
You can find me on Bluesky 🦋! bsky.app/profile/bunnech.bsk…
✓Assistant Professor at @EPFL in Computer Science and Life Sciences. PostDoc at @Genentech and @Stanford. PhD at @ETH and @BroadInstitute.

You can find me on Bluesky 🦋! bsky.app/profile/bunnech.bsk…

If we want AI agents to support Molecular Tumor Boards, we need benchmarks that capture real multimodal, longitudinal patient journeys. We introduce MTBBench (NeurIPS 2025 D&B track, Poster #1710) to evaluate AI agents in multimodal, sequential clinical decision-making settings.

Super important new benchmark is proposed here. MTBBench, creates the first test that actually checks whether an AI can handle real cancer cases that unfold over time with many data types, just like a molecular tumor board would. MTBBench builds a realistic cancer decision benchmark and shows current models struggle, but tool-using agents do better. Up to 9% better multimodal accuracy and 11.2% better longitudinal accuracy using tools. Current medical AI tests usually show a single question with all information given at once. In real cancer care, teams see scans, lab values, and reports over months, then adjust decisions. MTBBench copies this by giving an AI agent realistic cancer cases where new files appear over time. The agent sees only file names, must ask for what it wants, and remembers only its past reasoning. Cases mix many data types, including tissue images, blood tests, written reports, and DNA changes from sequencing. The authors also let the agent call extra tools, like pathology image models and drug or paper databases. Across many well known models, accuracy is middling and outcome or recurrence predictions are often near chance. Letting agents call these tools helps most on harder questions that need rich image and timeline context. --- Paper – arxiv. org/abs/2511.20490 Paper Title: "MTBBench: A Multimodal Sequential Clinical Decision-Making Benchmark in Oncology"

MTBBench: A Multimodal Sequential Clinical Decision-Making Benchmark in Oncology Going beyond the standard single-turn multiple choice benchmarks, this paper introduces a multimodal longitudinal agentic benchmark that simulates tumor boards, where oncologists review patient cases to make clinical decisions. Evaluates both VLMs and LLMs and the use of domain-specific foundation models as tools, which provides gains of up to 9.0% and 11.2% on multimodal and longitudinal reasoning, respectively.

Some news from the lab & collabs: - Med-PRM (medical reasoning with guideline-verified process reward models) received an 𝐨𝐫𝐚𝐥 𝐩𝐫𝐞𝐬𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧 at 𝐄𝐌𝐍𝐋𝐏! med-prm.github.io - SMMILE got accepted to 𝐍𝐞𝐮𝐫𝐈𝐏𝐒 2025 (D&B track) smmile-benchmark.github.io/ - another paper (not yet pre-printed) on benchmarking sequential & multimodal oncological decision-making with agents got accepted to 𝐍𝐞𝐮𝐫𝐈𝐏𝐒 2025 (D&B track) (1/2)

⌨️ 😇 Drafting for NeurIPS? Submit to #ICML2025 workshop on Scaling Up Intervention Models (SIM) too! Let’s enjoy some fun science in Vancouver this July. 🌞🌳 🗓️Workshop submission due on May 20 AOE 🔗More info: sites.google.com/view/sim-ic…

Excited to share that our work “Test-time view selection for multi-modal decision making” will be presented as an Oral at MLGenX, ICLR 2025 (27/04 9:15 am)!

🔥 Our workshop will take place in less than a week! In the meantime, check out the full schedule on our website and the strong lineup of invited speakers and panelists 💪 Looking forward to seeing you there!

What if AI could advance cell biology the way it did for protein folding? 🧬 In this video podcast, Head of Science @StephenQuake & EPFL’s @_bunnech join @EricTopol to discuss the virtual cell, a moonshot for digital biology. Watch: bit.ly/3FMXxq6

Researchers have developed a deep learning protein language model, ESM3, that enables programmable protein design. Learn more in this week's issue of Science: scim.ag/4b5IlQu

I’m excited to co-host the new #AIxBio conference alongside @Avsecz, @ferruz_noelia, @pdhsu, @BenLehner, @ml19, @deboramarks, @p_tingying, @CarolineUhlerat at @wellcomegenome/ @sangerinstitute! We’ll focus on modellingz and designing DNA, RNA, proteins, and cells, as well as producing data at scale for model training. Join our amazing speakers to explore the latest in #AI for biology! Register here and submit your abstract: bit.ly/4gh164E @pkoo562 @gokcen @mariabrbic

I spoke with @Nature about the transformative potential of foundation models for biology—particularly in personalized medicine and clinical diagnostics. Take a look at the technologies to watch in 2025. An exciting year ahead in AI, biology and medicine! nature.com/articles/d41586-0…

📄Read the full paper: arxiv.org/pdf/2501.06039v1 Would love to hear your thoughts! Feel free to reach out with questions.

This was a fantastic collaboration between researchers at @EPFL, the @EPFL_AI_Center, the @Unispital_USZ, and @UZH_en with co-authors @gabrigut , Andreas Wicki, Kiril Vasilev and @pariset_matteo! This is just the beginning - we’re excited of what is coming next!

🔎How? VirTues is built on a novel vision transformer architecture that efficiently processes high-dimensional data and flexible numbers of input channels while maintaining biological interpretability.

💡Why now? Spatial proteomics has just been announced ‘Method of the Year’ by @naturemethods due to its critical role in #clinical #diagnostics and #computational #pathology. VirTues is the first AI model that can handle the complexity and unique requirements of multiplex images