Multivariable Predictive Models to Identify the Optimal Biologic Therapy for Treatment of Patients With Psoriasis at the Individual Level.

Importance: Identifying the optimal long-term biologic therapy for patients with psoriasis is often done through trial and error. Objective: To identify the optimal biologic therapy for individual patients with psoriasis using predictive statistical and machine learning models. Design, Setting, and Participants: This population-based cohort study used data from Danish nationwide registries, primarily DERMBIO, and included adult patients treated for moderate-to-severe psoriasis with biologics. Data were processed and analyzed between spring 2021 and spring 2022. Main Outcomes and Measures: Patient clusters of clinical relevance were identified and their success rates estimated for each drug. Furthermore, predictive prognostic models to identify optimal biologic treatment at the individual level based on data from nationwide registries were evaluated. Results: Assuming a success criterion of 3 years of sustained treatment, this study included 2034 patients with a total of 3452 treatment series. Most treatment series involved male patients (2147 [62.2%]) originating from Denmark (3190 [92.4%]), and 2414 (69.9%) had finished an education longer than primary school. The average ages were 24.9 years at psoriasis diagnosis and 45.5 years at initiation of biologic therapy. Gradient-boosted decision trees and logistic regression were able to predict a specific cytokine target (eg, interleukin-17 inhibition) associated with a successful treatment with accuracies of 63.6% and 59.2%, and top 2 accuracies of 95.9% and 93.9%. When predicting specific drugs resulting in success, gradient boost and logistic regression had accuracies of 48.5% and 44.4%, top 2 accuracies of 77.6% and 75.9%, and top 3 accuracies of 89.9% and 89.0%. Conclusions and Relevance: Of the treatment prediction models used in this cohort study of patients with psoriasis, gradient-boosted decision trees performed significantly better than logistic regression when predicting specific biologic therapy (by drug as well as target) leading to a treatment duration of at least 3 years without discontinuation. Predicting the optimal biologic could benefit patients and clinicians by minimizing the number of failed treatment attempts.

DeepFRET, a software for rapid and automated single-molecule FRET data classification using deep learning.

Single-molecule Förster Resonance energy transfer (smFRET) is an adaptable method for studying the structure and dynamics of biomolecules. The development of high throughput methodologies and the growth of commercial instrumentation have outpaced the development of rapid, standardized, and automated methodologies to objectively analyze the wealth of produced data. Here we present DeepFRET, an automated, open-source standalone solution based on deep learning, where the only crucial human intervention in transiting from raw microscope images to histograms of biomolecule behavior, is a user-adjustable quality threshold. Integrating standard features of smFRET analysis, DeepFRET consequently outputs the common kinetic information metrics. Its classification accuracy on ground truth data reached >95% outperforming human operators and commonly used threshold, only requiring ~1% of the time. Its precise and rapid operation on real data demonstrates DeepFRET's capacity to objectively quantify biomolecular dynamics and the potential to contribute to benchmarking smFRET for dynamic structural biology.

Proteins are folded into particular shapes in order to carry out their roles in the cell. However, their structures are not rigid: proteins bend and rotate in response to their environment. Identifying these movements is an important part of understanding how proteins work and interact with each other. Unfortunately, when researchers study the structures of proteins, they often look at the 'average' shape a protein takes, missing out on other conformations the protein might only be in temporarily. An important technique for studying protein flexibility is known as single molecule Förster resonance energy transfer (FRET). In this technique, two light-sensitive tags are attached to the same protein molecule and give off a signal when they come into close contact. This nano-scale sensor allows structural biologists to get information from individual protein movements that can be lost when looking at the average conformations of proteins. Advances in the instruments used to perform FRET have made observing the motion of individual proteins more widely accessible to non-specialists, but the analysis of the data that these instruments produce still requires a high level of expertise. To lower the barrier for non-specialists to use the technology, and to ensure that experiments can be reproduced on different instruments and by different researchers, Thomsen et al. have developed a new way to automate the data analysis. They used machine learning technology to recognize, filter and characterize data so as to produce reliable results, with the user only needing to perform a couple of steps. This new analysis approach could help expand the use of single-molecule FRET to different fields , allowing researchers to investigate the importance of protein flexibility for certain diseases, or to better understand the roles that proteins have in a cell.