ProteoForge is out in Journal of Proteome Research. What changed during review, what is still broken, and where the package is headed.

After months on a feature branch, the new Zola-based website finally merged to main. Custom shortcodes, migrated old posts, 90ms build times, and a dark mode that actually looks good.

A note on making HarmonizePy public, and why we wanted a Python version of HarmonizR's missing-value-aware batch correction.

The story of a small ProteomeXchange metadata tool that started as a fragile Selenium parser and became a PyPI package the lab can actually depend on.

C and C++ cover the classic extension story. Rust through extendr and Savvy brings memory safety to the table. Both are modern, both are performant. Neither fits the R C API boundary as naturally as a language designed to treat C interop as a first-class feature. Early observations from zigr: six backends, 23 tasks, and a case for Zig as the missing option.

Release note for QuEStVar v0.1.0. Python package for paired equivalence and difference testing, now stable on PyPI.

z-toml reached a stable, tested version with TOML v1.1 support, writer support, and passing validation against the toml-test corpus.

A note on why I forked Andrew Kelley's poop into zebrac, what I have added so far, and the direction I want to take it next.

A personal note about a month of sickness, a long hospital stay, and the uncomfortable work of deciding what actually deserves my time.

Early prototype of a domain-specific compression codec for mass spectrometry data. Lossless path beats gzip and comes close to mzMLb. Lossy at q=4096 hits 13.17 MiB from 75.55 MiB mzML. Single DDA dataset. Decode is 27x faster than mzMLb. Early days.

Zero-dependency FASTA indexing that matches samtools byte-for-byte while running up to 17x faster. Built in Zig with SIMD and mmap. Also: a proof of concept that needs Zig 0.14, has no gzip support, and cannot extract sequences yet.

Proteomics data is growing faster than storage budgets. AI demand is driving up hardware prices. The format debate is a distraction. Compression tools exist but nobody uses them. The real problem is that storage is an unaccounted line item, and the field optimizes for generation, not retention.

A note on what I am doing with ZagPlot, a small private Zig plotting project. Not a finished tool, just an excuse to learn graphics, API design, and Zig's systems story from the inside.

Why vendor mass spectrometry formats bother me, why a small native bridge might matter, and why Zig feels like the right place to test the idea.

The search engine space is crowded, fast-moving, and genuinely competitive. DIA-NN, Spectronaut, FragPipe, MaxQuant, and Sage each carved a niche. A Zig-based search engine is not the right next step. The gap between engines might be.

Thermo locks its format behind Windows DLLs. Bruker does better. Spectronaut adds its own proprietary layer on top. Open formats are inevitable, but the real bottleneck is proprietary converters in the middle.

The story behind ProteoForge, a framework for finding differential proteoforms in bottom-up proteomics data. Why protein-level averages hide biology, how missing data broke every tool we tried, and what we found when we applied it to a hypoxia dataset.

Nextflow and Snakemake both get the big things right and the small things wrong. A bioinformatics-specific engine with no runtime and a single static binary might fix the part that actually costs time.

I already knew Rust. I'm learning Zig anyway. Not because it's better, but because it fits a specific kind of work: small, fast tools that live forever in pipelines. An early take, not a manifesto.

Probabilistic dropout analysis for label-free proteomics. Handles missing values by modeling dropout curves instead of imputing.

Deep learning for MS2 spectrum and retention time prediction. Accurate enough for rescoring and library generation.

A primer on fuzzy (soft) clustering, covering the Fuzzy C-Means algorithm with implementations in R and Python, the math behind membership coefficients, and how it differs from hard clustering.