Jekyll2026-09-19T19:20:46+00:00https://abidikhairi.github.io/feed.xmlKhairi Abidi | LLM / AI Engineer & NLP for Bioinformatics ResearcherAI/LLM engineer with experience shipping production ML and data systems, leading engineering teams, and researching protein language models for bioinformatics.Khairi AbidiKCNG Explorer: Probing a Protein Language Model on the ‘Silent’ Kv6 Channels2026-09-19T00:00:00+00:002026-09-19T00:00:00+00:00https://abidikhairi.github.io/blog/kcng-explorerI just shipped KCNG Explorer, a small research tool for poking at the KCNG subfamily of voltage-gated potassium channel subunits — KCNG1 through KCNG4, better known as Kv6.1–Kv6.4. They’re an odd corner of the Kv family: on their own they don’t form a functional channel at all. They’re “silent” subunits that only work as heterotetramers with Kv2, modulating its gating rather than conducting current by themselves. That makes them a nice, self-contained test case for asking what a protein language model actually understands about a sequence.

What it does

KCNG Explorer wraps a pretrained ESM Cambrian (ESMC) masked language model around a focused workflow:

  • Load a sequence — fetch directly by UniProt accession, or paste your own FASTA.
  • Score every point mutation — for the loaded sequence, the app computes a full log-likelihood-ratio (LLR) matrix over all 19 alternative residues at every position, so you get the model’s implied fitness landscape for the whole protein in one pass.
  • Cross-check against real variants — those LLR scores are joined against UniProt’s annotated natural variants (skipping any where the wild-type residue doesn’t match the given sequence, e.g. a mismatched isoform), so you can see whether the model’s notion of “tolerated” vs. “disruptive” lines up with what’s actually observed in nature.
  • Look at the structure — the same protein is rendered in 3D, preferring an experimental PDB structure via UniProt’s cross-references and falling back to an AlphaFold DB prediction (fetched live from AlphaFold’s API, since model versions get republished over time) when no experimental structure exists.

Because the mutation matrix is 19 × sequence length model calls, it’s cached per-session and only recomputed when the active sequence actually changes — everything else (variant table, heatmap, structure viewer) just re-renders off whatever’s already in state.

From Streamlit prototype to a Gradio Space

The app started life as a Streamlit prototype and got ported to Gradio so it could be deployed as a Hugging Face Space with GPU access (@spaces.GPU, ZeroGPU-backed). Both UIs still live side by side in the repo, following the same state-resolution pattern — a sidebar action resolves to (sequence, accession, matrix, matrix_sequence), and a single render function turns that into every panel. Keeping resolution and rendering separate is what let the migration be mostly a matter of swapping st.session_state for gr.State rather than rewriting the app logic.

The less glamorous part of shipping it was the dependency graph. esm >= 3.4.0 is a hard floor for the ESMC wrapper API this app needs, which drags in transformers < 5.0, which caps huggingface-hub < 1.0 — and Gradio raised its own huggingface-hub floor above that starting at 6.18.0. So the Space is deliberately pinned to gradio==6.17.3, the newest release where that three-way constraint is still satisfiable at all. Fun way to relearn that “just bump the version” isn’t always an option.

Try it

The Space is live at huggingface.co/spaces/khairi/KCNG-Explorer — load any KCNG1–KCNG4 accession and see where the model agrees (or disagrees) with nature on what this “silent” channel family can tolerate.

]]>Khairi AbidiWelcome to My Blog2026-09-19T00:00:00+00:002026-09-19T00:00:00+00:00https://abidikhairi.github.io/blog/welcome-to-my-blogThis is the first post on my blog. To add a new one, drop a Markdown file into _posts/ named YYYY-MM-DD-title.md with front matter like this one - Jekyll picks it up automatically and it shows up on the Blog page.

]]>Khairi Abidi