TopologyNet: Topology based deep convolutional and multi-task neural networks for biomolecular property predictions

May combine two claims.

TNet-MMP-2 achieved RP 0.57 and RMSE 1.09 on the M223 data set.

“| TNet-MMP-2d | 0.57 | 1.09 |”

May combine two claims.

On the S2648 set, TNet-MP-2 achieved RP 0.77 and RMSE 0.94.

“| TNet-MP-2 | 350 | 0.81 | 0.94 | 2648 | 0.77 | 0.94 |”

May combine two claims.

On the PDBBind v2016 refined set, TNet-BP achieved median RP 0.81 and RMSE 1.34 pKd/pKi units.

“The median RP and RMSE are 0.81 and 1.34 pKd/pKi units, respectively.”

May combine two claims.

On the PDBBind 2007 core set, TNet-BP achieved a median Pearson correlation coefficient of 0.826 and RMSE of 1.37 pKd/pKi units.

“| TNet-BP | 0.826a | 1.37 |”

• Also stated in: § Deep learning prediction of protein-ligand binding affinities · plain

  “a Median results (The best RP = 0.828 and best RMSE = 1.37 for this method).”

May combine two claims.

On the S350 set, TNet-MP-2 achieved RP 0.81 and RMSE 0.94.

“| TNet-MP-2 | 350 | 0.81 | 0.94 | 2648 | 0.77 | 0.94 |”

The quote only partly states this claim.

TNet-MP-2 outperforms other methods on the S350 and S2648 mutation prediction tasks.

“We demonstrate that TopologyNet outperforms the latest methods in the prediction of protein-ligand binding affinities, mutation induced globular protein folding free energy changes, and mutation induced membrane protein folding free energy changes.”

May combine two claims.

For protein-ligand binding v2007, the handcrafted features and gradient boosting trees method has RP (RMSE) 0.82 (1.40), the TopologyNet method has 0.83 (1.37), and averaging them achieves 0.84 (1.35).

“Our method based on handcrafted features and gradient boosting trees with performance 0.82 (1.40) [61] and the method presented in this work with performance 0.83 (1.37) can achieve improved performance of 0.84 (1.35) when the two results are combined by averaging.”

• cites [61] Z Cang, 2017, Integration of element specific persistent homology and machine learning for protein-ligand binding affinity prediction, International Journal for Numerical Methods in Biomedical EngineeringNot checked yet

  Not checked in the free run.

May combine two claims.

Averaging the results of the two methods for S350 improves performance to 0.83 (0.89), better than both individual methods.

“When the results are averaged for the two methods, the performance is improved to 0.83 (0.89) which is better than both individual methods.”

May combine two claims.

The slope of linear fitting of predicted values to experimental data is 0.66 for the neural network based method and 0.60 for the GBT based method.

“The slope of linear fitting of the predicted values to the experimental data is 0.66 for the neural network based method and 0.60 for the GBT based method which also illustrates that the neural network based method handles border cases better.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Dimension 1 and dimension 2 topological features characterize the hydrophobic network and geometric rings and voids.

“In contrast, dimension 1 and dimension 2 topological features characterize the hydrophobic network and geometric rings and voids.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Mutation data for the large globular protein data set are more reliable, while those for the small membrane protein data set are noisy and less reliable because current membrane protein mutagenesis technologies are immature.

“The mutation data of the large data set for globular proteins are more reliable, while those of the small data set for membrane proteins are noisy and less reliable due to the fact that the current technologies for membrane protein mutagenesis experiments are immature.”

• AssumptionRests on an assumption: Current technologies for membrane protein mutagenesis experiments are immature.

Probably not a substantial claim: may not pass S4 (based).

Neural networks have relatively high underlying randomness from random weight initialization and random mini-batch partition, so they can benefit from bagging even when individual models are trained on the same dataset.

“Specifically, as neural networks have relatively high underlying randomness caused by factors including the random weights initialization and the random mini-batch partition, it can benefit from bagging even if the individual models are trained on the same dataset.”

• AssumptionThe authors state this as an assumption: Models with randomness in the training process likely make different errors on testing data.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

There are strong, moderate and weak hydrogen bond interactions with donor-acceptor distances of 2.2-2.5Å, 2.5-3.2Å, and 3.2-4.0Å, respectively.

“Moreover, Jeffrey suggested that there are strong, moderate and weak hydrogen bond interactions with donor-acceptor distances of 2.2-2.5Å, 2.5-3.2Å, and 3.2-4.0Å, respectively [99].”

• cites [99] GA Jeffrey, 1997, An introduction to hydrogen bondingNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

The proposed method was evaluated using 5-fold cross validation on the S2648 and M223 datasets.

“The proposed method is tested on a data set of 2648 mutation instances of 131 proteins named “S2648” data set [86] in a 5-fold cross validation task over the “S2648” set”

• Also stated in: § Multi-task deep learning prediction of membrane protein mutation impacts · plain

  “The proposed method is tested on a set of 223 mutation instances of membrane proteins covering 7 protein families named “M223” data set [98] with 5-fold cross validation.”

• cites [86] Y Dehouck, 2009, Fast and accurate predictions of protein stability changes upon mutations using statistical potentials and neural networks: PoPMuSiC-2.0, Bioinformatics sourceNot checked yet

  Not checked in the free run.

• cites [98] BM Kroncke, 2016, Documentation of an Imperative To Improve Methods for Predicting Membrane Protein Stability, Biochemistry sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

The task of globular protein mutation predictions helps extract higher level features from low level topological representations.

“Intuitively, the task of globular protein mutation predictions help to extract higher level features from low level topological representations.”

• AssumptionRests on an assumption: The popular multi-task learning framework built on linear regression with regularization techniques lacks the ability to extract the relationship between very low level descriptors and the target quantity.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

0-dimensional TFs describe intramolecular interactions between the protein and ligand.

“Here, 0-dimensional TFs describe intramolecular interactions between the protein and ligand.”

May combine two claims.

Two sets of features, dimension 0 and dimensions 1 and 2, both contribute to predictions.

“The combination of all dimension features results in better RP (RMSE) of 0.74 (1.08) and 0.83 (1.37), respectively for the mutation and binding predictions, showing that two sets of features both contribute to predictions.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

The flexible and hierarchical structure of neural network allows seamless combination of automatically extracted features and handcrafted features.

“The flexible and hierarchical structure of neural network allows seamless combination of automatically extracted features and handcrafted features.”

May combine two claims.

Properties that reside in different distance scales are heterogeneous, so unlike images or videos, there is no obvious transferable property of convolution filters along the convolution dimension in the proposed method.

“Properties that reside in different distance scales are heterogeneous so unlike images or videos, there is no obvious transferable property of the convolution filters along the convolution dimension in the proposed method.”

Probably not a substantial claim: may not pass S4 (based).

Traditional machine learning approaches require manual extraction of features for each domain of application.

“Traditional machine learning approaches require manual extraction of features for each domain of application.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The dropout technique is an easy way of preventing network overfitting.

“The dropout technique is an easy way of preventing network overfitting [100].”

• cites [100] N Srivastava, 2014, Dropout: a simple way to prevent neural networks from overfitting, Journal of Machine Learning ResearchNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

All heavy atom TFs delineate the geometric effect of protein-ligand binding.

“All heavy atom TFs delineate the geometric effect of protein-ligand binding.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

ESTFs between nitrogen and oxygen atoms correlate to hydrophilic interactions and/or hydrogen bonds in biomolecules.

“Similarly ESTFs between nitrogen and oxygen atoms correlate to hydrophilic interactions and/or hydrogen bonds in biomolcules.”

Probably not a substantial claim: may not pass S4 (based).

Convolutional neural network facilitates automatic feature extraction from multichannel topological invariant inputs.

“Convolutional neural network facilitates the automatic feature extraction from multichannel topological invariant inputs.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Element-specific persistent homology reduces geometric and biological complexities and provides a sufficient and structured low level representation for neural networks.

“In this framework, element-specific persistent homology reduces geometric and biological complexities and provides a sufficient and structured low level representation for neural networks.”

Probably not a substantial claim: may not pass S4 (based).

The all heavy atom topological representation of proteins does not provide enough biological information about protein structures, such as bond length distribution of a given type of atoms, hydrogen bonds, hydrophobic and hydrophilic effects.

“The all heavy atom topological representation of proteins does not provide enough biological information about protein structures, such as bond length distribution of a given type of atoms, hydrogen bonds, hydrophobic and hydrophilic effects, etc.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Knowledge based scoring functions outperform other scoring functions.

“Given the current massive and complex data challenges, knowledge based scoring functions outperform other scoring functions. [65].”

• cites [65] Z Zheng, 2011, Ligand Identification Scoring Algorithm (LISA), Journal of Chemical Information and Model sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

When the convolutional neural network is applied, convolution layers identify local patterns of atomic interactions and fully connected layers extract higher level descriptions by combining local patterns at various distance scales.

“When the convolutional neural network is applied, the convolution layers identify local patterns of atomic interactions and the fully connected layers then extract higher level descriptions of the system by combining local patterns at various distance scales.”

Probably not a substantial claim: may not pass S4 (based).

It is important to keep track of the birth, death, and persistence patterns of the barcodes because this information is associated with the bond length, ring or cavity size, flexibility and steric effect.

“Additionally, as shown in Fig 2, it is important to keep track of the birth, death, and persistence patterns of the barcodes, because this information is associated with the bond length, ring or cavity size, flexibility and steric effect.”

Probably not a substantial claim: may not pass S4 (based).

Tasks with insufficient training data sets are more likely to benefit from information collected from tasks with large training sets in a multi-task learning framework.

“Finally, tasks with insufficient training data sets will be more likely to benefit from the information collected from tasks with large training sets in a multi-task learning framework.”

Probably not a substantial claim: may not pass S4 (based).

It is not valid to do bagging of predictors on different cross validation splits.

“Bagging of only 20 models is performed here because it is not valid to do bagging of predictors on different cross validation splits.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Vr reflects the 0th order topological connectivity information.

“Generally speaking, Vr also reflects the 0th order topological connectivity information.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Unfolded states of proteins are highly dynamic, which significantly increases modeling cost due to the need of sampling over large conformation space.

“Since unfolded states of proteins are highly dynamic which significantly increases the modeling cost due to the need of sampling over large conformation space, we only analyze the folded states of the mutants and the wild type proteins in this application.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Vr is not equivalent to Vd in most simplicial complex setups.

“It should be noted that Vr is not equivalent to Vd in most simplicial complex setups.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The flexible and hierarchical structure of neural network makes it easy to implement multi-task learning by combining related tasks to a desired level of model sharing by tuning the layer of model branching.

“It also makes it easy to implement multi-task learning by combining related tasks to a desired level of model sharing by tuning the layer of model branching.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Betti-1 and Betti-2 ESTFs from carbon atoms are associated with hydrophobic interaction networks in biomolecules.

“For example, Betti-1 and Betti-2 ESTFs from carbon atoms are associated with hydrophobic interaction networks in biomolecules.”

Probably not a substantial claim: may not pass S4 (based).

Common topological attributes and underlying physical interactions in features provide a basis for multi-task predictions.

“The common topological attributes and underlying physical interactions in features provide a basis for multi-task predictions.”

Probably not a substantial claim: may not pass S4 (based).

Neural networks with several convolution layers and fully connected layers possess a large number of degrees of freedom which can easily lead to overfitting.

“Neural networks with several convolution layers and fully connected layers possess a large number of degrees of freedom which can easily lead to overfitting.”

May combine two claims.

Methods of ensembles of trees and neural network based methods may result in different error characteristics for different samples and can potentially improve predictive power by correcting each other's error when averaged.

“Therefore, similar to the idea of bagging, methods of ensembles of trees and neural network based methods may result in different error characteristics for different samples and can potentially improve the predictive power by correcting each others’ error when the results from different models are averaged.”

• AssumptionRests on an assumption: In many applications, where among the dataset there are just a handful of samples with large absolute value for the target property, methods of ensembles of trees tend to overestimate (underestimate) the border cases with very negative (positive) target values.
• AssumptionRests on an assumption: Data samples are generally unevenly distributed.

May combine two claims.

The proposed TopologyNet methods have been shown to outperform other existing methods in protein-ligand binding affinity predictions and mutation induced protein stability change predictions.

“The proposed topology based neural network (TopologyNet) methods have been shown to outperform other existing methods in protein-ligand binding affinity predictions and mutation induced protein stability change predictions.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

TFs of carbon atoms account for hydrophobic effects and also implicitly reflect the solvation effects.

“The TFs of carbon atoms account for hydrophobic effects and also implicitly reflect the solvation effects.”

Probably not a substantial claim: may not pass S4 (based).

Dimension 0 inputs describe pairwise atomic interactions.

“Intuitively, the dimension 0 inputs describe pairwise atomic interactions, which clearly contribute to the prediction of the target properties.”

May combine two claims.

Convolutional neural networks can learn from data to extract high level representations of biomolecular systems while retaining spatial relationships, and construct mappings from these representations to target properties.

“Given this representation, convolutional neural networks can then learn from data to extract high level representations of the biomolecular systems, while retaining the spatial relationships, and construct mappings from these representations to the target properties.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

One advantage of multichannel topological convolutional deep neural networks is their ability to extract features hierarchically from low level topological representations.

“One of the advantages of multichannel topological convolutional deep neural networks is their ability to extract features hierarchically from low level topological representations.”