Research Rundown: AcademAI - AI in Research

AI-Powered Graphical Abstracts: Building Your Paper's Best Advertisement

Santhosh Eapen — Mon, 03 Aug 2026 02:45:47 GMT

I have never met a researcher who enjoyed making a graphical abstract. We are trained to write. Nobody trained us to draw.

And yet the figure is the part that travels. Your methods section will be read closely by perhaps a dozen people. The graphical abstract gets screenshotted, dropped into someone else’s slide deck, and shown at a conference you were not invited to. The evidence supports this, though less cleanly than the enthusiasts suggest. Papers carrying a graphical abstract reliably attract more online attention and more downloads. On citations the literature is split. A study in gastroenterology and hepatology journals found a lift; a larger propensity-matched analysis across fields found none. So I would not promise you citations. Visibility, yes.

My career began nearly forty years ago, a time when the graphical abstract was unheard of. Their widespread adoption is a recent phenomenon, gathered mostly over the last decade. For the better part of my professional life, the visibility they offer was locked behind software I simply could not use. We were a generation raised on Word and PowerPoint; the Adobe suite was a luxury we could rarely justify. The choice was binary: spend months mastering a vector editor, or hire a professional who already had. Most of us chose a third path and just sent the manuscript without a figure.

That barrier has now gone, and I think this is one of the genuinely good things AI has done for working researchers. Ananya Thakur made the case in Nature recently and called it a democratizing force. She is right. But she also wrote the sentence that ought to sit above every one of these tools: the graphic carries your name, not the AI’s.

What follows is the method I would use. It is not the only one.

Before anything else, read the journal’s rules

This is dull and it is also the step that decides what you are permitted to submit, so do it first.

The policies are not consistent with one another. Nature allows AI-assisted figures but wants the tool disclosed, usually in the Methods, and separates AI-assisted data visualisation (broadly acceptable) from AI-generated conceptual illustration (more scrutiny, and never standing in for real data). Science is stricter and treats undisclosed AI content as misconduct. Cell Press wants the figure labelled in the legend and bars AI images that purport to show primary experimental data.

The rule that keeps you safe under all three: AI builds the schematic, never the result. The arrows, the cartoon cell, the pathway diagram. Not the data. Keep a record of the tool and the prompts, the way you keep a lab notebook. You will need it for the disclosure statement, and it takes thirty seconds at the time and an afternoon to reconstruct later.

Decide the one thing you are saying

Most bad graphical abstracts are not ugly. They are overloaded.

A figure that faithfully maps every interaction in your paper communicates nothing at all. The reader’s eye has nowhere to go. A figure carrying one idea gets remembered for years. This is the hardest part of the whole exercise and it happens on paper, before you open anything.

Three questions get you there. What shape is the story: a mechanism, a process, a set-up, a workflow, a comparison, a timeline? Naming it narrows everything downstream.

Who is looking at it? This one is underrated. A plant pathologist wants the infection pathway. An extension officer wants to know what to spray and when. Same finding, two different figures, and trying to serve both produces a figure that serves neither.

And then the sentence. Force the paper down to one claim you could say out loud. Mine, for a worked example: Phytophthora moves through the black pepper vine’s vascular system, and a soil-applied biocontrol stops it at the root collar. One pathway. One intervention point. Everything else has to argue for its place against that sentence, and most of it loses.

Our recent Gene paper linked two sibling species - Phytophthora capsici and P. tropicalis - to black pepper foot rot disease. The core takeaway is that this disease is caused by a species complex sharing a convergent carbohydrate-degrading and effector arsenal, which management strategies must target. The graphical abstract should have highlighted this headline finding, omitting supporting details like scaffold counts, assembly sizes, and exact gene numbers.

A chat model is useful here, and this is where I find it most useful of all. Paste in your abstract and ask it to strip the paper to a single message, name the audiences, and propose ways of drawing an abstract concept like induced resistance or systems-level interaction. Treat what comes back as a first sketch from a well-read colleague who has not done the work. Sometimes it sees the framing you had missed. Often it does not.

“Here is the abstract of my paper. In one sentence, what is the single most important finding a graphical abstract should communicate? Then suggest three simple visual metaphors for [your concept], and tell me which audience each would suit.”

Match the tool to the job

There is no best tool. There is a best tool for each part of the work, and the commonest mistake is asking one tool to do all of it.

Broadly: a conversational model to think with, a science-figure platform to build the accurate schematic, a vector editor to finish. Generative image models belong at the sketching stage and nowhere near the final file. They will draw you a chloroplast with the wrong number of membranes and do it beautifully.

One thing worth insisting on regardless of what you pick: the final file must be editable vector. SVG, EPS or PDF. A journal will ask you to change one label at proof stage and you do not want to rebuild the figure to do it.

Briefing the tool

Output quality tracks brief quality almost exactly. Three decisions are worth making deliberately before you type a prompt.

Colour, and how little of it to use. Warm colours pull the eye and read as danger or activity, which is why they suit the pathogen or the key interaction. Cool colours read as healthy, normal, baseline. That much is standard. The part people get wrong is quantity. Two or three colours, high contrast, used consistently, will beat a rainbow every single time. And roughly one man in twelve has some colour-vision deficiency, so colour must never be the only thing separating two elements. Shape or label has to carry it too.

“I want the infected vascular pathway in warm tones and healthy tissue in cool tones, on white. Suggest a limited, high-contrast, colour-blind-safe palette with hex codes.”

Composition, which is really about where the eye goes. Left to right for a process with a start and a finish. A loop for a cycle. A fork for a comparison or a decision. Decide before you generate, because retrofitting a flow onto a finished figure rarely works.

“Here is my draft graphical abstract. Suggest two alternative compositions and better label placements so the eye moves from infection to intervention.”

Opacity, occasionally. You will not need this often. But fading a background structure, the plant or the soil or the cell, lets the central reaction dominate while the reader still knows where they are standing.

“Which elements should stay visually dominant and which can fade back? Do I need a context background at all, or would plain white serve the message better?”

Then check it properly

The first output is a draft. Treat it as one.

This next part is the bit that cannot be delegated, and it is where the scientist rather than the software does the work. My list:

Variation should mean something. A small set of text sizes, line weights, colours. Change an arrowhead from open to solid and a careful reader will spend twenty seconds hunting for the significance of it. There is none. You have cost them twenty seconds.
White space is not wasted. A clean background usually beats a busy one. Clutter is what makes a figure forgettable.
Print it in greyscale. This single test catches most accessibility failures. If two colours collapse into the same grey, you have lost the colour-blind reader and the person reading a photocopy at the same time.
Give it to someone outside your subfield. Ten seconds. If they cannot tell you the message, it is not finished. This is uncomfortable and it is the most useful thing on the list.
Proofread every character in the image. AI models are careless with text inside graphics in a way that is genuinely startling. Then check legend, scale bars, units, abbreviations, gene and protein formatting, and the journal’s own rules.
Confirm the science. Every arrow. Does the schematic imply a relationship your data do not support? A handsome figure that overstates the finding is worse than no figure, because it will be believed and shared by people who have never read the paper
.

What has changed and what has not

The barrier is down. Someone with no design training can now produce a clear, publishable graphical abstract in an afternoon, and I would encourage every author to take advantage of that.

What has not changed is who answers for it. The tool did the labour. It does not carry the judgement. That figure goes out under your name, makes a claim about your science to people who will never read the paper, and stands as the most widely seen thing you publish that year.

While it is futile to complain about the subpar performance of Indian academic journals, it is high time they embrace graphical abstracts. This single addition would significantly enhance their impact and global reach. Given the abundance of modern tools designed to elevate our skills, researchers should actively leverage them to deliver clear, compelling presentations of their findings.

Use the tools. Then check the figure the way you would check a result, because that is what it is.

The short version

Read the journal’s AI policy before you start, and keep a record of tools and prompts.
Write the one sentence. Do not open a tool until you can.
Chat model to think, science-figure platform to build, vector editor to finish. Export editable vector.
Two or three colours, one clear flow, plenty of white space.
Greyscale test, proofread the text, verify every arrow, then show a colleague.

Which "AI Scientist" Should You Actually Use? A Field Guide for the Perplexed Researcher

Santhosh Eapen — Tue, 21 Jul 2026 02:30:15 GMT

Earlier this month, Euan Ashley, a geneticist and cardiologist at Stanford, did something that would have been science fiction a decade ago. In 2010, he led the first clinical analysis of a human genome, a project that took his team of 31 scientists nine months. This July, while unpacking after a holiday, he handed his own genome to Anthropic’s Claude and asked it to do the same job. Thirty minutes later, the tool had flagged an Alzheimer’s risk allele and gene variants affecting how he metabolises drugs. “There is no world in which this is not utterly remarkable,” Ashley wrote.

I felt that story in my bones. For the better part of a decade, my team worked to decipher the genome of the Phytophthora species that devastates black pepper - the very crop that once drew the world to India’s shores. We tried outsourcing. Then we trained ourselves. Then we built our own computing facilities. It still took more than ten years. So reading that a holidaying scientist had matched a nine-month, 31-person effort in half an hour was, for me, equal parts marvel and quiet ache.

If you have spent any time in a lab, that story lands somewhere between exhilarating and unsettling. A new class of software - variously called “AI scientists,” “research agents,” or “co-scientists” - now promises to compress hours of literature review, data wrangling, and analysis into minutes. But a fresh wave of these tools launched in just the past few weeks, each pitching itself as the one your lab needs. Claude Science. OpenAI’s Prism. Google DeepMind’s Co-Scientist. The open-source Biomni. And, researchers will tell you, many more.

So which one suits your work? Drawing on a recent Nature explainer by Ewen Callaway and the launch details of each platform, here is a practical guide to sorting the genuinely useful from the merely impressive - without losing your scientific judgement in the process.

What these tools actually are

Start with what they are not. An “AI scientist” is not the same as a specialized model like AlphaFold, which does one narrow thing - predict protein structures - extraordinarily well. These new tools are generalists. They are built on the large language models that power ordinary chatbots, but wrapped in an agentic layer: give one a request, and it breaks the job into steps, then recruits external software, databases, and sometimes specialised models to carry each step out.

In practice that means help with the unglamorous middle of research - literature reviews, data analysis, figure generation, protocol drafting, manuscript preparation. Gabriele Corso, who runs the protein-design firm Boltz, set a Claude agent loose on designing an antibody that recognised two therapeutic targets, using his company’s own folding tools. The AI’s output matched what his experienced designers would have intuited.

“Work that usually takes me hours now takes minutes, I can really spend my time on the science that needs a human.” - Yuanhao Qu of the start-up Phylo

That last phrase is the whole game. These tools are not replacing the scientist. They are clearing the underbrush so the scientist can get to the clearing faster.

The main players, and what each is good at

The field is crowded, but the leading tools sort into recognisable roles. (I wrote about Claude Science in this newsletter not long ago; it is only one of a fast-growing crowd, so let me start there.)

Claude Science (Anthropic, launched 30 June 2026) is a workbench rather than a new model - it runs existing Claude models inside an environment purpose-built for biology. A coordinating agent draws on more than 60 curated skills and connectors for genomics, single-cell, proteomics, structural biology, and cheminformatics, and can spin up specialist sub-agents. It renders 3D protein structures, genome browser tracks, and chemical structures natively, and - importantly - a reviewer agent checks citations and calculations and flags errors. It is strong for hands-on analytical tasks late in a project.

Co-Scientist (Google DeepMind) sits at the opposite end of the workflow. It is a hypothesis machine: a multi-agent system that generates ideas, then debates and ranks them through a kind of internal tournament before evolving the best. Clare Bryant, a Cambridge immunologist and early adopter, fed it a grant application and her data. Some suggestions were unworkable; others were squarely in her wheelhouse. Her lab is now testing one - an experiment she reckons might otherwise have taken two years to arrive at. “You feel like you’re talking to an oracle,” says Stanford’s Gary Peltz, who used it to find existing drugs for a liver-fibrosis model.

Biomni (Stanford, published in Science in 2026) is the open-source option, already used by more than 15,000 scientists and now spun out into the start-up Phylo. It handles casual-language prompts - “Why are these patients responding differently to the drug?” - and was tested across more than 400 research tasks. It approaches human-level performance on database querying, sequence analysis, and molecular cloning, but the team is candid that it still struggles where nuanced clinical judgement or deep biological synthesis is needed.

Prism (OpenAI) is narrower in scope but useful: a free, LaTeX-native writing and collaboration workspace powered by GPT-5.2, aimed at drafting, revising, and preparing manuscripts for publication rather than running analyses.

How to actually choose

Here is the uncomfortable truth from the people building and studying these tools: fewer than 20% of labs have genuinely embedded AI scientists into their research. Most exposure is still the shallow kind - generating slides, drafting emails. “It’s really important that people actually try these things out,” says Ashu Singhal of the platform Benchling, “rather than simply trusting what gets shared in headlines.”

Three principles cut through the noise.

Match the tool to the stage of your project. Hypothesis-generating systems like Co-Scientist earn their keep at the very start, when you are still deciding what is worth doing. Analytical workbenches like Claude Science and Biomni pay off later, on defined tasks such as genomic data analysis. A writing tool like Prism belongs at the end. No single tool owns the whole pipeline, so expect to trial several.

Start small and verifiable. Corso’s advice is the most practical thing in the entire discussion: begin with tasks whose output you can check easily. “Worst case, you have to do them over.” A small, checkable task teaches you where a tool is reliable and where it quietly invents things - at a stage where a wrong answer costs you an afternoon, not a paper.

Weigh openness against convenience. An open-source tool like Biomni gives you transparency and no licence fee, at the cost of setup effort. A commercial workbench gives you polish and support, at the cost of dependence on a vendor. For an academic lab used to assembling its own pipelines, that trade-off may already feel familiar.

Can you trust an AI scientist?

The honest answer is: sometimes, and only if you check. The Science paper on Biomni found it matched experts on certain tasks - rare-disease diagnosis, cell-sequencing data analysis - often in a fraction of the time. On tasks demanding deep biological reasoning, it fell short.

Anthony Gitter, a computational biologist at the Morgridge Institute for Research in Wisconsin, put several general-purpose tools through a tougher test: a 2025 community challenge to predict preclinical drug properties such as toxicity, judged against unreleased experimental data from a biotech company. The tools did reasonably - but not as well as human experts, including those at large pharmaceutical firms. “They were doing generally reasonable things,” Gitter says, “but it was not pushing the boundaries.” His hands-on experience with Biomni was mixed: “Some things I wanted to try with it worked well, and some crashed pretty badly.”

That unreliability is survivable - a tool does not have to be right to be useful - but only if it shows its working. The feature researchers prize most is not the answer itself but the reasoning trail behind it. “When using any of these tools, it’s my responsibility to know what claims are being made and how those claims are being supported,” says Gitter. Some models do this far better than others, and that transparency should weigh heavily in which one you pick.

The catch nobody advertises

For Ashley, the productivity gain has been “unspeakably huge.” But the scientists Nature spoke to share a quieter worry worth taking seriously: what these tools do to the people still learning the craft. “You only learn to see certain types of mistakes by having made them yourself many times in the past,” Gitter notes - and an agent that clears away the grunt work also clears away part of the apprenticeship. The same risk shadows experienced scientists working outside their own fields, where the instinct for “that can’t be right” simply isn’t there.

Clare Bryant encourages her trainees to use Co-Scientist to form ideas, but she leans on decades of experience to catch its errors. “I have the experience to know when something’s not true,” she says. “The biggest question I have is, how do I train students.” That tension - between the acceleration these tools offer and the hard-won judgement they can quietly short-circuit - is the one every lab will have to manage for itself.

A note from where I sit

In India, where I have spent four decades in crop-protection research, AI tools are still treated by many as something close to taboo - a shortcut to be suspected rather than a scalpel to be mastered. I understand the caution; I lived the years of doing everything the hard way, and I know what it costs to trust a result you did not build by hand. But these tools are evolving at a pace that caution alone cannot outlast. Scientists across the globe now lean on them daily, and the refrain is remarkably consistent: work that used to take hours takes minutes, freeing them to spend their time on the science that genuinely needs a human. That is not a threat to good science. Used with eyes open, it may be the most time our discipline has ever been handed back.

What to do this week

If you are curious but unconvinced, four concrete steps will tell you more than any launch announcement:

Pick one real, small task from your current project - a literature scan, a tidy dataset, a figure - whose correct answer you already know or can quickly check.

Match a tool to it. Need ideas? Try a hypothesis generator. Need analysis? Try a workbench. Need a cleaner manuscript? Try a writing tool.

Run it, then audit the trace. Do not just read the answer - follow how it got there. That is where you learn whether to trust it on a task you can’t check.

Compare two tools on the same task. The differences in how they reason, cite, and fail will teach you more than any review, including this one.

The genome that took 31 people nine months now takes one person half an hour. That shift is real, and it is not slowing down. The researchers who benefit first will not be the ones who trust the headlines, nor the ones who dismiss the whole thing as hype - but the ones who roll up their sleeves, run a small experiment on the tools themselves, and keep their scientific scepticism switched firmly on.

Additional Reading

Anthropic. (2026, June 30). Claude Science, an AI workbench for scientists. Anthropic. https://www.anthropic.com/news/claude-science-ai-workbench
Callaway, E. (2026, July 10). Which ‘AI scientist’ suits your lab? A guide for the perplexed. Nature. https://doi.org/10.1038/d41586-026-02091-6
Huang, K., et al. (2026). Autonomous biomedical research with an artificial intelligence agent. Science. https://doi.org/10.1126/science.adz4351
Google DeepMind. (2026). Accelerating scientific discovery with Co-Scientist. Nature. https://doi.org/10.1038/s41586-026-10644-y
OpenAI. (2026). Introducing Prism. OpenAI. https://openai.com/index/introducing-prism/
Santhosh Eapen (2026, July 07). Claude Science: One workbench, not a dozen tabs. https://open.substack.com/pub/sanjoe

Subscribe now

Claude Science: One Workbench, Not a Dozen Tabs

Santhosh Eapen — Tue, 07 Jul 2026 01:30:35 GMT

Four decades in crop protection, biotechnology, and bioinformatics have taught me one unglamorous truth: research is rarely slowed down by hard questions. It is slowed down by fragmentation. A day’s work means logging into PubMed for one thing, opening a Jupyter notebook for another, switching to R for a plot, then SSH-ing into a cluster to submit a job and waiting to see if it succeeded. Each tool is fine on its own. Stitched together, they cost hours that should have gone to science.

Anthropic’s new Claude Science - released in beta on June 30, 2026 - is built squarely around that problem. It is not a chatbot that talks about biology. It is a desktop workbench that runs your analyses, queries the databases behind them, and keeps a full, checkable record of how every result was produced. Let’s go through what it does, how to install it, and how researchers are actually using it.

What It Is

Claude Science is a desktop application, available in beta for macOS and Linux, that pairs Claude with a real analysis environment on your own computer. You describe a task in plain language - “cluster these single-cell samples and annotate the marker genes” - and Claude writes and runs Python, R, or shell code inside a sandbox, reads whichever folders you grant it, pulls data from more than 60 scientific databases through built-in connectors, and saves the results as versioned artifacts with a complete provenance record. A background reviewer agent independently checks Claude’s claims against what actually ran, flagging anything that doesn’t hold up.

Your files stay on your machine. Code runs in a sandbox that is deny-by-default on the network; you approve each new folder, host, or remote job before Claude touches it. Anthropic is direct about its limits: the reviewer reduces errors, it doesn’t eliminate them, and Claude Science is explicitly a research tool, not intended for clinical or diagnostic use.

Key Features

Reproducible artifacts - every figure ships with its code, execution log, and environment.
Native renderers - proteins, alignments, genomic tracks, chemical structures, no extra viewers.
The reviewer - flags unsupported claims, bad citations, mismatched DOIs.
Scales compute - persistent kernels locally; SSH to your HPC/SLURM cluster or Modal for GPUs on demand.
Domain-ready connectors - Ensembl, UniProt, PDB, AlphaFold, PubChem, ClinVar, and more, on by default.
Skills - AlphaFold2, ESMFold2, ProteinMPNN, Evo 2, scGPT, and others; you can add your own.

Installation

macOS: download installer from claude.com/product/claude-science, double-click.

Linux:

curl -fsSL https://claude.ai/install-claude-science.sh | bash

claude-science serve

Windows: no native build yet - run the Linux binary under WSL 2 (Ubuntu 24.04+).

Requires Pro, Max, Team, or Enterprise. Sign in with your Claude account - no API key needed.

Workflows

Pre-configured for single-cell RNA-seq, phylogenetic/evolutionary analysis, protein structure work, and cheminformatics. Open a project, point Claude at a folder, approve its permission requests, and results land as artifacts in the Files panel. For heavier runs, connect your lab’s HPC cluster or Modal account under Settings > Compute, and Claude drafts, submits, and retrieves the job.

In beta use: Manifold Bio for drug-target nomination, the Allen Institute for a 20-skill literature-review pipeline, and UCSF for germline variant studies - including catching a virus contaminant in RNA-seq data that had stumped the team for a year.

Why It Matters Here

For crop protection and plant bioinformatics work, the same fragmentation applies - just with different databases. A single environment that traces every figure back to its code is exactly the reproducibility standard grant reviewers and journal editors expect.

Anthropic is also running an AI for Science grants program - up to $30,000 in credits, applications open through July 15, 2026.

Getting Started Today

Claude Science is in beta on Pro, Max, Team, and Enterprise plans. Documentation covering installation, connectors, and admin setup lives at claude.com/docs/claude-science. Download it, sign in with your existing Claude account, and open the bundled Example project to get a feel for how permission cards and artifacts work before pointing it at your own data.

The most interesting thing about the tool, in my early impression, isn’t any single feature - it’s that provenance is not an afterthought bolted onto the output. It’s structural. That is exactly the standard research has always needed and rarely gotten from software built for speed alone.