Matthew A. Clarke

MAGELLAN: Automated Generation of Interpretable Computational Models for Biological Reasoning Computational models have become essential tools for understanding signalling networks and their non-linear dynamics. However, these models are typically constructed manually using prior knowledge and...

Huge thanks to co-first authors @charliegbarker.bsky.social & Yuxin Sun for their exceptional efforts on this project, to our wonderful collaborators, and especially to the Jasmin Fisher lab at @ucl.ac.uk Cancer Institute!

The pre-print is open access & available at: www.biorxiv.org/content/10.1...

BioModelAnalyzer Logo

This model can then be further edited and interrogated using the BioModelAnalyzer GUI (biomodelanalyzer.org) and command-line tools (github.com/hallba/BioMo...).

Network visualization depicting model fitting and error distribution.

MAGELLAN allows the generation of a computational model from pathway databases, and the fitting of these models to a user-specified specification, resulting in a transparent and interpretable model.

Outline of the workflow and basic fitting of a computational model to the literature derived specification using MAGELLAN

Excited to share our new preprint showcasing a method to infer transparent, mechanistic network models of biological pathways using graph neural networks! www.biorxiv.org/content/10.1...

We then used our published method from @charliegbarker.bsky.social (Barker et al, Genome Research, 2022) to create a reference MASLD network by associating gene co-expression modules with histology scores, identifying key transcription factors regulating them and applying network reconstruction.

Excited to be working on gradient routing as part of Supervised Program for Alignment Research with Alex Cloud, Ariana Azarbal, Cailley Factor & Jorio Cocola. We hope to build safer AI by controlling where and how neural networks learn. See here for more: www.lesswrong.com/posts/tAnHM3...

🚨 New preprint!🚨 Tackling non-small cell lung cancer (NSCLC) adenocarcinoma with in silico models:

🌟 Models tumor pathways (EGFR, AKT, JAK/STAT, WNT)
🌟 personalized drug & radiotherapy combos
🌟 Identifies a signature for patient stratification

Special thanks to my co-first authors @charliegbarker.bsky.social and Ashley Nicholls for their incredible work on this project, and to the Jasmin Fisher lab and all our collaborators!

By predicting the effect of combining targeted therapies with radiotherapy, as well as over 10,000 combinations of drugs and driver mutations, we find promising strategies to improve treatment outcomes, and identify a 19-gene signature that could help predict response to radiotherapy.

Excited to share our new preprint exploring the potential to personalise treatment for non-small cell lung cancer (NSCLC) using in silico modeling: www.biorxiv.org/content/10.1...

Huge thanks to everyone who made the work above possible, especially the @cambiochem.bsky.social , @ucllifesciences.bsky.social, Joseph Bloom, Jasmin Fisher and the Fisher Lab (www.ucl.ac.uk/cancer/resea...)

Applying my experience in biological networks to LLMs, I find a promising method for extracting interpretable features, Sparse AutoEncoders, is more complex than expected. SAE latents are not always independent, forming clusters mapping interpretable subspaces. (www.lesswrong.com/posts/WNoqEi...)

Melanomas often escape immune surveillance. Rowan Howell led work to discover why using network modelling, finding that this rests on a TNF-alpha feedback loop that prevents Langerhans cell migration to the lymph nodes. (www.science.org/doi/full/10....)

The behaviour of myeloproliferative neoplasms depends not only on what mutations they have, but the order in which they acquired them. Laure Talarmain led research to work out why, finding that HOXA9 acts like a switch that remembers which mutation occurred first. (www.nature.com/articles/s41...)

During the pandemic, we rapidly pivoted our methods for finding cancer combination therapy to COVID-19, identifying treatments most likely to be effective for different stages of the disease, as well as remaining useful even with likely resistance mutations. (www.nature.com/articles/s41...)

Professor Jasmin Fisher and I reviewed the exciting progress in using executable computational models to understand cancer complexity. These offer the opportunity to integrate diverse data to predict response and resistance, paving the way for personalized treatment (www.nature.com/articles/s41...)

We investigated cooperation between cancer cells in silico and in vivo, showing how high-Myc expressing cells drive tumour growth but become dependent on WNT-expressing cells, and found combinations of drugs that best disrupted this cooperation. (www.pnas.org/doi/abs/10.1...)

Based on work by Garg et al. (academic.oup.com/bioinformati...), we took advantage of binary decision diagrams to efficiently search future and past possible states of genetic regulatory to understand how past mutations restrict future evolution. (api.repository.cam.ac.uk/server/api/c...)

To get things started on Bluesky, here is a thread of the research I have been part of so far! For details on all of these and more, please see my website: mclarke1991.github.io

I got a bsky so that i could repost this!! Full publication to follow shortly but, would appreciate your feedback on the methods and results! 🔭🧪