The whole is greater than its parts: ensembling improves protein contact prediction

Abstract

The prediction of amino acid contacts from protein sequence is an important problem, as protein contacts are a vital step towards the prediction of folded protein structures. We propose that a powerful concept from deep learning, called ensembling, can increase the accuracy of protein contact predictions by combining the outputs of different neural network models. We show that ensembling the predictions made by different groups at the recent Critical Assessment of Protein Structure Prediction (CASP13) outperforms all individual groups. Further, we show that contacts derived from the distance predictions of three additional deep neural networks—AlphaFold, trRosetta, and ProSPr—can be substantially improved by ensembling all three networks. We also show that ensembling these recent deep neural networks with the best CASP13 group creates a superior contact prediction tool. Finally, we demonstrate that two ensembled networks can successfully differentiate between the folds of two highly homologous sequences. In order to build further on these findings, we propose the creation of a better protein contact benchmark set and additional open-source contact prediction methods.

Stone decision engine accurately predicts stone removal and treatment complications for shock wave lithotripsy and laser ureterorenoscopy patients

“…The diverse models in our study were due to different random splits of the data, randomization of the SMOTE process, and randomization of the initial sets of weights and biases of the ANN models prior to training. Previous studies have used ensemble processes to improve predictions over those made by individually trained models [ 37 , 38 ]. An additional advantage of the ensemble process in our study was that the variability of the predictions for individual patient records could be determined.…”

Training undergraduate research assistants with an outcome-oriented and skill-based mentoring strategy

“…In late 2018, our team began the onboarding phase by attending the CASP13 conference and collecting relevant literature. After about six months, we understood the best methods in the field and designed new approaches based on them (Billings et al, 2019(Billings et al, , 2021Stern et al, 2021). Before the official CASP14 challenge began, the assessment center launched a communitywide experiment to help with the structure determination of COVID-19-related proteins (Kryshtafovych et al, 2021).…”

Improved the Protein Complex Prediction with Protein Language Models

“…When only sequences are given as inputs, PCP is even harder because the unbound structures of individual chains and auxiliary information on the complex interfaces are unavailable. Deep learning has enabled substantial progress in quite a few computational structural biology tasks, such as protein contact 8–10 , tertiary structure prediction 11–13 , and cryo-electron microscopy structure determination 14,15 . Recently, AlphaFold-Multimer 16 has been shown that it outperforms prior protein complex prediction systems, such as the fast Fourier transform-based method ClusPro 17–19 .…”

“…Deep learning has enabled substantial progress in quite a few computational structural biology tasks, such as protein contact [8][9][10] , tertiary structure prediction [11][12][13] , and cryo-electron microscopy structure determination 14,15 . Recently, AlphaFold-Multimer 16 has been shown that it outperforms prior protein complex prediction systems, such as the fast Fourier transform-based method ClusPro [17][18][19] .…”

“…In the last decades, deep learning has enabled substantial progress in quite a few computational structural biology tasks, such as protein contact [13][14][15], tertiary structure prediction [16][17][18], and cryo-electron microscopy structure determination [19,20]. Among these, co-evolution analysis based contact prediction [18,21,22] and structure prediction [23,24] have made substantial progress and demonstrated state-of-the-act accuracy for monomers.…”