This is the github regarding my work done for my Bachelors Thesis. You can find the theses itself here. You will also find the abstract there as well.
The datasets and the work regarding the datasets is in the BacDive+ and Merck&Co+ folders. You will find a more elaborate README regarding the work concerning them in the folders themselves.
Predicting protein thermostability change upon single mutation is a difficult interdisciplinary machine learning task.
Anyone attempting to succeed at it faces many challenges, including finding enough high-quality data to work with, building models that generalize well beyond the proteins that they were trained on, retaining performance while using more widely available features, and especially validating the models in a meaningful way.
While many machine learning architectures like support vector machines and random forests have been successfully applied to this problem domain, I tried to overcome the challenges and improve model generalization by relying on a state of the art gradient (tree) boosting framework XGBoost in combination with a novel transfer learning approach utilizing deep convolutional neural networks (CNNs), using two datasets, namely BacDive+ (for pre-learning) and Merck&Co+ (for dTm prediction). For a more realistic evaluation of performance, I conducted a leave-one-protein-out cross-validation, besides the common cross-validation.
While I was able to significantly improve performance over the baseline from McGuinness et al. [17] using the features from the Merck&Co+ dataset, models based on the transfer-learning approach alone were not able to compete on the same level. Nonetheless, the latter probably represent the best performing models in literature that do not rely on 3D protein structure information, and can thus be effectively applied to a roughly three orders of magnitude larger set of proteins. Additionally, the potential for further improvement using transfer learning approaches seems great, and is by no means exhausted.
However, the leave-one-protein-out cross-validation also shows, that probably none of these models is suited to reliably assist scientists in the wet lab by suggesting thermostability improving single point mutations, as opposed to very recent hybrid knowledge and machine-learning based models published by Pucci et al. [24] that seem more promising in that regard.