Biology subjects

Rousu, J.

Publications and source records attributed to Rousu, J..

14 recordsLinked to original sources

bioRxiv2026

Domain-adversarial learning predicts clinically actionable drug combination synergy in leukemia patients using bulk transcriptomics data

Deep learning has gained popularity in drug combination synergy prediction; however, DL models require large training datasets from cell line pharmacogenomic screens that poorly capture the heterogeneity in transcriptomic features and phenotypic responses seen in patients. To that end, we developed a domain-adversarial neural network (DANN) for personalized drug synergy prediction that accounts for systematic differences between cell line and patient domains. In applications to AML and CLL patient samples, we demonstrate how DANN boosts prediction accuracy under realistic data constraints. The model predictions demonstrated elevated synergy among clinically used combinations, such as venetoclax-based regimens, supporting its ability to identify both pharmaceutically and clinically meaningful combinations. DANN estimates prediction uncertainty and prioritizes high-confidence combination predictions to aid clinical translation. Together, DANN provides a systematic approach to improving accuracy and reliability in cross-domain drug synergy prediction, advancing the development of methods that are aligned with the translational requirements of precision hematology.

cancer biology↗

bioRxiv2025

Structured Chemical Reaction Modeling with Multitask Graph Neural Networks

Modeling chemical reactions requires connecting fine-grained atom-bond edits with broader semantic categories. Yet, most machine learning approaches model these aspects in isolation: atom mapping, reaction center identification, and reaction classification are treated as separate problems. This separation limits accuracy, interpretability, and generalization. In this work, we argue that multitask learning provides a natural and powerful framework for reaction modeling. By jointly predicting mappings, centers, and classes within a single graph neural network, models can leverage structural dependencies between tasks. We proposed MARCC (Mapping-Assisted Reaction Center and Classification), a multitask architecture that achieves state-of-the-art performance on the USPTO-50K benchmark. MARCC demonstrates that multitask supervision not only improves accuracy across all tasks but also provides a structured representation of reaction mechanisms.

bioinformatics↗

bioRxiv2024

Scaling up drug combination surface prediction

Drug combinations are required to treat advanced cancers and other complex diseases. Compared to monotherapy, combination treatments can enhance efficacy and reduce toxicity by lowering the doses of single drugs - and there especially synergistic combinations are of interest. Since drug combination screening experiments are costly and time consuming, reliable machine learning models are needed for prioritizing potential combinations for further studies. Most of the current machine learning models are based on scalar-valued approaches, which predict individual response values or synergy scores for drug combinations. We take a functional output prediction approach, in which full, continuous dose-response combination surfaces are predicted for each drug combination on the cell lines. We investigate the predictive power of the recently proposed comboKR method, which is based on a powerful input-output kernel regression technique and functional modelling of the response surface. In this work, we develop a scaled-up formulation of the comboKR, that also implements improved modeling choices: 1) we incorporate new modeling choices for the output drug combination response surfaces to the comboKR framework, and 2) propose a projected gradient descent method to solve the challenging pre-image problem that traditionally is solved with simple candidate set approaches. We provide thorough experimental analysis of comboKR 2.0 with three real-word datasets within various challenging experimental settings, including cases where drugs or cell lines have not been encountered in the training data. Our comparison with synergy score prediction methods further highlights the relevance of dose-response prediction approaches, instead of relying on simple scoring methods.

bioinformatics↗

bioRxiv2024

Predicting drug combination response surfaces

Prediction of drug combination responses is a research question of growing importance for cancer and other complex diseases. Current machine learning approaches generally consider predicting either drug combination synergy summaries or single combination dose-response values, which fail to appropriately model the continuous nature of the underlying dose-response combination surface and can lead to inconsistencies when a synergy score or a dose-response matrix is reconstructed from separate predictions. We propose a structured prediction method, comboKR, that directly predicts the drug combination response surface for a drug combination. The method is based on a powerful input-output kernel regression technique and functional modeling of the response surface. As an important part of our approach, we develop a novel normalisation between response surfaces that standardizes the heterogeneous experimental designs used to measure the dose-responses, and thus allows training the method with data measured in different laboratories. Our experiments on two predictive scenarios highlight the suitability of the proposed approach especially in the traditionally challenging setting of predicting combination responses for new drugs not available in the training data.

bioinformatics↗

bioRxiv2024

Attention-based approach to predict drug-target interactions across seven target superfamilies

MotivationDrug-target interactions (DTIs) hold pivotal role in drug repurposing and elucidation of drug mechanisms of action. While single-targeted drugs have demonstrated clinical success, they often exhibit limited efficacy against complex diseases, such as cancers, whose development and treatment is dependent on several biological processes. Therefore, a comprehensive understanding of primary, secondary, and even inactive targets becomes essential in the quest for effective and safe treatments for cancer and other indications. The human proteome offers over a thousand druggable targets. Yet, most FDA-approved drugs bind with only a small fraction of disease targets. ResultsThis study introduces an attention-based method to predict drug-target bioactivities for all human proteins across seven superfamilies. Nine different descriptor sets were meticulously examined to identify optimal signature descriptors for predicting novel DTIs. Our testing results demonstrated Spearman correlations exceeding 0.72 (P<0.001) for six out of seven superfamilies. The proposed method outperformed nine state-of-the-art deep learning and graph-based methods, and importantly, maintained relatively high performance for most target superfamilies when tested in independent sources of bioactivity data. We computationally validated 185,676 drug-target pairs from ChEMBL-V33, that were not available in model training, with a reasonable performance of Spearman correlation > 0.57 (P<0.001) for most superfamilies. This justifies the robustness of the proposed method for predicting novel DTIs. Finally, we applied our method to predict missing activities among 3,492 approved molecules in ChEMBL-V33, offering a valuable tool for advancing drug mechanism discovery and repurposing existing drugs for new indications. Availability and implementationThe codes and training datasets to reproduce and extend the results are available at https://github.com/AronSchulman/MMAtt-DTA.

systems biology↗

bioRxiv2022

Strain design optimization using reinforcement learning

Engineered microbial cells present a sustainable alternative to fossil-based synthesis of chemicals and fuels. Cellular synthesis routes are readily assembled and introduced into microbial strains using state-of-the-art synthetic biology tools. However, the optimization of the strains required to reach industrially feasible production levels is far less efficient. It typically relies on trial-and-error leading into high uncertainty in total duration and cost. New techniques that can cope with the complexity and limited mechanistic knowledge of the cellular regulation are called for guiding the strain optimization. In this paper, we put forward a multi-agent reinforcement learning (MARL) approach that learns from experiments to tune the metabolic enzyme levels so that the production is improved. Our method is model-free and does not assume prior knowledge of the microbes metabolic network or its regulation. The multi-agent approach is well-suited to make use of parallel experiments such as multi-well plates commonly used for screening microbial strains. We demonstrate the methods capabilities using the genome-scale kinetic model of Escherichia coli, k-ecoli457, as a surrogate for an in vivo cell behaviour in cultivation experiments. We investigate the methods performance relevant for practical applicability in strain engineering i.e. the speed of convergence towards the optimum response, noise tolerance, and the statistical stability of the solutions found. We further evaluate the proposed MARL approach in improving L-tryptophan production by yeast Saccharomyces cerevisiae, using publicly available experimental data on the performance of a combinatorial strain library. Overall, our results show that multi-agent reinforcement learning is a promising approach for guiding the strain optimization beyond mechanistic knowledge, with the goal of faster and more reliably obtaining industrially attractive production levels. Author summaryEngineered microbial cells offer a sustainable alternative solution to chemical production from fossil resources. However, to make the chemical production using microbial cells economically feasible, they need to be substantially optimized. Due to the biological complexity, this optimization to reach sufficiently high production is typically a costly trial and error process. This paper presents an Artificial Intelligence (AI) approach to guide this task. Our tool learns a model from previous experiments and uses the model to suggest improvements to the engineering design, until a satisfactory production performance is reached. This paper evaluates the behaviour of the proposed AI method from several angles, including the amount of experiments needed, the tolerance to noise as well as the stability of the proposed designs.

bioinformatics↗

bioRxiv2022

Joint structural annotation of small molecules using liquid chromatography retention order and tandem mass spectrometry data

We present LC-MS2Struct, a machine learning framework for structural annotation of small molecule data arising from liquid chromatography-tandem mass spectrometry (LC-MS2) measurements. LC-MS2Struct jointly predicts the annotations for a set of mass spectrometry features in a sample, using a novel structured prediction model trained to optimally combine the output of state-of-the-art MS2 scorers and observed retention orders. We evaluate our method on a dataset covering all publicly available reversed phase LC-MS2 data in the MassBank reference database, including 4327 molecules measured using 18 different LC conditions from 16 contributors, greatly expanding the chemical analytical space covered in previous multi-MS2 scorer evaluations. LC-MS2Struct obtains significantly higher annotation accuracy than earlier methods and improves the annotation accuracy of state-of-the-art MS2 scorers by up to 106%. The use of stereochemistry-aware molecular fingerprints improves prediction performance, which highlights limitations in existing approaches and has strong implications for future computational LC-MS2 developments.

bioinformatics↗

bioRxiv2021

Modeling drug combination effects via latent tensor reconstruction

AO_SCPLOWBSTRACTC_SCPLOWO_ST_ABSMotivationC_ST_ABSCombination therapies have emerged as a powerful treatment modality to overcome drug resistance and improve treatment efficacy. However, the number of possible drug combinations increases very rapidly with the number of individual drugs in consideration which makes the comprehensive experimental screening infeasible in practice. Machine learning models offer time- and cost-efficient means to aid this process by prioritising the most effective drug combinations for further pre-clinical and clinical validation. However, the complexity of the underlying interaction patterns across multiple drug doses and in different cellular contexts poses challenges to the predictive modelling of drug combination effects. ResultsWe introduce comboLTR, highly time-efficient method for learning complex, nonlinear target functions for describing the responses of therapeutic agent combinations in various doses and cancer cell-contexts. The method is based on a polynomial regression via powerful latent tensor reconstruction. It uses a combination of recommender system-style features indexing the data tensor of response values in different contexts, and chemical and multi-omics features as inputs. We demonstrate that comboLTR outperforms state-of-the-art methods in terms of predictive performance and running time, and produces highly accurate results even in the challenging and practical inference scenario where full dose-response matrices are predicted for completely new drug combinations with no available combination and monotherapy response measurements in any training cell line. Availability and implementationcomboLTR code is available at https://github.com/aalto-ics-kepaco/ComboLTR Contacttianduanyi.wang@aalto.fi; juho.rousu@aalto.fi

bioinformatics↗

bioRxiv2020

comboFM: leveraging multi-way interactions for systematic prediction of drug combination effects

We present comboFM, a machine learning framework for predicting the responses of drug combinations in preclinical studies, such as those based on cell lines or patient-derived cells. comboFM models the cell context-specific drug interactions through higher-order tensors, and efficiently learns latent factors of the tensor using powerful factorization machines. The approach enables comboFM to leverage information from previous experiments performed on similar drugs and cells when predicting responses of new combinations in so far untested cells; thereby, it achieves highly accurate predictions despite sparsely populated data tensors. We demonstrated high predictive performance of comboFM in various prediction scenarios using data from cancer cell line drug screening. Subsequent experimental validation of a set of previously untested drug combinations further supported the practical and robust applicability of comboFM. For instance, we confirmed a novel synergy between anaplastic lymphoma kinase (ALK) inhibitor crizotinib and proteasome inhibitor bortezomib in lymphoma cells. Overall, our results demonstrate that comboFM provides an effective means for systematic pre-screening of drug combinations to support precision oncology applications.

bioinformatics↗

bioRxiv2020

Probabilistic Framework for Integration of Mass Spectrum and Retention Time Information in Small Molecule Identification

MotivationIdentification of small molecules in a biological sample remains a major bottleneck in molecular biology, despite a decade of rapid development of computational approaches for predicting molecular structures using mass spectrometry (MS) data. Recently, there has been increasing interest in utilizing other information sources, such as liquid chromatography (LC) retention time (RT), to improve the MS based identifications. ResultsWe put forward a probabilistic modelling framework to integrate MS and RT data of multiple features in an LC-MS experiment. We model the MS measurements and all pairwise retention order information as a Markov random field and use efficient approximate inference for scoring and ranking potential molecular structures. Our experiments show improved identification accuracy by combining tandem mass spectrometry data (MS2) and retention orders using our approach, thereby outperforming state-of-the-art methods. Furthermore, we demonstrate the benefit of our model when only a subset of LC-MS features have MS2 measurements available besides MS1. Availability and implementationSoftware and data is freely available at https://github.com/aalto-ics-kepaco/msms_rt_score_integration. Contacteric.bach@aalto.fi

systems biology↗

bioRxiv2020

Ranking microbial metabolomic and genomic links using correlation-based and feature-based scoring functions

Specialised metabolites from microbial sources are well-known for their wide range of biomedical applications, particularly as antibiotics. When mining paired genomic and metabolomic data sets for novel specialised metabolites, establishing links between Biosynthetic Gene Clusters (BGCs) and metabolites represents a promising way of finding such novel chemistry. However, due to the lack of detailed biosynthetic knowledge for the majority of predicted BGCs, and the large number of possible combinations, this is not a simple task. This problem is becoming ever more pressing with the increased availability of paired omics data sets. Current tools are not effective at identifying valid links automatically, and manual verification is a considerable bottleneck in natural product research. We demonstrate that using multiple link-scoring functions together makes it easier to prioritise true links relative to others. Based on standardising a commonly used score, we introduce a new, more effective score, and introduce a novel score using an Input-Output Kernel Regression approach. Finally, we present NPLinker, a software framework to link genomic and metabolomic data. Results are verified using publicly available data sets that include validated links. Author summaryIn this article, we introduce NPLinker, a software framework to link genomic and metabolomic data, to link microbial secondary metabolites to their producing genomic regions. Two of the major approaches for such linking are analysis of the correlation between sets of strains, and analysis of predicted features of the molecules. While these methods are usually used separately, we demonstrate that they are in fact complementary, and show a way to combine them to improve their performance. We begin by demonstrating a weakness in the most common method of strain correlation analysis, and suggest an improvement. We then introduce a new feature-based analysis method which, unlike most such methods, does not directly depend on the natural prodcut compound class. Finally, we demonstrate that the two are complementary and proceed to combine them into a single scoring function for genomic and metabolomic links, which shows improved performance over either of the individual approaches. Verification is done using curated databases of genomic and metabolomic data, as well as public data sets of microbial data including verified links.

systems biology↗

bioRxiv2020

Classes for the masses: Systematic classification of unknowns using fragmentation spectra

Metabolomics experiments can employ non-targeted tandem mass spectrometry to detect hundreds to thousands of molecules in a biological sample. Structural annotation of molecules is typically carried out by searching their fragmentation spectra in spectral libraries or, recently, in structure databases. Annotations are limited to structures present in the library or database employed, prohibiting a thorough utilization of the experimental data. We present a computational tool for systematic compound class annotation: CANOPUS uses a deep neural network to predict 1,270 compound classes from fragmentation spectra, and explicitly targets compounds where neither spectral nor structural reference data are available. CANOPUS even predicts classes for which no MS/MS training data are available. We demonstrate the broad utility of CANOPUS by investigating the effect of the microbial colonization in the digestive system in mice, and through analysis of the chemodiversity of different Euphorbia plants; both uniquely revealing biological insights at the compound class level.

bioinformatics↗

bioRxiv2017

Principal Metabolic Flux Mode Analysis

MotivationIn the analysis of metabolism using omics data, two distinct and complementary approaches are frequently used: Principal component analysis (PCA) and Stoichiometric flux analysis. PCA is able to capture the main modes of variability in a set of experiments and does not make many prior assumptions about the data, but does not inherently take into account the flux mode structure of metabolism. Stoichiometric flux analysis methods, such as Flux Balance Analysis (FBA) and Elementary Mode Analysis, on the other hand, produce results that are readily interpretable in terms of metabolic flux modes, however, they are not best suited for exploratory analysis on a large set of samples.\n\nResultsWe propose a new methodology for the analysis of metabolism, called Principal Metabolic Flux Mode Analysis (PMFA), which marries the PCA and Stoichiometric flux analysis approaches in an elegant regularized optimization framework. In short, the method incorporates a variance maximization objective form PCA coupled with a Stoichiometric regularizer, which penalizes projections that are far from any flux modes of the network. For interpretability, we also introduce a sparse variant of PMFA that favours flux modes that contain a small number of reactions. Our experiments demonstrate the versatility and capabilities of our methodology.\n\nAvailabilityMatlab software for PMFA and SPMFA is available in https://github.com/ aalto-ics-kepaco/PMFA.\n\nContactsahely@iitpkd.ac.in, juho.rousu@aalto.fi, Peter.Blomberg@vtt.fi, Sandra.Castillo@vtt.fi\n\nSupplementary informationDetailed results are in Supplementary files. Supplementary data are available at https://github.com/aalto-ics-kepaco/PMFA/blob/master/Results.zip.

bioinformatics↗

bioRxiv2017

Predicting Protein Producibility In Filamentous Fungi

In this paper we study the problem of predicting the producibility of recombinant proteins in filamentous fungi, especially T. reesei, using machine learning methods. We train supervised and semi-supervised support vector machines with protein sequences, represented by their amino acid composition as well as protein family and domain information. Our results indicate, somewhat surprisingly, that quite modest amount of proteins with experimental data are required to build a state-of-the-art classifier and that additional unlabeled sequences in semi-supervised models do not bring increased predictive performance. Our experiments in cross-species prediction show that models trained for the filamentous fungus A. niger protein dataset can be generalized to predict protein producibility in T. reesei, and vice versa, without sacrificing too much accuracy, regardless of their approximately 500 millions years of divergence. However, predictors trained on E. coli and S. cerevisiae datasets gave variable performance when applied to the filamentous fungi datasets, indicating that while protein producibility prediction can be generalized accross related species, fully generic prediction tools applicable to any protein production host may not be realistic to achieve.

bioinformatics↗