Despite the fundamental role of experimental protein structure determination, computational methods are of essential importance to bridge the ever growing gap between available protein sequence and structure data. Common structure prediction methods rely on experimental data, which is not available for about half of the known protein families.

Recent advancements in amino acid contact prediction have revolutionized the field of protein structure prediction. Contacts can be used to guide template-free structure predictions that do not rely on experimentally solved structures of homologous proteins. Such methods are now able to produce accurate models for a wide range of protein families.

We developed PconsC2, an approach that improved existing contact prediction methods by recognizing intra-molecular contact patterns and noise reduction. An inherent problem of contact prediction based on maximum entropy models is that large alignments with over 1000 effective sequences are needed to infer contacts accurately. These are however not available for more than 80% of all protein families that do not have a representative structure in PDB. With PconsC3, we could extend the applicability of contact prediction to families as small as 100 effective sequences by combining global inference methods with machine learning based on local pairwise measures.

By introducing PconsFold, a pipeline for contact-based structure prediction, we could show that improvements in contact prediction accuracy translate to more accurate models. Finally, we applied a similar technique to Pfam, a comprehensive database of known protein families. In addition to using a faster folding protocol we employed model quality assessment methods, crucial for estimating the confidence in the accuracy of predicted models. We propose models tobe accurate for 558 families that do not have a representative known structure. Out of those, over 75% have not been reported before.

Proteins are the basic molecular machines of the cell, performing a broad range of tasks, from structural support to catalysisof chemical reactions. Their function is determined by their 3D structure, which in turn is dictated by the order of their components, the amino acids.

This thesis is dedicated to applications of machine learning to the problems of contact prediction, ab-initio, and model quality assessment. In particular, my research has been focused on developing methods that are both effective, and easy to use.

In the first paper, we improved the already state-of-the-art model quality assessment (MQA) program ProQ3 replacing the underlying machine learning algorithm from svm to Deep Learning, baptised ProQ3D. The correlation between predicted and true scores was improved from 0.85 to 0.90, using the same training data and features.

The second paper joined several programs into a single pipeline for ab-initio structure prediction: contact prediction,folding, and model selection. We attempted to predict the structures of all 6379 PFAM families with unknown structure, ofwhich 558 we believe to be accurate. Of these, 415 had not been reported before.

The third paper uses advances in machine learning to build a contact predictor, PconsC4, that is fast and easy to deployin large-scale studies, since it requires a single Multiple Sequence Alignment (MSA), and no external dependencies. The predictions are state-of-the-art, yielding a 12% improvement in precision over PconsC3, and 244 times faster.

With ProQ4, in the fourth paper, we introduce a novel way of training deep networks for MQA in a way that minimises the bias of the training data, and emphasises model ranking, and demonstrate its viability with a minimal description ofthe protein. The ranking correlation was improved with respect to ProQ3D from 0.82 to 0.90.

Lastly, in the fifth paper, weshow the results of ProQ3D and ProQ4 in a completely blind test: CASP13.