Optimal inference of molecular interaction from live-cell FRET imaging

Signal-to-noise ratio (SNR) ultimately limits what we can learn from data. Intensity-based fluorescence resonance energy transfer (FRET) is a widely-used technique to convert the molecular interaction or conformational change in living cells into fluorescent signals that are detected under a microscope, allowing experimenters to infer molecular states in a non-invasive manner. However, existing data-analysis methods turn to algebraic manipulation of the noise-corrupted observables, resulting in the loss of information stored in the raw data at the cost of interpretability of results. Here, we present a novel computational FRET method by exploiting the framework of Bayesian filtering. Based on the direct computation of the Bayesian posterior, the method provides statistically optimal inference of molecular interaction, and thus achieves significantly higher SNR than existing methods. We quantify how the new approach outperforms existing methods using artificial data with various properties, and demonstrate the efficacy using real data obtained from both unimolecular and bimolecular FRET reporter molecules in living cells.