
https://www.youtube.com/watch?v=UvAccQQgqMg
TL;DR: We're building a world model for biology and the tools that both human and agentic scientists need to use it: biological data, SOTA sequence and structure search, clustering, and alignment, optimized AI4Science model inference, and a shared discovery workbench. Today we're launching Umi and Sora, the Orikata web toolkit, and hosted model inference with Noto, alongside the open-source release of Hikoboshi and Orihime. The full Noto discovery workbench follows in mid-November.
For Umi and Sora access, email info@orikata.ai. The public search demo currently covers TEDBench.
Hey everyone! I'm Casey, founder of Orikata Bio PBC, the company behind Orikata Bio and Megure Labs. At Orikata, we're building Ibuki, a world model for biology that learns from evolution itself to help scientists understand and engineer biology. Our first focus is mastering the art of folding proteins.
At Megure, we're building the scientific software to train and run those models, and Noto, a shared multiplayer discovery workbench for human and agentic scientists.
I'm a former MD/PhD student at UPenn. As an undergraduate at Berkeley, I was jointly advised by Matt Francis and Jennifer Doudna, one of the 2020 Nobel Prize in Chemistry laureates; that work led to a first-author JACS paper, two additional publications, and a co-invented patent underlying Catena Bio (YC S21). I later pursued a fully supported MPhil at the University of Cambridge, where I initiated a collaboration between my Cambridge supervisors and David Baker, 2024 Nobel Laureate also in Chemistry. That collaboration later became part of Archon Bio. I'm also first author of an ICML 2026 paper, and a coauthor of a 2026 Nature Biotechnology paper on AI-guided in-silico evolution. My work has taken me from wet-lab biochemistry to machine learning and the systems needed to run it efficiently.
Building those models meant solving many problems underneath them first.
We needed data to learn from, tools to find, organize, and compare the relevant biology, and software that could run the whole thing efficiently. And scientists, both agentic and human, need somewhere to work with those results. These are the pieces we're bringing together across Orikata Bio and Megure Labs.
What has evolution already tried? Evolution has spent billions of years exploring ways to build living systems. The sequences, structures, and other forms of biological data we collect are a record of that exploration. If we can learn how they relate, we can use that context to understand what exists and design what comes next.
We started by assembling Umi, with over 100 billion unique protein sequences, and Sora, with nearly 2 billion predicted and experimental protein structures. Umi brings together more than 340 billion raw sequence records before deduplication. That's larger than the datasets used to train leading protein models, and on par with today's largest assembled sequence collections.
We built these collections by bringing together biological data scattered across public repositories, legacy tape archives and old servers. Recovering the files was only the beginning. We converted, normalized, deduplicated, and stored those records in an efficient format so many scientists and agents can work with the same collection at once.
Having all that data raises the next question: how do you search through it?
The tools scientists use to discover and compare genes and proteins were built for much smaller databases. Search, clustering, and alignment supply the biological context for structure-prediction models, generative-design models, and, increasingly, agentic discovery pipelines. Those steps matter during training and inference, and they are becoming the bottleneck.
Looking at how those tools compare proteins, we realized that the fixed substitution matrices used by classical sequence-alignment tools are secretly just the simplest possible example of a neural network. They assign a fixed score to each pair of amino acids, scoring how similar the two amino acid tokens are, wherever those amino acids appear in a sequence.
What if we replaced that fixed matrix with a contextual substitution operator: a neural network that learns how to compare positions using their sequence, structure, or both?
Classical alignment can also be understood as structured cross-attention. It finds and highlights corresponding positions across related sequences. Across many proteins, those correspondences form multiple sequence alignments, or MSAs, the evolutionary context used by AlphaFold and other similar biological ML models.
That gives us a different learning objective: predict which positions correspond. We can train on known alignments, learn directly from structure through geometric self-supervision, or combine the two depending on what's available. We can also learn from alignment examples from proteins that have no structures.
To train through those calculations, we built Orihime, our GPU-backed structured-attention and differentiable-alignment library at Megure Labs. Orihime lets us train the models through the alignment calculation itself, with CPU, CUDA, and HIP implementations. The research behind it is described in our first-author ICML 2026 paper.
At Orikata, we used those learned comparisons to build Hikoboshi, our neural-network bioinformatics suite. We're releasing tools for:
The alignment tools accept sequences, structures, or both, and run locally on CPU. With sequence input, our models recover structural relationships without requiring an input structure. With structure input, our alignments reach TMalign-level full-protein geometric quality.
LDDT measures structural agreement, and F1 measures how well we recover corresponding positions. Both range from 0 to 1, with 1 the maximum.
On the SABmark twilight benchmark, sequence-only Hikoboshi delivers 62% higher full-protein LDDT than MAFFT and 48% higher than MUSCLE5 for pairwise alignment. For multiple sequence alignment on the same benchmark, the gains are 57% and 50%, respectively.
With structures on BaliFam1000, our self-supervised geometric-consistency model reaches full-protein LDDT 0.785 versus the recently released LoL-Align's 0.777, and F1 0.95 versus 0.94. On AFDBv4, our structure-assisted multiple sequence alignments reach full-protein LDDT 0.62 against FoldMason's 0.54.
Email info@orikata.ai for the full benchmark results.
The Orikata web toolkit brings these workflows into your browser. Run analyses on your own data, compare our methods with established tools, and inspect the sequences, structures, and results together in one workspace.
To find relevant biology in these collections, we're extending Hikoboshi with vector search. We retain one learned vector per residue for alignment and project those representations into compact protein vectors for candidate search. That lets us use approximate nearest-neighbor (ANN) vector search to retrieve candidates before comparing them in detail.
Using precomputed representations for 462K proteins, ANN candidate search on a single CPU core was 500 times faster than Foldseek's prefilter, with an index over an order of magnitude smaller. The compact vectors retain biological similarity for protein search. You can try out the search function over the TEDBench dataset today, with larger datasets coming soon.
Those search and alignment models are part of the foundation we're building for Ibuki, the 'breath of life,' our world-model family for biology. We want it to learn across the breadth of life: how sequence, structure, and function relate, and how to use those relationships to understand and engineer biology. That connects folding, inverse folding, and generative design with multimodal search, clustering, alignment, ancestral sequence and structure reconstruction, evolutionary rate prediction, conservation, and phylogenetic tree building. The tools find evolutionary context, the models use it to predict and design, and better representations improve the tools in turn.
We're pre-training then fine-tuning our own models to use alongside the ESM- and ProteinMPNN-based models in today's Hikoboshi release. Ibuki-based workflow models are trained but not part of this initial release. Our goal is to connect what the family learns about sequence, structure, and evolution with the workflows scientists use to understand and engineer biology.
At Megure Labs, we're building Hataya around a simple idea: AI models should ship as software. Hataya is an inference engine that compiles models and their dependencies into optimized standalone programs for CPUs and GPUs. We're building it to run Hikoboshi and Ibuki alongside public scientific ML models.
The code that runs, the exact weights, and the evidence comparing its outputs with the original model ship together in a .hya model package. Scientists can inspect what ran, check the results, and run the model on their own hardware or through a hosted service.
AI should be reproducible. AI should be verifiable. AI should be yours.
Scientists and their agents also need a place to run the models and work with their results. We're launching Noto with hosted model inference, so you can run biological AI models without setting up the models and compute yourself.
The full Noto discovery workbench follows in mid-November, bringing human and agentic scientists into the same multiplayer workspace. Plan an experiment, work with its generated data, run an analysis or model, maybe even train a new model, and keep all the results in the notebook. The inputs, methods, and results stay connected across local and hosted compute, so both human and agentic scientists can see how a result was produced and build on it.
We call it a discovery workbench because we're building it for human and agentic scientists to work together as peers. Both can plan experiments, run analyses and models, and build on each other's work in the same shared notebook. We're starting with biology and building toward other areas of research.
Here's what we're launching today:
Hosted inference and agentic workflows are online now in Noto, with 45+ models and tools in addition to the Hikoboshi toolkit. The agentic Noto discovery workbench launches in mid-November, with new features added daily. You can use Noto on the web now. A desktop app is coming soon, with native iOS and Android support planned for early 2027.
Our curated corpora, precomputed representations, indexes, and hosted infrastructure are paid-access offerings. We offer individual and team plans and metered API usage. We're expanding web, API, and MCP access for scientists and their agents.
If you're doing protein engineering, biological ML, or agentic discovery, try the Orikata toolkit, run a model with Noto, or build with the open-source tools. We'd love to hear what you're working on. If you know a research group or biotech team that would find this useful, I'd love an introduction. For Umi and Sora access or the full Noto workbench, reply here or email info@orikata.ai.
We're strong believers in open science, and we hope you'll build the future of biology with us.
Orikata Bio: https://orikata.ai
Megure Labs: https://megure.ai
Umi: https://orikata.ai/umi
Sora: https://orikata.ai/sora
Orikata toolkit: https://orikata.ai/tools
Hikoboshi: https://github.com/orikata-bio/hikoboshi
Orihime: https://github.com/megure-labs/orihime
Ibuki: https://orikata.ai/ibuki
Noto: https://megure.ai/noto