Periodic Reporting for period 1 - RNA-RAPSODY (RNA Bioinformatics by Robustness Analysis of Parameter Space Optimization for Dynamic Programming)
Reporting period: 2021-05-01 to 2023-04-30
The proposed bioinformatics project has strong ties to even several diverse disciplines, namely computer science, molecular biology, genetics, mathematics and statistics.
RNA Biology plays a central role in bioinformatics research such that numerous model-driven algorithms and methods are developed to predict and calculate structures and functions or compare sequences of RNA molecules that are essential in molecular and genetic processes and thus for medical and pharmaceutical applications.
Many of these problems can be cleanly phrased as optimization problems with 'optimal substructure' and thus solved exactly and efficiently by dynamic programming (for arbitrary parametrization of the objective
function). This project will systematically explore the impact of parameter changes on the quality of results of DP optimization methods, such as predictions of molecule structures or comparison of sequences. The
approach strongly relies on algebraic dynamic programming (ADP), which decouples the decomposition of the search space from the algebra used to compute a final result. Thus, the ADP framework provides a unified
setting and a generic implementation to quickly test working hypotheses. Here, it enables naturally implementing the suggested parametric optimization by developing novel algebras.
Those include the polytope algebra, which allows to segment the parameter space based on its impact of the final prediction, and a formal derivative algebra, which allows to compute the derivative of ensemble predictions with respect to a given parameter. Conversely, those methods can be
used to learn the optimal parameter sets based on a reference set of instances. This will result in deeper insights into robustness of the algorithms to changes of parameters or input data and hence, results can be assessed based on robustness measures and leads to the calculation of biologically more meaningful results.
The overall methodology will be applied to selected problems in RNA Bioinformatics.
At the end of the MSCA project, it can be concluded that project has achieved some of its objectives and milestones. The career related objectives have all been achieved as the researcher has secured a permanent assistant professor position at the host institution.
Some scientific objectives could not fully be achieved, and significant deviations had to be explored, as explained in more detail in the final report. Specifically, the computational complexity of a key operation envisioned as part of this project turned out to be considerably
more involved than initially foreseen, leading to the exploration of alternative paradigms and algorithmic problems. Moreover, the (unfunded) experimental partners experienced difficulties while collecting SHAPE data, leading to a data collection effort that is still being completed at the moment.
The project was still productive, with significant algorithmic contributions being achieved in connection with polytope membership problems, RNA-RNA interactions prediction, RNA folding in extended conformational spaces and sequence design.
RNA Biology plays a central role in bioinformatics research such that numerous model-driven algorithms and methods are developed to predict and calculate structures and functions or compare sequences of RNA molecules that are essential in molecular and genetic processes and thus for medical and pharmaceutical applications.
Many of these problems can be cleanly phrased as optimization problems with 'optimal substructure' and thus solved exactly and efficiently by dynamic programming (for arbitrary parametrization of the objective
function). This project will systematically explore the impact of parameter changes on the quality of results of DP optimization methods, such as predictions of molecule structures or comparison of sequences. The
approach strongly relies on algebraic dynamic programming (ADP), which decouples the decomposition of the search space from the algebra used to compute a final result. Thus, the ADP framework provides a unified
setting and a generic implementation to quickly test working hypotheses. Here, it enables naturally implementing the suggested parametric optimization by developing novel algebras.
Those include the polytope algebra, which allows to segment the parameter space based on its impact of the final prediction, and a formal derivative algebra, which allows to compute the derivative of ensemble predictions with respect to a given parameter. Conversely, those methods can be
used to learn the optimal parameter sets based on a reference set of instances. This will result in deeper insights into robustness of the algorithms to changes of parameters or input data and hence, results can be assessed based on robustness measures and leads to the calculation of biologically more meaningful results.
The overall methodology will be applied to selected problems in RNA Bioinformatics.
At the end of the MSCA project, it can be concluded that project has achieved some of its objectives and milestones. The career related objectives have all been achieved as the researcher has secured a permanent assistant professor position at the host institution.
Some scientific objectives could not fully be achieved, and significant deviations had to be explored, as explained in more detail in the final report. Specifically, the computational complexity of a key operation envisioned as part of this project turned out to be considerably
more involved than initially foreseen, leading to the exploration of alternative paradigms and algorithmic problems. Moreover, the (unfunded) experimental partners experienced difficulties while collecting SHAPE data, leading to a data collection effort that is still being completed at the moment.
The project was still productive, with significant algorithmic contributions being achieved in connection with polytope membership problems, RNA-RNA interactions prediction, RNA folding in extended conformational spaces and sequence design.
The main results in line with the DoA include the exploration of the state of the art and current literature, as well as a thorough examination of putative next steps and possible implementations, namely, along with Doan Dai Nguyen, an excellent student doing his Bachelor thesis under the researcher’s supervision. The team showed that the initial strategy, i.e.studying parameters independently by setting others to their usual value, would lead to omitting a large area of the search space and hence, will not offer the theoretical guarantees sought in the initial research program. Hence, the project has been reoriented towards a more theoretical exploration of possible solutions, as explained in the final report. As a consequence, two manuscripts are currently written along the lines of the DoA, that will hopefully be submitted within the current year. Finally, the researcher participated in two thematically-close projects, whose results have been presented at WABI 2022 and WABI 2024, with extended journal versions currently being under review.
The project had high measurable impact in the researcher’s career. Namely, the researcher now holds a permanent position as assistant professor in the host institution.
Given the algorithmic difficulties encountered during the implementation, the contributions mainly pertain to algorithmic theory, and the effort currently resides at the research prototype/proof of concept level.
Translation into innovative methodologies, including medical applications, could be foreseen but would require extensive further developments;
The natural users of the algorithmic developments, and their implementations, are researchers in molecular biology/biochemistry interested in RNA folding, interaction and design, revolving around an early implementation of the RNA-RNA interaction prediction algorithm.
Given the algorithmic difficulties encountered during the implementation, the contributions mainly pertain to algorithmic theory, and the effort currently resides at the research prototype/proof of concept level.
Translation into innovative methodologies, including medical applications, could be foreseen but would require extensive further developments;
The natural users of the algorithmic developments, and their implementations, are researchers in molecular biology/biochemistry interested in RNA folding, interaction and design, revolving around an early implementation of the RNA-RNA interaction prediction algorithm.