Table of Contents: :: Library Catalog

Saved in:

Bibliographic Details
Main Author:	van Tilborg, Derek
Format:	Recurso digital
Language:
Published:	Zenodo 2025
Online Access:	https://doi.org/10.5281/zenodo.16735685
Tags:	Add Tag No Tags, Be the first to tag this record!

Table of Contents:

This repository contains all data and code required to replicate the figures from my PhD thesis. The original, raw data for each chapter can be found in their respective published paper. Thesis Summary: Molecular machine learning has the potential to revolutionize drug discovery by enabling the identification of bioactive molecules from a practically infinite combinatorial space of chemical structures. Yet, these predictive models face fundamental challenges related to generalization that limit their adoption. Unlike other domains in artificial intelligence, such as natural language processing or computer vision, this field operates under extreme data constraints. Molecular structures are discrete, datasets are small, and chemical space is vast. As a result, models often fail to extrapolate beyond the chemical matter they are trained on. Just as the Kármán line marks the boundary between Earth’s atmosphere and outer space – where traditional aerodynamics no longer apply – there exists a "chemical Kármán line", beyond which molecular machine learning models fail to generalize. Operating beyond this boundary requires fundamentally rethinking how models handle the unknown. This thesis addresses this issue through better model evaluation, improved reliability estimation, and data-efficient learning strategies. First, chapters one and two outline the central challenge of molecular machine learning for drug discovery and present the theoretical fundamentals of the methods used in this thesis. Next, chapter three exposes the failure modes of existing methods, particularly deep learning approaches, in capturing complex structure-activity relationships. Stress-testing well-established approaches reveals that predictive performance is driven by how molecular structures are encoded, rather than by the machine learning algorithm itself. Additionally, training set size emerges as a critical limiting factor in training robust models.    Chapters four and five explore active learning as a practical strategy to overcome data scarcity, where models are iteratively updated by acquiring new experimental results. Chapter four establishes a real-world testbed for active learning under extreme low-data conditions, using nanoparticle design as an illustrative and controlled case study. This is extended in chapter five, which systematically evaluates active deep learning strategies for small molecule discovery in low-data regimes, demonstrating that the acquisition strategy – not model architecture – is the primary factor driving hit rate. This chapter also shows how active learning dynamically adapts to emerging structure-activity patterns and compensates for limited molecular diversity in the initial starting data. Chapter six reframes the low-data problem as a broader issue of model reliability. It introduces a joint modeling framework that performs molecular reconstruction alongside property prediction. A model’s ability to reconstruct a molecule is used to quantify its "unfamiliarity" – a new metric reflecting the distance between a molecule and the model’s learned data distribution. This unfamiliarity score proves highly informative for detecting poorly learned molecules and complements uncertainty-based methods for estimating prediction reliability. Prospective validation of unfamiliarity-based molecule screening in the wet lab on two pharmacologically relevant target proteins yielded seven compounds with low micromolar potency. Chapter seven presents an approach for designing nanobodies using a protein language model. Based on a reconstruction-based scoring function, 900 nanobody designs and 100 candidate binders are prioritized for experimental validation, offering a machine learning-guided alternative to traditional mutagenesis during affinity maturation. Finally, chapter eight concludes the findings of this thesis and offers strategies for reliably applying molecular machine learning in prospective drug discovery campaigns. Collectively, the research in this thesis charts a path towards more robust, data-efficient, and generalizable machine learning systems – capable of operating beyond the current boundaries of chemical knowledge. Hopefully, this work contributes to the discovery of better drugs using fewer resources.

Similar Items