Deep learning models are seeing increased use as methods to predict mutational effects or allowed mutations in proteins. The models commonly used for these purposes include large language models (LLMs) and 3D Convolutional Neural Networks (CNNs). These two model types have very different architectures and are commonly trained on different representations of proteins. LLMs make use of the transformer architecture and are trained purely on protein sequences whereas 3D CNNs are trained on voxelized representations of local protein structure. While comparable overall prediction accuracies have been reported for both types of models, it is not known to what extent these models make comparable specific predictions and/or generalize protein biochemistry in similar ways. Here, we perform a systematic comparison of two LLMs and two structure-based models (CNNs) and show that the different model types have distinct strengths and weaknesses. The overall prediction accuracies are largely uncorrelated between the sequence- and structure-based models. Overall, the two structure-based models are better at predicting buried aliphatic and hydrophobic residues whereas the two LLMs are better at predicting solvent-exposed polar and charged amino acids. Finally, we find that a combined model that takes the individual model predictions as input can leverage these individual model strengths and results in significantly improved overall prediction accuracy.
- Award ID(s):
- 1660648
- NSF-PAR ID:
- 10381983
- Editor(s):
- Fariselli, Piero
- Date Published:
- Journal Name:
- PLOS Computational Biology
- Volume:
- 16
- Issue:
- 11
- ISSN:
- 1553-7358
- Page Range / eLocation ID:
- e1008291
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
Abstract -
de Meaux, Juliette (Ed.)Abstract Genetic redundancy refers to a situation where an individual with a loss-of-function mutation in one gene (single mutant) does not show an apparent phenotype until one or more paralogs are also knocked out (double/higher-order mutant). Previous studies have identified some characteristics common among redundant gene pairs, but a predictive model of genetic redundancy incorporating a wide variety of features derived from accumulating omics and mutant phenotype data is yet to be established. In addition, the relative importance of these features for genetic redundancy remains largely unclear. Here, we establish machine learning models for predicting whether a gene pair is likely redundant or not in the model plant Arabidopsis thaliana based on six feature categories: functional annotations, evolutionary conservation including duplication patterns and mechanisms, epigenetic marks, protein properties including post-translational modifications, gene expression, and gene network properties. The definition of redundancy, data transformations, feature subsets, and machine learning algorithms used significantly affected model performance based on hold-out, testing phenotype data. Among the most important features in predicting gene pairs as redundant were having a paralog(s) from recent duplication events, annotation as a transcription factor, downregulation during stress conditions, and having similar expression patterns under stress conditions. We also explored the potential reasons underlying mispredictions and limitations of our studies. This genetic redundancy model sheds light on characteristics that may contribute to long-term maintenance of paralogs, and will ultimately allow for more targeted generation of functionally informative double mutants, advancing functional genomic studies.more » « less
-
Abstract Advances in genome sequencing and annotation have eased the difficulty of identifying new gene sequences. Predicting the functions of these newly identified genes remains challenging. Genes descended from a common ancestral sequence are likely to have common functions. As a result, homology is widely used for gene function prediction. This means functional annotation errors also propagate from one species to another. Several approaches based on machine learning classification algorithms were evaluated for their ability to accurately predict gene function from non‐homology gene features. Among the eight supervised classification algorithms evaluated, random‐forest‐based prediction consistently provided the most accurate gene function prediction. Non‐homology‐based functional annotation provides complementary strengths to homology‐based annotation, with higher average performance in Biological Process GO terms, the domain where homology‐based functional annotation performs the worst, and weaker performance in Molecular Function GO terms, the domain where the accuracy of homology‐based functional annotation is highest. GO prediction models trained with homology‐based annotations were able to successfully predict annotations from a manually curated “gold standard” GO annotation set. Non‐homology‐based functional annotation based on machine learning may ultimately prove useful both as a method to assign predicted functions to orphan genes which lack functionally characterized homologs, and to identify and correct functional annotation errors which were propagated through homology‐based functional annotations.
-
Structural information of biological macromolecules is crucial and necessary to deliver predictions about the effects of mutations—whether polymorphic or deleterious (i.e., disease causing), wherein, thermodynamic parameters, namely, folding and binding free energies potentially serve as effective biomarkers. It may be emphasized that the effect of a mutation depends on various factors, including the type of protein (globular, membrane or intrinsically disordered protein) and the structural context in which it occurs. Such information may positively aid drug-design. Furthermore, due to the intrinsic plasticity of proteins, even mutations involving radical change of the structural and physico–chemical properties of the amino acids (native vs. mutant) can still have minimal effects on protein thermodynamics. However, if a mutation causes significant perturbation by either folding or binding free energies, it is quite likely to be deleterious. Mitigating such effects is a promising alternative to the traditional approaches of designing inhibitors. This can be done by structure-based in silico screening of small molecules for which binding to the dysfunctional protein restores its wild type thermodynamics. In this review we emphasize the effects of mutations on two important biophysical properties, stability and binding affinity, and how structures can be used for structure-based drug design to mitigate the effects of disease-causing variants on the above biophysical properties.more » « less
-
Cancer results from an evolutionary process that typically yields multiple clones with varying sets of mutations within the same tumor. Accurately modeling this process is key to understanding and predicting cancer evolution. Here, we introduce clone to mutation (CloMu), a flexible and low-parameter tree generative model of cancer evolution. CloMu uses a two-layer neural network trained via reinforcement learning to determine the probability of new mutations based on the existing mutations on a clone. CloMu supports several prediction tasks, including the determination of evolutionary trajectories, tree selection, causality and interchangeability between mutations, and mutation fitness. Importantly, previous methods support only some of these tasks, and many suffer from overfitting on data sets with a large number of mutations. Using simulations, we show that CloMu either matches or outperforms current methods on a wide variety of prediction tasks. In particular, for simulated data with interchangeable mutations, current methods are unable to uncover causal relationships as effectively as CloMu. On breast cancer and leukemia cohorts, we show that CloMu determines similarities and causal relationships between mutations as well as the fitness of mutations. We validate CloMu's inferred mutation fitness values for the leukemia cohort by comparing them to clonal proportion data not used during training, showing high concordance. In summary, CloMu's low-parameter model facilitates a wide range of prediction tasks regarding cancer evolution on increasingly available cohort-level data sets.