Biology subjects

Kroll, A.

Publications and source records attributed to Kroll, A..

8 recordsLinked to original sources

bioRxiv2026

Predicting turnover number fold-changes to recover true mutation effects and overcome biases in mutant datasets

Machine learning is increasingly used to guide protein engineering by predicting how mutations affect desired properties. Recent models for the turnover number (kcat) of enzymes report high accuracy, suggesting that mutation effects can be inferred directly from protein sequence. However, these approaches are typically evaluated on heterogeneous datasets of enzyme variants, where closely related sequences and systematic reporting patterns may confound model performance. A central challenge is therefore to determine whether current models truly capture mutation-specific effects or instead exploit statistical regularities in the data. Here we show that much of the reported accuracy in mutant kcat prediction arises from two pervasive biases: variants of the same enzyme occupy a narrow activity range, and mutations within a group often share a common direction of change. Simple baselines that exploit these biases match or exceed the performance of existing models, indicating that high apparent accuracy does not imply mechanistic understanding. To address this limitation, we introduce a bias-aware framework that reformulates prediction as a pairwise fold-change task and evaluates performance on unseen mutant-mutant pairs, thereby isolating mutation-specific signal. A proof-of-principle implementation explains approximately one-third of the variance under these conditions and outperforms existing models on leakage-controlled benchmarks. More broadly, this work establishes a general strategy for evaluating and modeling mutation effects in biochemical datasets, with implications for protein engineering and related fields.

bioinformatics↗

bioRxiv2026

FoldVision: A compute-efficient atom-level 3D protein encoder

Protein function emerges from three-dimensional structure, yet many large-scale protein prediction pipelines still rely solely on linear sequence embeddings. Although multiple structure-aware protein networks have been proposed, they often omit atom-level details and struggle to capture the detailed chemistry of binding sites. Here, we introduce FoldVision, a compute-efficient 3D convolutional neural network that voxelizes every heavy atom, learns rotation-robust representations, and is pre-trained on over 500 000 AlphaFold-2 structures, which is more than two orders of magnitude less data than used to train modern protein language models. Despite its compact size of 123 million parameters, FoldVision outperforms or matches state-of-the-art protein encoders on four benchmarks that require fine structural resolution: enzyme-substrate classification, transporter-substrate classification, drug-kinase inhibition, and drug-target inhibition prediction. A simple ensemble with a sequence-based model consistently improves performance across all benchmarks beyond any individual model. This indicates that FoldVision learns structural signals that are complementary to those extracted by sequence-based models. This study demonstrates that full-atom protein 3D CNNs are both tractable and superior to protein language models alone for structure-dependent tasks.

biochemistry↗

bioRxiv2023

A general prediction model for substrates of transport proteins

Transport proteins play a crucial role in cellular metabolism and are central to many aspects of molecular biology and medicine. Determining the function of transport proteins experimentally is challenging, as they become unstable when isolated from cell membranes. Machine learning-based predictions could provide an efficient alternative. However, existing methods are limited to predicting a small number of specific substrates or broad transporter classes. These limitations stem partly from using small datasets for model training and a choice of input features that lack sufficient information about the prediction problem. Here, we present SPOT, the first general machine learning model that can successfully predict specific substrates for arbitrary transport proteins, achieving an accuracy above 92% on independent and diverse test data covering widely different transporters and a broad range of metabolites. SPOT uses Transformer Networks to represent transporters and substrates numerically. To overcome the problem of missing negative data for training, it augments a large dataset of known transporter-substrate pairs with carefully sampled random molecules as non-substrates. SPOT not only predicts specific transporter-substrate pairs, but also outperforms previously published models designed to predict broad substrate classes for individual transport proteins. We provide a web server and Python function that allows users to explore the substrate scope of arbitrary transporters.

bioinformatics↗

bioRxiv2023

A multimodal Transformer Network for protein-small molecule interactions enhances drug-target affinity and enzyme-substrate predictions

Most drugs are small molecules, with their activities typically arising from interactions with protein targets. Accurate predictions of these interactions could greatly accelerate pharmaceutical research. Current machine learning models designed for this task have a limited ability to generalize beyond the proteins used for training. This limitation is likely due to a lack of information exchange between the protein and the small molecule during the generation of the required numerical representations. Here, we introduce ProSmith, a machine learning framework that employs a multimodal Transformer Network to simultaneously process protein amino acid sequences and small molecule strings in the same input. This approach facilitates the exchange of all relevant information between the two types of molecules during the computation of their numerical representations, allowing the model to account for their structural and functional interactions. Our final model combines gradient boosting predictions based on the resulting multimodal Transformer Network with independent predictions based on separate deep learning representations of the proteins and small molecules. The resulting predictions outperform all previous models for predicting drug-target interactions, and the model demonstrates unprecedented generalization capabilities to unseen proteins. We further show that the superior performance of ProSmith is not limited to drug-target interaction predictions, but also leads to improvements in other protein-small molecule interaction prediction tasks, the prediction of Michaelis constants KM of enzyme-substrate pairs and the identification of potential substrates for enzymes. The Python code provided can be used to easily implement and improve machine learning predictions of interactions between proteins and arbitrary drug candidates or other small molecules.

bioinformatics↗

bioRxiv2023

Machine learning models for the prediction of enzyme properties should be tested on proteins not used for model training

The recently published DLKcat model, a deep learning approach for predicting enzyme turnover numbers (kcat), claims to enable high-throughput kcat predictions for metabolic enzymes from any organism and to capture kcat changes for mutated enzymes. Here, we critically evaluate these claims. We show that DLKcat predictions become positively misleading for enzymes with less than 60% sequence identity to the training data, performing worse than simply assuming a mean kcat value for all reactions. Furthermore, DLKcats ability to predict mutation effects is much weaker than implied, capturing only 3% of the experimentally observed variation across mutants not included in the training data. These findings highlight significant limitations in DLKcats generalizability and its practical utility for predicting kcat values for novel enzyme families or mutants, which are crucial applications in fields such as metabolic modeling.

bioinformatics↗

bioRxiv2022

Turnover number predictions for kinetically uncharacterized enzymes using machine and deep learning

The turnover number kcat, a measure of enzyme efficiency, is central to understanding cellular physiology and resource allocation. As experimental kcat estimates are unavailable for the vast majority of enzymatic reactions, the development of accurate computational prediction methods is highly desirable. However, existing machine learning models are limited to a single, well-studied organism, or they provide inaccurate predictions except for enzymes that are highly similar to proteins in the training set. Here, we present TurNuP, a general and organism-independent model that successfully predicts turnover numbers for natural reactions of wild-type enzymes. We constructed model inputs by representing complete chemical reactions through difference fingerprints and by representing enzymes through a modified and re-trained Transformer Network model for protein sequences. TurNuP outperforms previous models and generalizes well even to enzymes that are not similar to proteins in the training set. Parameterizing metabolic models with TurNuP-predicted kcat values leads to improved proteome allocation predictions. To provide a powerful and convenient tool for the study of molecular biochemistry and physiology, we implemented a TurNuP web server at https://turnup.cs.hhu.de.

bioinformatics↗

bioRxiv2022

The substrate scopes of enzymes: a general prediction model based on machine and deep learning

For a comprehensive understanding of metabolism, it is necessary to know all potential substrates for each enzyme encoded in an organisms genome. However, for most proteins annotated as enzymes, it is unknown which primary and/or secondary reactions they catalyze [1], as experimental characterizations are time-consuming and costly. Machine learning predictions could provide an efficient alternative, but are hampered by a lack of information regarding enzyme non-substrates, as available training data comprises mainly positive examples. Here, we present ESP, a general machine learning model for the prediction of enzyme-substrate pairs, with an accuracy of over 90% on independent and diverse test data. This accuracy was achieved by representing enzymes through a modified transformer model [2] with a trained, task-specific token, and by augmenting the positive training data by randomly sampling small molecules and assigning them as non-substrates. ESP can be applied successfully across widely different enzymes and a broad range of metabolites. It outperforms recently published models designed for individual, well-studied enzyme families, which use much more detailed input data [3, 4]. We implemented a user-friendly web server to predict the substrate scope of arbitrary enzymes, which may support not only basic science, but also the development of pharmaceuticals and bioengineering processes.

bioinformatics↗

bioRxiv2020

Prediction of Michaelis constants from structuralfeatures using deep learning

The Michaelis constant KM describes the affinity of an enzyme for a specific substrate, and is a central parameter in studies of enzyme kinetics and cellular physiology. As measurements of KM are often difficult and time-consuming, experimental estimates exist for only a minority of enzyme-substrate combinations even in model organisms. Here, we build and train an organism-independent model that successfully predicts KM values for natural enzyme-substrate combinations using machine and deep learning methods. Predictions are based on a task-specific molecular fingerprint of the substrate, generated using a graph neural network, and the domain structure of the enzyme. Model predictions can be used to estimate enzyme efficiencies, to relate metabolite concentrations to cellular physiology, and to fill gaps in the parameterization of kinetic models of cellular metabolism.

systems biology↗