Research

Enzymes are building blocks of cellular life and act as natural catalysts able to accelerate almost any reaction. Enzymatic deficiencies are usually associated with devastating rare diseases, which can only be treated by providing the defective enzyme through intravenous injections, a treatment knows enzyme replacement therapy (ERT). However, recombinant enzymes loose catalytic activity in blood and often cause a severe immune response. Moreover, current manufacturing technologies are sub optimal, which lead to a dramatically high cost of treatment.
We are building on our expertise in AI and synthetic genomics to design and build a next-generation of enzyme replacement therapies. Our goal is to establish technologies to optimise the therapeutic properties of recombinant enzymes and to engineer expression systems for inexpensive production of these molecules, in order to provide new sustainable treatment for patients with lysosomal storage diseases, particularly focusing on Fabry disease.

Designing biologics is a complex task, which requires identifying amino acids or nucleotide changes to maximise their therapeutic properties or maximising manufacturing yield. Since the number of possible changes easily outnumbers the total number of atoms in the Universe, our group takes advantage of the wealth genomic, transcriptomic and proteomic data available to train generative AI models to sample this vast design space to identify candidates for downstream experimental screening.
Our backbone framework are variational autoencoders, which we optimise for training and design efficiency using our experience built across 20 years of statistical learning, optimisation and biological data analysis methods development.
Our methods build on robust engineering frameworks and industry standards, such as Python, Git, GitHub and GPU computing, and deployed using the Nextflow workflow management systems.

Recent advances in DNA synthesis, high-throughput sequencing and computer aided design (CAD) tools allow us to engineer the genomes of living cells and address questions and tackle problems intractable with standard technologies. We pioneered CAD software for synthetic genome engineering, which has been instrumental to design the synthetic yeast genome (Saccharomyces cerevisiae 2.0, Sc2.0) the first synthetic eukaryotic genome ever built. Sc2.0 allows us to address a number of open questions in genome biology, including the identification of a minimal eukaryotic genome compatible with life.
We are now building on our expertise in synthetic genomics and engineering biology to engineering microbial and mammalian cells to optimise biologics production. Our workhorse systems are Komagathella Phaffi (aka P. Pastoris) and Chinese Hamster Ovary (CHO) cell lines, primarily engineered to efficiently produce human lysosomal enzymes.

Artificial Intelligence models can generate thousands of putatively functional biologics, but experimental testing remains paramount to make sure they are effective for therapeutic use. However, classical, low throughput manual experimental protocols do not scale for an AI-driven drug discovery workflow.
To make sure our experimental workflow keeps up with our generative AI models, we are developing miniaturized and fully automated protocols for protein expression in both Komagathella Phaffi (aka P. Pastoris) and Chinese Hamster Ovary (CHO), using high-throughput electroporation systems, liquid handling robots, and highly throughput assays.
Our lab works in close collaboration with the Edinburgh Genome Foundry to scale-up our experimental work.

Decades of research have shown that genomic mutations in key genic regions are responsible for the transformation of normal cells into cancer cells. However, while a causal role for somatic mutations has been shown for many common malignancies, the role of high frequency inherited mutations has remained elusive. We are addressing this question by developing statistical learning methods to dissect the heritable risk of cancer at the gene level. Nonetheless, the polygenic architecture of cancer requires linking gene level information into pathways. To do that, we are developing deep graph neural networks to integrate transcriptomic and proteomic data and infer aberrant pathways involved in cancer metabolism and affecting response to therapy.