This thesis contains three main sections. In the first section, we examine methodologies to discriminate Type II protein kinase inhibitors from the Type I inhibitors. We have studied the structure of 55 Type II kinase inhibitors and have notice specific descriptive geometric features. Using this information we have developed a pharmacophore and a shape based screening approach. We have found that these methods did not effectively discriminate between the two inhibitor types used independently, but when combined in a consecutive way – pharmacophore search first, then shape based screening, we have found a method that successfully filtered out all Type I molecules. The effect of protonation states and using different conformer generators were studied as well. This method was then tested on a freely available database of decoy molecules and again shown to be discriminative. In the second section of the thesis, we implement and assess swarm-based docking methods. We implement a repulsive particle swarm optimization (RPSO) based conformational search approach into Autodock 3.05. The performance of this approach with different parameters was then tested on a set of 51 protein ligand complexes. The effect of using different factoring for the cognitive, social and repulsive terms and the importance of the inertia weight were explored. We found that the RPSO method gives similar performance to the particle swarm optimization method. Compared to the genetic algorithm approach used in Autodock 3.05, our RPSO method gives better results in terms of finding lower energy conformations. In the final, third section we have implemented a Monte Carlo (MC) based conformer searching approach into Gaussian03. This enables high level quantum mechanics/molecular mechanics (QM/MM) potentials to be used in docking molecules in a protein active site. This program was tested on two Zn2+ ion-containing complexes, carbonic anhydrase II and cytidine deaminase. The effects of different QM region definitions were explored in both systems. A consecutive and a parallel docking approach were used to study the volume of the active site explored by the MC search algorithm. In case of the carbonic anhydrase II complex, we have used 1,2-difluorobenzene as a ligand to explore the favourable interactions within the binding site. With the cytidine deaminase complex, we have evaluated the ability of the approach to discriminate the native pose from other higher energy conformations during the exploration of the active site of the protein. We find from our initial calculations, that our program is able to perform a conformational search in both cases, and the effect of QM region definition is noticeable, especially in the description of the hydrophobic interactions within the carbonic anhydrase II system. Our approach is also able to find poses of the cytidine deaminase ligand within 1 Å of the native pose.

In the field of mechanical engineering, particularly in managing 3D CAD model databases, there's a pressing challenge in efficiently identifying and reusing duplicate models instead of the costly process of creating new ones from scratch. Shape-based similarity retrieval stands out as a crucial solution for industry players like Epiroc, offering a strategy for duplicate detection and overcoming the constraints of text-based searches. In this thesis, we explore an innovative Machine Learning-based framework for shape retrieval, leveraging UV-Net to encode 3D models for effective shape description and incorporating Self-Supervised Learning to improve the identification of model similarities without relying heavily on labeled data. To assess the effectiveness of our approach and understand its performance, we tested it using three foundational ML models for a quantitative evaluation of the shape embeddings. The findings highlight the framework's ability to accurately identify similar models within extensive datasets. Moreover, we applied this methodology to retrieve similar models of ’Epirocs product 3D models within ’Epirocs Standard Part Dataset, demonstrating the successful integration of ML in mechanical design for efficient shape retrieval and presenting a viable industrial application for Epiroc. Despite its achievements, the study acknowledges limitations in system integration and computational efficiency, suggesting future research directions such as improving access to various file formats, enhancing the user interface, and optimizing computational resources. / Inom området maskinteknik, särskilt när det gäller hantering av 3D CAD-modelldatabaser, finns det en akut utmaning i att effektivt identifiera och återanvända dubbletter av modeller istället för den kostsamma processen att skapa nya från grunden. Formbaserad likhetshämtning framstår som en avgörande lösning för branschaktörer som Epiroc, och erbjuder en strategi för dubblettdetektering och att övervinna begränsningarna med textbaserade sökningar. I det här examensarbetet utforskar vi ett innovativt Maskininlärning-baserat ramverk för formåtervinning, utnyttjar UV-Net för att koda 3D-modeller för effektiv formbeskrivning och införlivar Självledd Inlärning för att förbättra identifieringen av modelllikheter utan att förlita oss mycket på märkta data. För att bedöma effektiviteten av vårt tillvägagångssätt och förstå dess prestanda, testade vi det med tre grundläggande ML-modeller för en kvantitativ utvärdering av forminbäddningarna. Resultaten belyser ramverkets förmåga att exakt identifiera liknande modeller inom omfattande datauppsättningar. Dessutom tillämpade vi denna metod för att hämta liknande modeller av Epirocs produkt-3D-modeller inom Epirocs standarddeldataset, vilket visar den framgångsrika integrationen av ML i mekanisk design för effektiv formhämtning och presenterar en hållbar industriell applikation för Epiroc. Trots sina framgångar erkänner studien begränsningar i systemintegration och beräkningseffektivitet, och föreslår framtida forskningsriktningar som att förbättra tillgången till olika filformat, förbättra användargränssnittet och optimera beräkningsresurser.