Neural networks predict protein structure and function

Functional and genomic analyses of alpha-solenoid proteins

Alpha-solenoids are flexible protein structural domains formed by ensembles of alpha-helical repeats (Armadillo and HEAT repeats among others). While homology can be used to detect many of these repeats, some alpha-solenoids have very little sequence homology to proteins of known structure and we expect that many remain undetected. We previously developed a method for detection of alpha-helical repeats based on a neural network trained on a dataset of protein structures. Here we improved the detection algorithm and updated the training dataset using recently solved structures of alpha-solenoids. Unexpectedly, we identified occurrences of alpha-solenoids in solved protein structures that escaped attention, for example within the core of the catalytic subunit of PI3KC. Our results expand the current set of known alpha-solenoids. Application of our tool to the protein universe allowed us to detect their significant enrichment in proteins interacting with many proteins, confirming that alpha-solenoids are generally involved in protein-protein interactions. We then studied the taxonomic distribution of alpha-solenoids to discuss an evolutionary scenario for the emergence of this type of domain, speculating that alpha-solenoids have emerged in multiple taxa in independent events by convergent evolution. We observe a higher rate of alpha-solenoids in eukaryotic genomes and in some prokaryotic families, such as Cyanobacteria and Planctomycetes, which could be associated to increased cellular complexity. The method is available at http://cbdm.mdc-berlin.de/~ard2/.

Machine learning on normalized protein sequences

BACKGROUND

Machine learning techniques have been widely applied to biological sequences, e.g. to predict drug resistance in HIV-1 from sequences of drug target proteins and protein functional classes. As deletions and insertions are frequent in biological sequences, a major limitation of current methods is the inability to handle varying sequence lengths.

FINDINGS

We propose to normalize sequences to uniform length. To this end, we tested one linear and four different non-linear interpolation methods for the normalization of sequence lengths of 19 classification datasets. Classification tasks included prediction of HIV-1 drug resistance from drug target sequences and sequence-based prediction of protein function. We applied random forests to the classification of sequences into "positive" and "negative" samples. Statistical tests showed that the linear interpolation outperforms the non-linear interpolation methods in most of the analyzed datasets, while in a few cases non-linear methods had a small but significant advantage. Compared to other published methods, our prediction scheme leads to an improvement in prediction accuracy by up to 14%.

CONCLUSIONS

We found that machine learning on sequences normalized by simple linear interpolation gave better or at least competitive results compared to state-of-the-art procedures, and thus, is a promising alternative to existing methods, especially for protein sequences of variable length.

Predicting Bevirimat resistance of HIV-1 from genotype

BACKGROUND

Maturation inhibitors are a new class of antiretroviral drugs. Bevirimat (BVM) was the first substance in this class of inhibitors entering clinical trials. While the inhibitory function of BVM is well established, the molecular mechanisms of action and resistance are not well understood. It is known that mutations in the regions CS p24/p2 and p2 can cause phenotypic resistance to BVM. We have investigated a set of p24/p2 sequences of HIV-1 of known phenotypic resistance to BVM to test whether BVM resistance can be predicted from sequence, and to identify possible molecular mechanisms of BVM resistance in HIV-1.

RESULTS

We used artificial neural networks and random forests with different descriptors for the prediction of BVM resistance. Random forests with hydrophobicity as descriptor performed best and classified the sequences with an area under the Receiver Operating Characteristics (ROC) curve of 0.93 +/- 0.001. For the collected data we find that p2 sequence positions 369 to 376 have the highest impact on resistance, with positions 370 and 372 being particularly important. These findings are in partial agreement with other recent studies. Apart from the complex machine learning models we derived a number of simple rules that predict BVM resistance from sequence with surprising accuracy. According to computational predictions based on the data set used, cleavage sites are usually not shifted by resistance mutations. However, we found that resistance mutations could shorten and weaken the alpha-helix in p2, which hints at a possible resistance mechanism.

CONCLUSIONS

We found that BVM resistance of HIV-1 can be predicted well from the sequence of the p2 peptide, which may prove useful for personalized therapy if maturation inhibitors reach clinical practice. Results of secondary structure analysis are compatible with a possible route to BVM resistance in which mutations weaken a six-helix bundle discovered in recent experiments, and thus ease Gag cleavage by the retroviral protease.