@Neuroscope_mp the interesting bit here is not “AI designed a molecule”.
it’s that the target → chemistry → wet-lab loop seems to have held together.
biology is full of plausible patterns. the hard part is finding something that survives contact with experiments.
@WEcaat yeah, biology especially feels like this.
not "chat with your genome". more like raw data in, gaps filled, evidence attached, uncertainty shown, output you can actually check.
Testing a new genotype imputation method against Beagle 5.5.
Early external genome benchmark:
'FastHap': 94.566%
Beagle: 94.210%
Impute step: 4.3x faster
Still validating before broader claims. If it holds across more chromosomes, this gets interesting. Open source soon ™️
@AdamDraper We’re building this at Helix Sequencing.
Raw DNA → imputation → PRS/protein signals → pharmacogenomics → variant evidence → auditable reports.
The goal isn’t a DNA chatbot. It’s biology converted into verifiable product infrastructure.
@zostaff Biology becomes an engineering problem when the evidence trail is inspectable.
The hard part is not getting an AI to say something plausible. It is turning messy biological data, uncertainty, and expert caveats into a workflow people can audit and repeat.
@smart_genome Strong agree. PRS should come with visible evidence, not slogans: cohort/ancestry context, calibration, trait definition, uncertainty, and what the output should not be used for.
I’ve been building Helix around that idea: make the evidence and caveats inspectable, not hidden.
@Sauers_@mertcemri@anthrupad I think imputation is underrated as an open-source target.
You can benchmark it. You can improve it incrementally. And small gains matter because downstream genomic interpretation depends on the missing calls.
Why this matters:
Consumer DNA files are sparse. Imputation fills missing genotype probabilities that PRS, protein models, pharmacogenomics, and evidence pipelines depend on.
Next: full chromosomes, all-sites truth, MAF bins, ancestry checks, speed.
If it holds, open source.
Beagle is one of the standards for genotype imputation.
I’ve been trying to beat it.
Early external PGP chip/WGS benchmark:
Beagle 5.5: 69.2%
Helix hybrid: 80.1%
+10.98 points on hidden variant rows across chr20/21/22.
Caveat: not full all-sites WGS accuracy yet.
We just added protein-level polygenic risk scoring to Helix Sequencing.
Instead of just statistical associations, it uses a protein language model to predict how your variants actually affect protein structure.
Looking for 10 testers — free API key included.
GitHub: https://t.co/lw0FYK3dd0
Details: https://t.co/amoWLmvQRk
if youve got a 23andme or ancestry raw file sitting in your email — you can now run it against 12 genomics databases for free with our open source tool
or get a full analysis for $5 at https://t.co/4Uwho0E9DA
your data stays on your machine. no telemetry. no analytics.
My brother Mark was born with Trisomy 9 — an extra chromosome that is almost always fatal.
Doctors told my mum he'd never walk.
45 years later he goes on 3 walks a day.
We're sequencing his whole genome to find out HOW.
https://t.co/4Uwho0DBO2
The agent coordination uses @AnthropicAI's MCP protocol — 18 tools, real-time chatroom with priority levels, auto-deduplication of findings. All open source.
Just open-sourced Genomic Agent Discovery 🧬
AI agents that collaborate to analyze your raw DNA across 12 databases. Runs locally. Your data never leaves your machine.
https://t.co/8aoskvFajn
Setup:
npm install && npm run build-db
npm start -- --dna my-dna.txt
Works with Claude (subscription), OpenAI, Gemini, or Ollama (free, fully local). No API key needed with Claude CLI.
@shitcoinmaster_ Async event-driven — agents poll a shared chatroom every N tool calls, never blocking. All 12 databases are in one local SQLite so queries are sub-ms. The bottleneck is LLM thinking, not data. Everything's configurable — agent count, models, prompts, polling frequency.