CHIMERA RESEARCH
Sign up

We adapt biofoundation models to each scientist's problem.

One system for scientists, agents, and labs, so every result improves the models.

Request a Demo
Working with
  • MD Anderson
  • UCLA

Science is specific. Its tools are not.

Models tailored to each scientist's workflow, data, and constraints are largely unavailable.

  • Expertise

    Adapting a frontier model takes research skill most labs do not have on staff.

  • Cost

    A single training run can cost more than the project it was meant to serve.

  • Infrastructure

    Hosting and serving a frontier model is an engineering problem of its own.

  • Time

    Weeks of wall clock per attempt, with no guarantee the result beats where you started.

Chimera workbench
Training lossValidation loss
Training and validation loss over training stepsTwo curves descending from left to right and flattening, the validation curve sitting slightly above the training curve. Illustrative, not measured data.losstraining steps
Run
Adaptation run, generic model A
Steps
Placeholder
Status
Placeholder

Train

Adapt a frontier model to your target.

  1. Model agnosticModel agnostic at any scale. We host the frontier open-source models.
  2. Managed computeTraining compute and infrastructure abstracted away.
  3. Your dataCustom, data-driven adaptation techniques built around the data you already have.
De Novo Antibody Design. De novo antibody / binder design pipelines.
Autoregressive Sequence Models. Autoregressive sequence models.
Sequence & Structure Retrieval. Sequence and structure database access.
Gene Annotation. Sequence annotation.
Inverse Folding. Sequence design from structures.
Masked Language Models. Masked language models.
Molecular Docking. Protein-ligand binding pose prediction.
Mutagenesis. Random sequence mutagenesis.
ORF Prediction. Open reading frame detection.
RNA Splicing. RNA splice site prediction.
Sequence Alignment. Sequence search and multiple sequence alignment.
Sequence Scoring. Genomic and regulatory scoring.
Structure Alignment. Structure comparison.
De Novo Structure Design. De novo structure generation.
Structure Dynamics. Conformational dynamics.
Structure Prediction. 3D structure prediction.
Structure Scoring. Structure quality scoring.

De Novo Antibody Design

  • bindcraft
  • freebindcraft
  • germinal

01 / 17 tool families

A single job lane resolving into a deep grid of identical compute nodes receding into the dark.
Assorted data tables and assay plates funnelling into a model block whose weights are partly rewritten, above a descending training curve.
Protein crystals seen through a microscope at high magnification. Angular, transparent crystals of varying size scattered across the field.

Example

A small-molecule hit design model for one protein target.

Fine-tuned on a partner's cryo-EM, crystallography and surface plasmon resonance data.

Compute to get that model

Pose accuracy on that target

Full retrain~50,000 GPU-hours
Chimera fine-tuningPose accuracy50 GPU-hours71%+6
Pose accuracy71%+6

1,000×less compute

Deploy

Put that model in front of the people and labs doing the work.

  1. Same environmentModels run in the same environment they trained in, so nothing breaks in between.
  2. VersionedEvery model version is checkpointed and labelled, so choosing what to run is one decision.
  3. Lab readyAutomated SOPs that instruct human wet labs and cloud labs within their real constraints.
Two identical enclosures holding the same stack of layers, with one unbroken line running straight through both.
A vertical spine of evenly spaced checkpoints, each with a label tag, one picked out by a selection bracket.
A protocol step list routing into two destinations, a well plate with a pipette and an automated instrument, each bounded by constraint brackets.

Example

Proposed molecules routed from the model to three benches.

Cell biologists and medicinal chemists approve the designs, then the platform sends them to human wet labs, cloud labs, and structure verification.

Iterate

Every result feeds the next version of the model.

  1. Side by sideRun a new model version alongside the current one and compare on live work before switching.
  2. Always growingAdd new data, collaborators, or literature at any point and the models improve with it.
  3. Agent reviewAgents trace errors to root cause and propose the fix.
One incoming stream splitting into two parallel lanes carrying the same work, a difference strip between them, and a routing gate still set to the current lane.
A dense core of connected nodes ringed by newer ones, with data tables, people and papers arriving inward from every edge.
A dependency tree with a failure at one leaf and a lit path traced back to the ringed origin node, beside a small proposed patch.

Example

Agent-driven design improvement.

An agent redirected the screen to a different binding pocket, and it returned confirmed binders.

Your model stays with you.

Every model is containerized by default. Open-source it only if you choose to, so the community and other partners can build on it.

Request a Demo