Inbound volume of messages, DMs, and emails about personalized cancer vaccines is…uh, very high.

One of the grad students in our lab pointed out there's a $200 credit in the Usage settings so I'm back to the races

There's something a little shameful about reaching the usage limit on a Claude Max subscription...

Data curation at 2x MHCflurry for mono-allelic MS, 4M+ MS observations overall, lots of allele annotations missing from IEDB manually curated from papers.

Total number of MHC seqs curated up to 60k.

“sidecar”

It's such a high stakes bit of information that I've defaulted to typing through HistoGenetics or Red Cross whenever possible

bsky.app/profile/alex...

I'm working towards a synthesis of a lot of the software we've written that can be a one stop / best available way for finding tumor specific T-cell targets on cancers.

...it needs a name though!

Any suggestions?

Going to dissolve the distinction between OpenVax and PIRL by adding support for single cell, long read, structural variants, &c into the neoantigen logic as well as non-mutant targets (CTAs & viruses)

Also curating data for mhcflurry 3 and other new ML models.

Also: new mhctools (added BigMHC and Pepsickle), Topiary (created DSL for sorting/filtering predicted T-cell epitopes), mhcgnomes (curated gene ontology of MHC genes for hundreds of species with flexible parsing logic) and a bunch of other tools in the PIRL/OpenVax universe.

The neoantigen vaccine selection logic at the end of the OpenVax pipeline (code behind Mount Sinai PGV trials) is called Vaxrank

inputs: HLA + VCF + RNA BAM
outputs: vaccine design output

Hadn't changed much since 2010s

Just shipped 2.0 (new predictors, per-trial configs, &c)

Generalized YAML config and CLI redesign by iskandr · Pull Request #215 · openvax/vaxrank Summary This PR starts redesigning Vaxrank's YAML config / CLI surface, but keeps the runtime on a strict unversioned current schema for now. The goal in this slice is to centralize config load...

Request for feedback from anyone using the OpenVax pipeline / its neoantigen vaccine design tool Vaxrank

Vaxrank is getting a more sophisticated config layer (so you can save a YAML per trial without using lots of CLI flags), check out this PR: github.com/openvax/vaxr...

What about other quantities, not just expression?

There's also Processing, Presentation and going to keep adding these as we get richer predictors into mhctools

What about the raw [0,1] scores or percentile rank?

The DSL expressions also have Affinity.rank and Affinity.score

What if you want multiple predictors?

eg TopiaryPredictor(models=[NetMHCpan, MHCflurry], alleles=["HLA-A*02:01", "HLA-B*07:02"])

Can make filtering/ranking expressions in terms of Affinity["netmhcpan"] & Affinity["mhcflurry"]

(from CLI netmhcpan_ba & mhcflurry_ba)

Resurrected Topiary with a pMHC expression DSL eg:

Affinity.logistic(350, 150) - WT(Affinity).logistic(350, 150) - 0.1 * Column("cysteine_count") + 0.1 * Column("aromaticity")

As a str: "ba.logistic(350,150) - wt_ba.logistic(350,150) - 0.1 * cysteine_count + 0.1 * aromaticity"

The thing I feel most about coding agents is tremendous gratitude. Every half-finished project that has haunted me for a decade gets to actually come to life

...but what's been happening today seems extraordinarily bad.

Sudden surprise degradation from experienced software engineer to "I dunno, maybe there's a library called wuzzle with a function called nuzzle? We could try calling it in a loop wrapped in try/catch and see what happens"

I'm surprised by how little this seems to happen for me, but I have been trying to dump more state to gh issues and PRs so information is tracked in a publicly legible way between contexts

I've been working with many Claude Code tabs at once and noticed this morning that at least one of them in particular has gone rock-dumb.

Maybe context gone bad? Maybe lower thinking time behind the scenes?

But very consistently hallucinating APIs, refusing to search, &c

Feeling the AGI

(after like 5+ years of being too stubborn to install any of the horrible windows programs for shotgun MS peptide-spectrum matching, I might get to finally reanalyze raw data with heavy stochastic parrot assistance)

Dog making mRNA vaccine at computer

PuppyVaxx

AI powered mRNA vaccines
made FOR dogs / BY dogs

(raising at $1B valuation)

Def a lot Claude still doesn’t get from the land of wetlab failures

Previously unfathomable depths of obsessive data curation

If you're an MHC sequence, I will find you.

Wetlab immunologist in our lab meeting of computational folks is just absolutely miserable hearing us talk about Claude Code in our updates.

Grad student: "the moment we're experiencing now feels like humans discovering fire"

Immunologist: groans, dies, withers to ashes

It’s unfortunate that industry buries negative results like FixVac, making sense of those trials is the route to therapeutic cancer vaccines which actually work

Any immunologists looking for a 1yr staff scientist job in NC?

I’m not sure if I’m staying at UNC past 2027 but I’m feeling the weight of accumulating experimental backlog.

Ideally already in the area, experienced in T-cell assays like IFNg ELISpot, ICS, MHC multimer flow, &c

a black and white photo of a man with a stethoscope around his neck screaming . ALT: a black and white photo of a man with a stethoscope around his neck screaming .

mhcflurry v2.2.0 is live!

%s/TensorFlow/PyTorch/g/

Thanks to my collaborators Codex and Claude, who accomplished what no human could.

Release 2.2.0: PyTorch migration docs and metadata by iskandr · Pull Request #263 · openvax/mhcflurry Summary Prepare the 2.2.0 stable release -- the first version using PyTorch instead of TensorFlow/Keras. Version bump: 2.2.0rc3 → 2.2.0 README: Replace RC notice with release notes covering the ba...

Alright, here's a PR to officially modernize MHCflurry: github.com/openvax/mhcf...

Keras & TensorFlow are out

PyTorch is in

MPS added as a back-end

Min Python >= 3.10, Pandas >= 2.0

This PR is mostly doc updates, real changes in 2.2.0rc1/2/3 already