IDA - Department of Computer and Information Science

We present a novel method that ranks a set of user specified drugs based on user specified single cell data from one or more sick individuals and knowledge about the regulatory networks and the targets of the drugs in these networks. The most efficient drug is ranked first, and least efficient drug is ranked last. The efficiency of the method is demonstrated by in vitro (in lab) and in vivo (mouse) experiments.

Multicellular Disease Model of multiple tissues reveals that inflammations switch on/off based on altered balance between pro- and anti-inflammatory upstream regulators  and downstream pathways

Here we propose a new approach based on Multicellular Disease Models to detect early biomarkers of the disease. The approach employs longitudinal single cell data for patients and controls.

Do you have data with many features where some or all features are observed subject to detection limits, for example some features are not observed below some threshold? This publication may help you to build your prediction model more efficiently! The method is based on a combination of an 'improper' multiple imputation approach and subspace k-Nearest Neighbors.

This paper uses conditional inference tree framework to build an interpretable machine learning model of adolescent pregnancy based on the data from the Northern Nicaragua.

In this paper we search for biomarkers of nasal polyps by studying regulatory mechanisms. It is concluded that there are multiple components, and combinations of biomarkers is needed for successful diagnostics

This paper describes a generic framework called Digital Twins that can be used to create "digital copies" of any individual patient, treat the person computationally with thousands of different drugs, and then select a treatment that would be optimal for that particular patient. Possible ways to implement this generic framework are explained.

This is an applied study where k-means algorithm is applied to discover clusters of various variables associated with poverty. Descriptive analyses are done to characterize the obtained clusters.

In this applied study, conditional inference trees were used to build a decision tree model that predicts poverty status from multiple predictors, such as education, emigration and food consumption.

This is an applied study where conditional inference random forests are applied to identify the most important risk factors of child stunting. In addition, conditional inference trees are used to build interpretable predictive models.

Do you have data in which multiple treatments or interventions were applied to a group of individuals and you want to discover subgroups that share similar characteristics and that also benefit from the same kinds of treatments? A novel interpretable machine learning (decision tree) framework is proposed in this paper to discover such subgroups and estimate the probabilities of certain treatments to be the best ones in these subgroups.

These two publications introduce a new regression method to estimate the target variable as a smooth and monotonically increasing function of the predictor. The probabilistic model and hyperparameter selection strategy is presented in 2018 paper, while some additional mathematical proofs regarding complexity estimates are given in 2017 paper.

Do you have two observed features that correspond to two latent (unobserved) features and you want to find out whether the latent features are monotonically related to each other? The hypothesis testing framework developed in this paper can help you! The method is based on a modification of the monotonic regression approach and bootstrap methods.

The paper introduces new methods to compute confidence intervals for monotonic regression in a multivariate setting. The methods are based on bootstrap and can handle large data sets and focus on inference of the expected target.

If AIC criterion is used for model selection, it is important to be able to estimate variance of the error in the regression model. This work introduces a variance estimator for monotonic regression models, and this estimator can be used for example for feature selection in monotonic regression models. The estimator has been demonstrated to have good finite sample properties.

A new method for solving large-scale monotonic regression problems is proposed here. The approach acts by splitting the data into smaller segments, solving regression problem in these segments, and then modifying these local solutions in a special manner into a global monotonic prediction.

• (2006) Burdakov, A. Grimvall and O. Sysoev. Data preordering in generalized PAV algorithm for monotonic regression. Journal of Computational Mathematics. 24, No. 6, pp. 771-790.
• (2006) Burdakov O. , Sysoev O. ,Grimvall A. and Hussian M. An O(n2) algorithm for isotonic regression. In: G. Di Pillo and M. Roma (Eds) Large-Scale Nonlinear Optimization. Series: Nonconvex Optimization and Its Applications, Springer-Verlag, 83, pp. 25-33.
• (2005) Hussian M. ,Grimvall A. ,Burdakov O. and Sysoev O. Monotonic regression for the detection of temporal trends in environmental quality data. MATCH Commun. Math. Comput. Chem. 54, pp. 535-550.

Do you have data in which target variable is expected to be an increasing or decreasing function of multiple features, but the observed data is not monotonic due to noise? Our publication from 2006 proposes a new method which has quadratic complexity and is able to find an approximate solution to the given monotonic regression problem. The choice of hyperparameter (data preordering) is investigated in our second 2006 publication. An empirical study using the general methodology was published already in 2005.

Do you have large-scale data with thousands of observations and hundreds of features in which some or all features are censored subject to detection limits, for example they are not observed below some threshold? We propose a new variational autoencoder framework that is able to handle that scenario! It can find a latent representation of the features, denoise the features and make probabilistic predictions of the target variable, including uncertainty estimates.

Single cell data is composed of cells that belong to different cell types which are not measured. To define the cell types for each single cell, we propose to use other kind of data (bulk data) where the observations are labeled with their cell types. In this paper, a novel machine learning method is proposed to "project" bulk data into clusters of single cells and, by doing this, predicting the cell types of single cells from these clusters.

In this paper, we aim to predict whether the signal on the current frequency is stronger than the signal on an alternative frequency, based on the current frequency signal characteristics. This is formulated as a classification problem which is modeled by various machine learning methods such as random forests and neural networks. The quality of the prediction appears to be poor, and we introduce a duo-threshold approach which is based on the rejection option principle. This allows us to increase the accuracy of predictions dramatically while the costs increase only marginally.