ScienceMachine

AI + data + biology

[ScienceMachine @SciMac]

Big thanks to the Inoviv team for the collaboration so far. Excited for what’s ahead!

Full case study here: https://t.co/T3ev6TRO6G

[ScienceMachine @SciMac]

Excited to announce ScienceMachine’s partnership with Inoviv to bring AI automation into proteomics workflows. 🚀

Together, we’ve reduced time spent on QC and reporting by ~80% and increased lab capacity by 20% — giving scientists more time to focus on science, not admin.

[Zane Koch @zanehkoch]

How capable are AI agents at bioinformatics research?

I tested 9 LLMs on BixBench: 205 complex tasks from real-world research – the kind I do as a compbio PhD

Results:
- 63% accuracy (3x higher than the SOTA 10 months ago)
- The harness mattered as much as the model

[ScienceMachine @SciMac]

So we are asking for your advice: try out our platform with your most tedious scientific tasks and let's schedule a call on how we can improve the platform!

[ScienceMachine @SciMac]

🤯 “Really, really good results!” — we’re hearing this more and more from our users and customers

> see below

[ScienceMachine @SciMac]

But with her and others’ expert feedback, we went heads-down building and shipped a fine-tuned product, fit-for-purpose for complex scientific work. And we’re only getting started!

[Ben Tenmann @BTenmann]

🚨 Studies suggest that writing a thesis or a paper increases your risk of sleep deprivation by 100%

Or at least that was the case for me…

I would be pulling all-nighters, days before the big deadline

And nothing was more nerve-wracking than finding a bit of contradictory evidence or a bug in the data pipeline at the eleventh hour

It was this personal trauma that motivated @SciMac. And so it is super fun to see folks using it to win back those precious sleeping hours!

PhDs, Post-docs and PIs have used our platform to create publication-ready plots, interpret biological mechanisms, and automate painful workflows

We’ve upped the free usage tier, so feel free to give it a spin (here: https://t.co/eF641zTWTq)

[pic is of me after submitting my master's thesis w/ 0h of sleep]
@lorenzosani_ @rekatronics @robbiemccorkell

[ScienceMachine @SciMac]

From Richard Littlewood: "ScienceMachine transformed our research process at Flow Health Science Inc., streamlining the complex data analytics needed to launch our first product, Klario™”

🧬🚀

[ScienceMachine @SciMac]

📝 ScienceMachine powering real research!

Flow Health Science used SM to analyse and interpret their data, and helped them get published much quicker!

Read the paper here: https://t.co/PACjMWHGL1

[ScienceMachine @SciMac]

This is just the beginning. And we need your help to make it even better.

📣 So here's the ask 📣: try out our platform with your most tedious scientific tasks!

https://t.co/KAYbVaaUIP

[ScienceMachine @SciMac]

🚀🧬 Today, we’re launching the next generation of work in life sciences R&D

Over the past months we have been heads-down building, working closely with domain experts and leading biotech companies to define how AI will augment research. And today we're making it public!

[ScienceMachine @SciMac]

To this end, we:

🤖 Built a faster, smarter, and more accurate AI agent, using the latest models and techniques
🧪 Added enhanced support for scientific workflows, such as ELISA, flow cytometry, and proteomics
💾 Made it easier to work where your data already lives

[Bo Wang @BoWang87]

Over the past year, there’s been a surge of excitement around agentic AI — systems that don’t just answer questions, but can act: reading instructions, running code, designing pipelines, and making decisions.

In biomedicine, this raises a provocative question:

💡 Could the next member of your ML team be an AI agent?
The honest answer — not yet.

Today, we share BioML-bench, a new open benchmark to measure how far today’s agents are from this vision, and what it will take to get there.

📄 Paper : https://t.co/MI6Wxq3CWK

💻 Code: https://t.co/yIG7JIOKjm

Why this matters
Biomedical discovery doesn’t happen in a single step.
It’s messy, iterative, and deeply interdisciplinary: cleaning data, choosing models, validating results, integrating diverse domains like genomics, imaging, and clinical records.

Existing evaluations — mostly Q&A or coding challenges — don’t capture this complexity.
We needed a testbed that reflects the real work of biomedical ML.

What we built

BioML-bench is a suite of 24 real biomedical ML tasks where agents must:
--Parse nuanced task descriptions
--Build and train models end-to-end
--Compete against human leaderboards populated by domain experts

It’s the first benchmark designed to ask: Can an agent truly operate like a biomedical data scientist?

What we learned
Our experiments with four different agents — from general-purpose systems to biomedical specialists — reveal a sobering truth:
--Current agents operate at ~35% of human expert performance.
--Domain specialization alone isn’t enough. Success comes from flexible, creative strategies, not rigid pipelines.
--Even on imaging tasks, deep learning was underutilized, highlighting a gap between human and agent intuition.

Looking ahead
The promise of agentic AI isn’t to replace human scientists — it’s to amplify them.
Imagine a future where an agent can set up a first-pass analysis overnight, freeing a scientist to focus on questions, not debugging scripts.

We’re not there yet. But with BioML-bench, we now have a shared yardstick to track progress, spark innovation, and bring accountability to this emerging field.

Grateful to our amazing team — led by @Henrymiller2012 , with contributions from Matthew Greenig, Benjamin Tenmann, and support from @SciMac.

This work is a small but necessary step toward a future where AI becomes a true partner in biomedical discovery. 🌱

#AI #Biomedicine #Agents #MachineLearning #BioML

[Henry E. Miller @Henrymiller2012]

🤖 Could an 𝐀𝐈 𝐚𝐠𝐞𝐧𝐭 be your biotech's newest 𝐌𝐋 𝐒𝐜𝐢𝐞𝐧𝐭𝐢𝐬𝐭? ✨

We built a benchmark to find out.

𝘛𝘓𝘋𝘙: 𝘛𝘩𝘦 𝘰𝘱𝘦𝘯 𝘴𝘰𝘶𝘳𝘤𝘦 𝘢𝘨𝘦𝘯𝘵𝘴 𝘸𝘦 𝘵𝘳𝘪𝘦𝘥 𝘥𝘰𝘯'𝘵 𝘲𝘶𝘪𝘵𝘦 𝘩𝘪𝘵 𝘵𝘩𝘦 𝘮𝘢𝘳𝘬 𝘺𝘦𝘵. 𝘉𝘶𝘵, 𝘯𝘰𝘸 𝘸𝘦'𝘷𝘦 𝘨𝘰𝘵 𝘢 𝘣𝘦𝘯𝘤𝘩𝘮𝘢𝘳𝘬𝘪𝘯𝘨 𝘴𝘶𝘪𝘵𝘦 𝘵𝘰 𝘵𝘦𝘭𝘭 𝘶𝘴 𝘸𝘩𝘦𝘯 𝘸𝘦'𝘳𝘦 𝘰𝘯 𝘵𝘩𝘦 𝘳𝘪𝘨𝘩𝘵 𝘵𝘳𝘢𝘤𝘬 𝘢𝘴 𝘸𝘦 𝘤𝘰𝘯𝘵𝘪𝘯𝘶𝘦 𝘵𝘳𝘺𝘪𝘯𝘨 𝘯𝘦𝘸 𝘢𝘨𝘦𝘯𝘵𝘴.

🔎  𝗪𝗵𝗮𝘁 𝘄𝗲 𝗳𝗼𝘂𝗻𝗱
• We benchmarked two biomedical specialists (Biomni, STELLA) and two ML generalist agents (AIDE, MLAgentBench).
• Biomedical specialization did not confer a consistent advantage as Biomni and AIDE generally performed better than STELLA and MLAgentBench overall.
• Agents consistently underperform human baselines in general (avg. 34-37% of human leaderboard performance).
• The best-performing agents tried more diverse ML strategies (feature engineering, model selection, stacking) rather than sticking to a single approach.
• Deep learning was rarely used by agents, even on imaging tasks, despite the fact that human leaderboards were dominated by DL models.

🛠️  𝗪𝗵𝗮𝘁 𝘄𝗲 𝗯𝘂𝗶𝗹𝘁
• BioML-bench – a benchmarking suite for agentic BioML.
• Built upon MLE-bench, agents must parse task descriptions, build & train models, and submit predictions graded against human leaderboards.
• 24 biomedical-specific ML tasks covering multiple biomedical domains, with human leaderboards mostly populated by experts.
• A software package to lower benchmarking barriers and enable reproducible evaluation.

🔭 𝗪𝗵𝘆 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀
Agentic AI holds promise for automating biomedical R&D. However, realizing this promise will require agents that can reliably complete end-to-end data analysis workflows and build predictive models. Currently, most agents are evaluated by text-based question answering, leaving a gap where practical evaluations of end-to-end BioML capability are needed. BioML-bench provides the first benchmarking suite to fulfill this need. As agents become a stronger focus for biomedical researchers in the coming years, we hope BioML-bench will serve as the standard benchmark for evaluating them in their biomedical ML capabilities.

👇 𝗗𝗶𝘃𝗲 𝗶𝗻
• 𝐏𝐚𝐩𝐞𝐫: https://t.co/hQezVbGeNb
• 𝗖𝗼𝗱𝗲 & 𝗱𝗼𝗰𝘀: https://t.co/kckFkfDwze

Congrats to my co-authors Matthew Greenig, Benjamin Tenmann, and @BoWang87.

Thanks to @g27182818 for invaluable feedback. And thanks to @SciMac for providing compute and LLM API resources for this work.

[ScienceMachine @SciMac]

Nothing more annoying than a plot that is just not quite right 🤬

We’ve added simple *plot editing* for your most important figures, so you can visually edit things that are almost perfect!

https://t.co/KAYbVaamTh

[Lorenzo @lorenzosani_]

thanks @MartinJBCoulter for writing about @SciMac and how we accelerate life sciences research with AI agents! 🧬🧪

@Siftedeu @BTenmann
https://t.co/TxgbPHT46A

[Ben Tenmann @BTenmann]

thanks @MartinJBCoulter for writing about ScienceMachine and how we accelerate life sciences research with AI agents! 🧬🧪

@Siftedeu @lorenzosani_
https://t.co/zrLlBCFPZR