayesian approach to generate probabilistic fitness landscapes from noisy experimental data.
• Applied FLIGHTED to single-step selection assays (phage display, SELEX) and the DHARMA assay.
• Benchmarked standard ML models using fitness landscapes generated with and without FLIGHTED.

Main Results:

• FLIGHTED significantly enhances ML model performance, particularly for Convolutional Neural Network (CNN) architectures.
• Accounting for experimental noise alters model rankings in benchmarking studies.
• Benchmarking indicates data size, not model scale, is the primary performance limitation; model architecture is more critical than protein language model embeddings.

Conclusions:

• FLIGHTED provides a straightforward method to incorporate experimental noise into protein fitness modeling.
• The approach is broadly applicable to various high-throughput assays and ML models.
• FLIGHTED facilitates more accurate and reliable protein design by addressing data uncertainty.