Accurate prediction of B-form/A-form DNA conformation propensity from primary sequence: A machine learning and free energy handshake

DNA carries the genetic code of life, with different conformations associated with different biological functions. Predicting the conformation of DNA from its primary sequence, although desirable, is a challenging problem owing to the polymorphic nature of DNA. We have deployed a host of machine learning algorithms, including the popular state-of-the-art LightGBM (a gradient boosting model), for building prediction models. We used the nested cross-validation strategy to address the issues of "overfitting" and selection bias. This simultaneously provides an unbiased estimate of the generalization performance of a machine learning algorithm and allows us to tune the hyperparameters optimally. Furthermore, we built a secondary model based on SHAP (SHapley Additive exPlanations) that offers crucial insight into model interpretability. Our detailed model-building strategy and robust statistical validation protocols tackle the formidable challenge of working on small datasets, which is often the case in biological and medical data.