AllResearchEngineering
Featured

Notes from building an AI co-scientist.

Research, method and engineering writing from the team building the co-scientist.

Browse archive →
ResearchFeatured

A random forest beat two frontier co-folders. That is the benchmark talking.

Affinity models report one held-out correlation near 0.6. Split that number by how many times a protein family already appears in the PDB and it comes apart. MIRAGE measures the split, and finds part of the deficit is a missing input rather than a hard limit.

Mehdi Yazdani-Jahromi · Sep 6, 2026 / 5 min read

Engineering posts

View all →
Jul 24, 2026What we keep in the knowledge graph, including the things that did not workMost biomedical graphs remember hits. Ours keeps curated negatives alongside them, types every edge by how it was established, and refuses to settle a disagreement by deleting one side of it.Engineering
2 min read
News

Releases, milestones and announcements.

5 items
ProductSep 11, 2026
Early access to Helixir opens to research teams
Research and biotech teams can apply for early access to the Co-Scientist and run hypothesis discovery, de novo design and literature review on their own targets.
ResearchSep 6, 2026
MIRAGE: a benchmark for protein-family generalization in affinity prediction
Our preprint treats historical public protein-family support as an explicit experimental variable, for affinity and for pose, and reports what happens to frontier co-folders on families they have not seen.
ResearchSep 1, 2026
OIL: an open codon-resolved nucleotide framework, presented at MoML 2026
The Omics Integrated Library assembles 2.35 trillion nucleotides across fifteen sources, codon-resolved and with provenance on every record. Scripts and the automated procedure follow later this year.
ResearchDec 1, 2025
First equivariant foundation model for mRNA design at NeurIPS 2025
Our foundation model learns biology in a geometry-aware representation: the technical core the co-scientist platform is built on.
ResearchAug 21, 2025
Equi-mRNA: the first codon-level equivariant mRNA language model
Our paper encodes synonymous-codon symmetry directly into the representation as SO(2) cyclic subgroups, rather than leaving a language model to infer it from data.

Most read

01Codon resolution at metagenomic scale: why we built OILResearch · 5 min02Codon-level equivariance: what it fixes, and what it does notResearch · 2 min03Introducing Equi-mRNA: the first codon-level equivariant mRNA language modelResearch · 8 min04What we keep in the knowledge graph, including the things that did not workEngineering · 2 min

Topics

Research4 posts

Foundation models, geometry-aware representations, and the science behind the co-scientist.

Latest: Sep 6, 2026
Engineering1 post

Agent orchestration, pipeline design, and the systems that keep reasoning auditable.

Latest: Jul 24, 2026

Deep dives

View all →
Research

Codon resolution at metagenomic scale: why we built OIL

Codon corpora stop at reference organisms. Metagenomic corpora throw the codons away by storing coding sequence as protein. OIL is 2.35 trillion nucleotides that keep both, with provenance on every record and nothing back-translated.

Sep 1, 2026 / 5 min read
Research

Codon-level equivariance: what it fixes, and what it does not

The genetic code is degenerate, and most mRNA models learn that the expensive way. Equi-mRNA encodes it in the representation instead. The mechanism, the numbers, and the thing people get wrong about synonymous codons.

Aug 14, 2026 / 2 min read
Research

Introducing Equi-mRNA: the first codon-level equivariant mRNA language model

Most sequence models treat the genetic code as text and relearn its geometry from scratch. Equi-mRNA builds that geometry in: a 15M-parameter model that outperforms baselines 3-5x its size, peer-reviewed at NeurIPS 2025.

Aug 4, 2026 / 8 min read
Newsletter

Research notes, straight to your inbox.

One email every few weeks: new posts, model releases, and what we are reading. No spam, unsubscribe anytime.