RNA Design is a crucial bioinformatics problem to tailor specific RNA sequences into structures that guide our biology and medicine. Because these computer generated sequences are often statistically different from observed RNA sequence and do not fold as intended in real lab conditions, methods are developed to ensure more biological consistency. One approach is substructure restriction where any solutions are composed of observed RNA substructures recorded in a database. However, studies on this restricted problem are limited: while our Applied Research Lab's Simulated Annealing solution (SIMARD) uses substructures it only considers a single mutation policy and single method of generating substructure consistent sequences. We therefore propose two new policies of mutation: uniform, randomly modifying any structure, and length proportional, substructures are swapped randomly in proportion to their RNA length to target substructures that cover the most bases of a problem. In experiments on roughly fifty RNA Design problems, we conclude the potential value of these substructure based mutation methods resulting in solutions potentially hundreds of bases closer to the target folded structure than previous Simulated Annealing solutions.

RNA Design is an essential problem in bioinformatics to tailor RNA Sequences that guide our biology and medicine. While there are many RNA Design solutions in literature, the field has been limited by the high runtime to estimate how a candidate RNA sequence will fold, especially with "long sequences" over 750 "bases"; as these folding algorithms are called millions of times per RNA Design problem, this limits RNA Design to shorter sequences despite the prevalence of long sequences in nature; for example, computationally designing COVID's over 20000 bases RNA sequence is not feasible with current RNA Design algorithms. To address this issue, we are the first work to integrate LinearFold, a faster RNA prediction algorithm used by Baidu to analyze COVID with a higher efficiency, into RNA Design. We compare the runtime and solution quality of our Applied Research Lab's Simulated Annealing solution (SIMARD) with and without LinearFold to design sequences thousands of bases long that timeout after weeks of runtime with current algorithms. We also survey challenges in terms of solution quality and adjusting the cooling schedule parameters. This work is thus a first step into RNA Design for longer RNA sequences.

Computational Intelligence is frequently applied to solve RNA design problems, to construct RNA sequences that fold into biochemically useful structures and alignments. RNA Design's NP-Hardness means that heuristic solutions, such as Evolutionary algorithms' Simulated Annealing, are commonly used to more effectively search for RNA sequences that fold into the target structure. Examples of Simulated Annealing in RNA Design include SIMARD, the ERD approach, and RNAPredict, all which aim to return RNA Sequences as close as possible to the target structure. However, such methods only use a single simulated annealing cooling schedule even though literature covers many schedules with varied convergence and performances guarantees. Since existing RNA Design cooling schedule surveys only cover at most four RNA design problems over two simulated annealing variants, we investigate the performance of four major simulated annealing schedules with ten variants on twenty-nine RNA design sequences. Relevant findings include a) the insensitivity of geometric schedule parameters, b) that logarithmic cooling schedules can solve RNA Design problems not solved by other schedules, c) suggestions for adjusting geometric schedule stopping conditions, and d) identifying common issues in popular adaptive and non-adaptive schedules for RNA Design.

We formulate the Cost-Sensitive Learning to Rank problem of learning to prioritize limited resources to mitigate the most costly outcomes. We develop improved ranking models to solve this problem, as verified by experiments in diverse domains such as forest fire prevention, crime prevention, and preventing storm caused outages in electrical networks.

In real life applications, we often face the following risk clearance problem: given a set of instances with a known numeric outcome Y (e.g., a toxicity level), we want to learn a model to identify new instances that have a low risk, i.e., the probability of the Y value exceeding a certain maximum MAX is less than some threshold t. This problem guarantees that the cleared instances have the minimum precision of 1 – t for Y ≤ MAX. By clearing such low risk instances, we can allocate costly resources to the remaining high risk instances. In this work, we formulate this problem as Risk Clearance with a goal of maximizing the clearance of low risk instances. Existing classification models fail to solve Risk Clearance adequately, so we develop algorithms designed specifically for this problem. We then validate that our approach improves on existing work via experiments on an industrial case study.

Identifying bad objects hidden amidst many good objects is important for public safety and decision-making. These problems are complicated in that the cost of leaving a bad object unidentified may not be specified easily, making it difficult to apply existing cost-sensitive classification that depends on knowing a cost matrix or cost distribution. A compelling case for this "illusive cost" issue is presented in our project of identifying contaminated transformers with an industrial partner. To address this problem, we present an alternative formulation of cost-sensitive classification, Clearance Under Threshold (CUT) Classification. Given a training set, CUT classification is to partition the attribute space such that a partition is cleared if the probability of a future object in this partition being bad is less than a user-specified threshold. The goal is to clear many low-risk objects so that users can more effectively target high-risk objects. We present a solution to this problem and evaluate it on a case study for clearing contaminated transformers and on public benchmarks from UC Irvine's Machine Learning Repository. According to the experiments, our algorithms performed far better than the baselines derived from previous classification approaches.