Inspect protein structures, compare molecules, explore genomic regions, and create figures with the code behind them—in a research workspace your team can share.
A useful research discussion should survive three questions: What are we looking at? Where did it come from? Can someone else check it?
That is the standard we want AI for scientific research to meet.
Today, we’re introducing CoChat Science Workbench, bringing interactive protein structures, molecular drawings, a genome browser, and Python-generated figures into the conversation where you investigate a research question. These tools extend CoChat’s existing literature search, citation checking, shared projects, and choice of AI models.
The point is not to replace scientific judgment with a more confident answer. It is to make the material behind an answer easier to inspect, question, revise, and share.
Start with one structure, one genomic region, or one CSV. Build from there.
From an AI answer to something you can inspect
Science Workbench adds four ways to work directly with scientific material.
| Your next task | What you can do in CoChat | Why it is useful |
|---|---|---|
| Inspect a protein model | Open a 3D structure and view AlphaFold predictions with confidence coloring. | Put the model—and its uncertainty—into the discussion. |
| Compare chemical structures | Draw molecules alongside calculated properties and their SMILES representations. | Check molecular identity before interpreting a property comparison. |
| Investigate a genomic region | Open a genome browser, add supported tracks, and mark a region for follow-up. | Give a collaborator or the assistant a precise location to discuss. |
| Revise a scientific figure | Run Python, inspect the executed code, and update the existing figure through a code change. | Review the transformation from data to image, not just the final picture. |
The specialist viewers focus on biological and chemical research. The figure workflow is useful more broadly: an environmental time series, a materials experiment, or a behavioral dataset can also start with a table and an explicit analysis question.
Protein structure visualization: keep confidence in view
A structure discussion is more useful when everyone can inspect the model being discussed.
CoChat can find experimental structures through the Protein Data Bank, open uploaded PDB or mmCIF files, and display predicted models from AlphaFold DB. The interactive viewer supports different representations and residue or ligand highlights, so you can focus the view on the question rather than describing the entire protein.
For example:
Show the AlphaFold DB model for human p53, P04637, colored by pLDDT. Explain what the confidence colors tell me—and what they do not establish.
In the example prepared for this launch, CoChat retrieves the predicted p53 model and displays it with the confidence legend and source attribution. You can inspect the structure without treating every region as equally reliable.
That distinction matters. pLDDT describes local, per-residue prediction confidence; it does not establish the relative orientation of separate domains. Low confidence can reflect flexibility or insufficient information for a confident prediction, rather than a simple verdict that a region is “wrong.”
For a structural biologist or protein biochemist, this provides a starting point for checking a model before using it to motivate an experiment. Ask which structure was retrieved, which residues are represented, and what additional evidence would be needed for the interpretation you are considering.
The viewer gives you an inspectable object. Biological function still needs biological evidence.
Molecular visualization: check the structure before comparing the numbers
A chemical name, a structure, and a calculated property should stay connected.
CoChat can look up a compound and turn its returned structure into a 2D molecular drawing with calculated descriptors, including molecular weight, logP, hydrogen-bond counts, and polar surface area. The molecule cards keep the SMILES representation alongside the drawing.
Try:
Look up aspirin, ibuprofen, and naproxen. Draw their structures side by side, highlight the carboxylic acid group, and compare their RDKit-calculated properties.
The example uses retrieved PubChem structures rather than molecular strings invented by the assistant. Naproxen’s specified stereochemistry is preserved; the ibuprofen representation does not specify a single enantiomer.
For a chemist reviewing a compound list, this makes an important check part of the workflow: is the structure being analyzed actually the structure you intended?
It is also useful in a research meeting. Instead of presenting a property table without its chemical context, you can discuss the drawings and the descriptors together. When reviewing your own analogue series, that gives you a clearer basis for deciding which comparisons deserve further analysis.
Calculated descriptors are not experimental potency, selectivity, or safety results. Use them to organize questions—not to declare a molecule successful.
An interactive genome browser: make “this region” precise
A marked genomic location is a better starting point than a description of where someone thinks they saw something.
CoChat brings an igv.js genome browser into the research conversation, with a reference assembly, gene annotations, and support for adding your own genomic tracks. You can navigate to a region and mark it for a follow-up question.
Start with a reference-only example:
Open TP53 in the hg38 genome browser.
Then, for an analysis using your own files:
Add my alignment and its index. Keep the reference assembly explicit, and help me inspect the region I mark.
The workflow is useful for discussing a coverage change, a candidate variant, or a feature that needs closer inspection. A student can point a supervisor to the same interval. A computational collaborator can ask for a quantitative check of that interval without first reconstructing its coordinates from a screenshot.
The browser locates the question; analysis must still answer it. If you need read depth, allele counts, or an overlap calculation, ask for a computation from the underlying files rather than an estimate from the picture.
File preparation remains important. Use the assembly your data were aligned to, and supply the required index files for indexed access; igv.js documents that missing index information can lead to entire-file reads. Efficient browsing is not a promise that every dataset, file size, or hosting configuration will work without preparation.
Nor does a visible variant establish pathogenicity. Treat the browser as a way to inspect evidence, not as a diagnostic result.
Scientific figures with receipts
A figure should come with a way to inspect how it was made.
CoChat executes Python against supplied data and returns the resulting figures to the conversation. The figure’s Show code control exposes the program that ran, and the provenance record includes input-file and environment information.
For the launch example, we generated a clearly labeled synthetic assay dataset, then plotted the individual observations with means and standard deviations. The demonstration is not a study, a drug comparison, or evidence of an experimental effect.
A useful starting prompt for your own data is:
Inspect this CSV before plotting. Tell me which columns you will use and how you will handle missing values. Show individual observations, define any error bars, and preserve the code behind the figure.
The benefit is not just getting a chart. It is being able to ask the questions that make a chart reviewable:
- Which rows and columns did the code use?
- Were any observations excluded?
- Does an error bar represent SD, SEM, or a confidence interval?
- Was a curve fitted, or were group means simply connected?
- Does the caption describe the calculation that actually ran?
Then revise it:
Remove the gridlines. Keep the observations, summary statistics, axes, and labels unchanged.
CoChat updates the plotting code and reruns it to replace the existing figure, rather than painting over the image. We tested that exact edit on the synthetic example: the figure kept the same artifact identity, and the run still plotted all 64 simulated observations.
This is particularly useful during manuscript revision. A request to change an axis or clarify a caption can be addressed in the same conversation as the original analysis, with the code available for review.
For more precise feedback, CoChat also supports comments pinned to locations on a figure and sent back to the conversation. A note about “this peak” or “this label” can refer to a specific spot instead of forcing a collaborator to describe its position.
Traceability supports reproducibility; it does not replace the work of making an analysis reproducible. Keep the underlying data, record external dependencies and random seeds where relevant, and review the method before using an output in a publication.
Keep the literature, analysis, and review together
The new viewers are most useful when they remain connected to the research question.
Science Workbench sits alongside CoChat’s scientific research tools, including literature search, structured review tables, and citation checking. That creates a practical path from reading a claim to inspecting the material relevant to it.
For example, a research team could:
- Find papers relevant to its question and record what each actually reports.
- Open a reported structure or an appropriate reference model.
- Inspect a supplied dataset and generate a figure with explicit analysis choices.
- Review the draft interpretation against the sources and the code.
Keep the boundaries between those steps visible. A paper’s reported finding, a computed result from your data, and a new hypothesis are different kinds of information. They should not become interchangeable simply because they appear in the same conversation.
The same principle applies to citation checks: establishing that a reference exists is different from establishing that it supports a particular sentence. Use source checking to support review, not as permission to stop reviewing.
For the literature side of the workflow, see our guide to building a structured AI literature review.
Built for a research group, not just one person’s chat history
The useful handoff is the context behind the result—not only the exported image.
CoChat supports shared conversations and projects, so collaborators can work with the same research context rather than pass around disconnected summaries. The new scientific artifacts belong in that shared workflow.
Consider a lab-meeting follow-up. A postdoc prepares a figure, a statistician questions the summary, and a principal investigator asks how the result relates to a cited paper. The goal is to keep those questions attached to the analysis they concern, rather than turn each into a new thread with a new explanation.
CoChat also lets you work with Claude, GPT, and Gemini within its research workspace. That can be useful for seeking a second critique of an interpretation or checking whether another model notices an unstated assumption.
Agreement between models is not independent scientific validation. The deciding evidence remains the source material, the method, and the result you can inspect.
Looking for a Claude Science alternative?
CoChat Science Workbench offers a browser-based option for teams that want scientific artifacts, shared research context, and a choice of AI models.
Anthropic’s Claude Science also supports scientific viewers, code-backed figures, annotations, and reviewer agents. These are useful capabilities, not features we claim to have invented.
The difference to evaluate is how the workbench fits your team’s work. CoChat brings the scientific tools described here into its existing shared research workspace, with browser access and multiple AI models.
Anthropic’s announcement describes a macOS and Linux application with connections to lab HPC resources over SSH and on-demand compute through Modal. If cluster orchestration is central to your workflow, compare those requirements directly; this release should not be read as a claim of identical compute capabilities.
A note on local computing: We’re also introducing a separate companion app—a local program that allows remote CoChat to access your machine and perform tasks there. We’ll cover its setup, access boundaries, and workflows in a dedicated launch post.
Start with one research question
You do not need to move an entire project to try the workbench.
Choose a first task:
- Inspect a prediction: “Show the AlphaFold DB model for p53, P04637, colored by pLDDT.”
- Compare structures: “Look up aspirin, ibuprofen, and naproxen, and draw them with calculated properties.”
- Open a locus: “Show TP53 in the hg38 genome browser.”
- Review a figure: Upload a suitable CSV and ask for a plot with explicit data handling, labeled uncertainty, and inspectable code.
Start with public or non-sensitive material. Before uploading unpublished, proprietary, or human-subject data, check your institution’s requirements and the relevant processing and access arrangements.
The goal is straightforward: make it easier to move from a question to something your team can inspect.
