Abstract
Machine learning (ML) for protein design requires large protein fitness datasets generated by high-throughput experiments for training and benchmarking models. However, most models do not account for experimental noise inherent in these datasets, thereby harming model performance. Here, we develop fitness landscape inference generated by high-throughput experimental data (FLIGHTED), a Bayesian method of accounting for uncertainty by generating probabilistic fitness landscapes from noisy experiments. We demonstrate how FLIGHTED can improve model performance on two experiments: single-step selection assays, such as phage display, and a high-throughput assay that ties activity to base editing. We compare the performance of standard ML models on fitness landscapes with and without FLIGHTED. Accounting for noise statistically significantly improves model performance. Based on our new benchmarking with FLIGHTED, data size, not model scale, is limiting protein fitness model performance. Our results indicate that FLIGHTED can be applied to any high-throughput assay and any ML model, making it straightforward for protein designers to account for noise.