iiinova — where your digital self keeps evolving

Lucas van Dijk

@lucasvandijk.nl

206 followers 386 following 67 posts

Bioinformatics Software Engineer @ @pacbio.bsky.social 🧬🖥️. Previously at the data sciences platform and bacterial genomics group @ @broadinstitute.org. Interested in algorithm development, human genetics, and the microbiome.

Posts Replies Media

Lucas van Dijk @lucasvandijk.nl · 09/02/2025

For those using WDL for their bioinformatic workflows: I wrote an extension for the @zed.dev code editor enabling syntax highlighting and integration with wdl-lsp. It also supports bash highlighting in command blocks. Installable through the Zed extensions pane! github.com/broadinstitu...

Screenshot of the Zed editor showcasing syntax highlighting for the Workflow Description Language.

150

Lucas van Dijk @lucasvandijk.nl · 27/01/2025

A printed thesis! Did not expect it would feel so good to have the final version physically in your hands 🙌

120

Lucas van Dijk @lucasvandijk.nl · 19/11/2024

Found ChatGPT in Fallout New Vegas: "Are you some kind of AI?" "Regrettably not, sir. All modules in this habitat are synthetic personalities atop a mundane operating system. There is no intelligence here, sir." www.youtube.com/watch?v=4XJw...

A view of the Central Intelligence Unit in Fallout New Vegas: Old World Blues.

110

Lucas van Dijk @lucasvandijk.nl · 12/11/2024

I think Philip Eng deserves a statue in the Boston common

MBTA subway map showing only a handful of remaining slow zones

032

Lucas van Dijk @lucasvandijk.nl · 26/10/2024

Working through Prince's excellent "Understanding Deep Learning" book. Full of great insights like these: by optimizing using SGD we implicitly regularize the weights towards values where the gradients are similar across batches 🤯

The equation shows the updated loss function for SGD with the regularization term.

010

Lucas van Dijk @lucasvandijk.nl · 28/05/2024

For those interested in the technical details, POASTA achieves its performance through three techniques: 1) its minimum remaining gap heuristic, 2) depth-first greedy alignment of matches and 3) superbubble topology-informed pruning.

110

Lucas van Dijk @lucasvandijk.nl · 28/05/2024

In low-diversity settings, POASTA's memory-scaling characteristics also enable alignment of much longer sequences, and we constructed megabase-length alignments of 342 M. tuberculosis sequences. This would require terabytes of memory with SPOA.

110