Pattern Identification in Sequence Activity Data

National Institute Of Diabetes And Digestive And Kidney Diseases

Investigators

Linked publications, trials & patents

Abstract

A key application of this framework is forecasting drug resistance in HIV. We developed a probabilistic large-deviation model to analyze genotypes observed under different Protease Inhibitor (PI) therapies. By simulating stochastic evolutionary paths, we showed that low-probability mutations are required for the virus to evolve to diverse, fit genotypes. We then trained a classification model using a clinical dataset of in vitro susceptibility tests to infer the drug resistance of a given genotype. By analyzing resistance along the simulated evolutionary paths, we predicted that the combination PI-therapy of Atazanavir (ATV) and Ritonavir (RTV) is the least susceptible to developing resistance. Critically, without any prior knowledge of PI-associated mutations, our model predicted known primary and secondary resistance mutations as being essential for drug resistance. This result validates that our model can learn mechanistic relationships from sparse sequence data, a significant challenge in protein evolution. This work is under review. We also applied this technique as an extension of a similar analysis on sequence evolution of the active lineage of LINE-1 retrotransposons. We find that some mutation trajectories show large deviations and escape to chronologically adjacent LINE-1 families, predicting evolutionary pathways. Such pathways require mutations that are probabilistically rare. We used our model to compute the probabilities of these trajectories and determine the ones that are least unlikely, which we define as large deviation evolutionary pathways. Additionally, we are studying the coevolution of the coiled-coil domain and the C-terminus because of their strong functional codependence as follows. The coiled-coil domain in the middle of the open reading frame 1 protein (ORF1p) acts as a scaffold, assembling three ORF1p molecules into a stable trimer and brings the three individual C-terminal domains into close proximity. This correctly positions them to collectively bind to the LINE-1 RNA and perform their nucleic acid chaperone duties. This is ongoing work. We introduced Dynamical Systems Machine Learning (DynML), a novel hybrid framework that integrates nonlinear dynamical systems with machine learning for forecasting and classification of complex time-evolving biological signals. In this architecture, input signals are projected into an ensemble of chaotic dynamical cores (Lorenz and Rössler systems) that evolve in parallel, encoding multiscale temporal features. Simple linear readouts are trained to predict future trajectories or classify signal patterns, reducing computational burden while preserving interpretability. DynML was first validated on synthetic chaotic systems (Lorenz and double pendulum), where it achieved near-perfect trajectory reconstruction (r = 0.98 and 0.97, respectively). The framework was then applied to the MIT-BIH Arrhythmia Database for ECG time-series prediction and arrhythmia classification. Across 20 patient datasets, DynML consistently achieved high-fidelity extrapolation of ECG trajectories, with an average correlation of 0.87 on held-out test segments. Arrhythmia classification from extrapolated signals yielded an average accuracy of 0.80 and F1-score of 0.84, outperforming conventional baselines including MLP, LSTM, GRU, and TCN models. Benchmarking showed that DynML maintains stable predictive accuracy across long prediction horizons where conventional networks degrade, underscoring the importance of embedding chaotic dynamics into predictive models. The results establish DynML as a robust and interpretable framework for physiological signal prediction, with potential applicability to other biological time series such as gene expression or neural activity. A manuscript is in preparation.