We study the prediction of T-cell response for specific given peptides, which could, among other applications, be a crucial step towards the development of personalized cancer vaccines. It is a challenging task due to limited, heterogeneous training data featuring a multi-domain structure; such data entail the danger of shortcut learning, where models learn general characteristics of peptide sources, such as the source organism, rather than specific peptide characteristics associated with T-cell response. Using a transformer model for T-cell response prediction, we show that the danger of inflated predictive performance is not merely theoretical but occurs in practice. Consequently, we propose a domain-aware evaluation scheme. We then study different transfer learning techniques to deal with the multi-domain structure and shortcut learning. We demonstrate a per-source fine tuning approach to be effective across a wide range of peptide sources and further show that our final model is competitive with existing state-of-the-art approaches for predicting T-cell responses for human peptides.

To learn a Bayesian network structure from data, one popular approach is to maximize a decomposable likelihood-based score. While various scores have been proposed, they usually assume a uniform prior, or “penalty,” over the possible directed acyclic graphs (DAGs); relatively little attention has been paid to alternative priors. We investigate empirically several structure priors in combination with different scores, using benchmark data sets and data sets generated from benchmark networks. Our results suggest that, in practice, priors that strongly favor sparsity perform significantly better than the uniform prior or even the informed variant that is conditioned on the correct number of parents for each node. For an analytic comparison of different priors, we generalize a known recurrence equation for the number of DAGs to accommodate modular weightings of DAGs, a result that is also of independent interest.

MOTIVATION: Predictive models are a powerful tool for solving complex problems in computational biology. They are typically designed to predict or classify data coming from the same unknown distribution as the training data. In many real-world settings, however, uncontrolled biological or technical factors can lead to a distribution mismatch between datasets acquired at different times, causing model performance to deteriorate on new data. A common additional obstacle in computational biology is scarce data with many more features than samples. To address these problems, we propose a method for unsupervised domain adaptation that is based on a weighted elastic net. The key idea of our approach is to compare dependencies between inputs in training and test data and to increase the cost of differently behaving features in the elastic net regularization term. In doing so, we encourage the model to assign a higher importance to features that are robust and behave similarly across domains. RESULTS: We evaluate our method both on simulated data with varying degrees of distribution mismatch and on real data, considering the problem of age prediction based on DNA methylation data across multiple tissues. Compared with a non-adaptive standard model, our approach substantially reduces errors on samples with a mismatched distribution. On real data, we achieve far lower errors on cerebellum samples, a tissue which is not part of the training data and poorly predicted by standard models. Our results demonstrate that unsupervised domain adaptation is possible for applications in computational biology, even with many more features than samples. AVAILABILITY AND IMPLEMENTATION: Source code is available at https://github.com/PfeiferLabTue/wenda. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Modern exact algorithms for structure learning in Bayesian networks first compute an exact local score of every candidate parent set, and then find a network structure by combinatorial optimization so as to maximize the global score. This approach assumes that each local score can be computed fast, which can be problematic when the scarcity of the data calls for structured local models or when there are both continuous and discrete variables, for these cases have lacked efficient-to-compute local scores. To address this challenge, we introduce a local score that is based on a class of classification and regression trees. We show that under modest restrictions on the possible branchings in the tree structure, it is feasible to find a structure that maximizes a Bayes score in a range of moderate-size problem instances. In particular, this enables global optimization of the Bayesian network structure, including the local structure. In addition, we introduce a related model class that extends ordinary conditional probability tables to continuous variables by employing an adaptive discretization approach. The two model classes are compared empirically by learning Bayesian networks from benchmark real-world and synthetic data sets. We discuss the relative strengths of the model classes in terms of their structure learning capability, predictive performance, and running time.

Parsimonious context trees, PCTs, provide a sparse parameterization of conditional probability distributions. They are particularly powerful for modeling context-specific independencies in sequential discrete data. Learning PCTs from data is computationally hard due to the combinatorial explosion of the space of model structures as the number of predictor variables grows. Under the score-and-search paradigm, the fastest algorithm for finding an optimal PCT, prior to the present work, is based on dynamic programming. While the algorithm can handle small instances fast, it becomes infeasible already when there are half a dozen four-state predictor variables. Here, we show that common scoring functions enable the use of new algorithmic ideas, which can significantly expedite the dynamic programming algorithm on typical data. Specifically, we introduce a memoization technique, which exploits regularities within the predictor variables by equating different contexts associated with the same data subset, and a bound-and-prune technique, which exploits regularities within the response variable by pruning parts of the search space based on score upper bounds. On real-world data from recent applications of PCTs within computational biology the ideas are shown to reduce the traversed search space and the computation time by several orders of magnitude in typical cases.

To compare learning algorithms that differ by the adopted statistical paradigm, model class, or search heuristic, it is common to evaluate the performance on training data of varying size. Measuring the performance is straightforward if the data are generated from a known model, the ground truth. However, when the study concerns real-world data, the current methodology is limited to estimating predictive performance, typically by cross-validation. This work introduces a method to compare algorithms’ ability to learn the model structure, assuming no ground truth is given. The idea is to identify a partial structure on which the algorithms agree, and measure the performance in relation to that structure on subsamples of the data. The method is instantiated to structure learning in Bayesian networks, measuring the performance by the structural Hamming distance. It is tested using benchmark ground truth networks and algorithms that maximize various scoring functions. The results show that the method can produce evaluation outcomes that are close to those one would obtain if the ground truth was available.

Summary: Recent studies have shown that the traditional position weight matrix model is often insufficient for modeling transcription factor binding sites, as intra-motif dependencies play a significant role for an accurate description of binding motifs. Here, we present the Java application InMoDe, a collection of tools for learning, leveraging and visualizing such dependencies of putative higher order. The distinguishing feature of InMoDe is a robust model selection from a class of parsimonious models, taking into account dependencies only if justified by the data while choosing for simplicity otherwise. Availability and Implementation: InMoDe is implemented in Java and is available as command line application, as application with a graphical user-interface, and as an integration into Galaxy on the project website at http://www.jstacs.de/index.php/InMoDe . Contact: [email protected].

Context trees (CT) are a widely used tool in machine learning for representing context-specific independences in conditional probability distributions. Parsimonious context trees (PCTs) are a recently proposed generalization of CTs that can enable statistically more efficient learning due to a higher structural flexibility, which is particularly useful for small-data settings. However, this comes at the cost of a computationally expensive structure learning algorithm, which is feasible only for domains with small alphabets and tree depths. In this work, we investigate to which degree CTs can be generalized to increase statistical efficiency while still keeping the learning computationally feasible. Approaching this goal from two different angles, we (i) propose algorithmic improvements to the PCT learning algorithm, and (ii) study further generalizations of CTs, which are inspired by PCTs, but trade structural flexibility for computational efficiency. By empirical studies both on simulated and real-world data, we demonstrate that the synergy of combining of both orthogonal approaches yields a substantial improvement in obtaining statistically efficient and computationally feasible generalizations of CTs.

BACKGROUND: Statistical modeling of transcription factor binding sites is one of the classical fields in bioinformatics. The position weight matrix (PWM) model, which assumes statistical independence among all nucleotides in a binding site, used to be the standard model for this task for more than three decades but its simple assumptions are increasingly put into question. Recent high-throughput sequencing methods have provided data sets of sufficient size and quality for studying the benefits of more complex models. However, learning more complex models typically entails the danger of overfitting, and while model classes that dynamically adapt the model complexity to data have been developed, effective model selection is to date only possible for fully observable data, but not, e.g., within de novo motif discovery. RESULTS: To address this issue, we propose a stochastic algorithm for performing robust model selection in a latent variable setting. This algorithm yields a solution without relying on hyperparameter-tuning via massive cross-validation or other computationally expensive resampling techniques. Using this algorithm for learning inhomogeneous parsimonious Markov models, we study the degree of putative higher-order intra-motif dependencies for transcription factor binding sites inferred via de novo motif discovery from ChIP-seq data. We find that intra-motif dependencies are prevalent and not limited to first-order dependencies among directly adjacent nucleotides, but that second-order models appear to be the significantly better choice. CONCLUSIONS: The traditional PWM model appears to be indeed insufficient to infer realistic sequence motifs, as it is on average outperformed by more complex models that take into account intra-motif dependencies. Moreover, using such models together with an appropriate model selection procedure does not lead to a significant performance loss in comparison with the PWM model for any of the studied transcription factors. Hence, we find it worthwhile to recommend that any modern motif discovery algorithm should attempt to take into account intra-motif dependencies.

Inhomogeneous parsimonious Markov models have recently been introduced for modeling symbolic sequences, with a main application being DNA sequence analysis. Structure and parameter learning of these models has been proposed using a Bayesian approach, which entails the practically challenging choice of the prior distribution. Cross validation is a possible way of tuning the prior hyperparameters towards a specific task such as prediction or classification, but it is overly time-consuming. On this account, robust learning methods, which do not require explicit prior specification and – in the absence of prior knowledge – no hyperparameter tuning, are of interest. In this work, we empirically investigate the performance of robust alternatives for structure and parameter learning that extend the practical applicability of inhomogeneous parsimonious Markov models to more complex settings than before.