Alejandro Lozano

🧬 What if we could build a virtual cell to predict how it responds to drugs or genetic perturbations? Super excited to introduce CellFlux at #ICML2025 — an image generative model that simulates cellular morphological changes from microscopy images. yuhui-zh15.github.io/CellFlux/ 💡 Key Innovation: We frame perturbation prediction as a distribution-to-distribution learning problem, mapping control cells to perturbed cells within the same batch to mitigate biological batch artifacts, and solve it using flow matching. 📊 Results: ✅ 35% higher image fidelity ✅ 12% greater biological accuracy ✅ New capabilities: batch effect correction & trajectory modeling 🎯 Impact: Paving the way for faster drug discovery and deeper biological insights. 📍 Poster: Tue 7/15, 11:00 AM – 1:30 PM, #301 🙏Huge thanks to all my brilliant collaborators: @hhhhh2033528 (co-lead; now Harvard AIM PhD! 🎉) @ChenyuW64562111 @TianhongLi6 @zoewefers @jnirsch @jmhb0 @DaisyYDing @Ale9806_ @Prof_Lundberg @yeung_levy! #MachineLearning #ComputerVision #ComputationalBiology #DrugDiscovery #AIforScience #CVforScience

@jmhb0 Damn two papers? Congrats!

I'm at CVPR! Come see me at one of my posters, or reach out for a chit chat. MicroVQA: reasoning llm benchmark in biology Sat 5-7pm, hall D, poster #357 jmhb0.github.io/microvqa/ BIOMEDICA: a massive vision-language dataset Sat 5-7pm, hall D, poster #374 minwoosun.github.io/biomedica-webs…

@AkliluJosiah2 Great work @AkliluJosiah2 and Team!

There’s growing excitement around VLMs and their potential to transform surgery🏥—but where exactly are we on the path to AI-assisted surgical procedures? In our latest work, we systematically evaluated leading VLMs across major surgical tasks where AI is gaining traction..🧵

@xhluca Not only useful but necessary! Thanks for making such an amazing piece of code.

@Ale9806_ Congratulations! Glad to know BM25S was useful for designing the index :)

Earlier this year, we released the BIOMEDICA dataset, featuring 24 million unique image caption pairs and 30 million image references derived from open-source biomedical literature. It's been great to see the community engaging with it—we're currently seeing around 6K downloads per month. I'm excited to share our latest update, showcasing the new additions to the BIOMEDICA framework: 📄 Brief Announcement: lnkd.in/gVjgT8me 🗄️ BIOMEDICA Index:lnkd.in/gxD4YwDa 🌐 Website: lnkd.in/gGrbZaRu Contributions: 🗄️ The BIOMEDICA Index: A hybrid vector-BM25 database capable of querying similar images, captions, and full-text articles with language and image queries! 🎯 BMC-SmolVLM: A small yet powerful biomedical VLM. With only 2.2B parameters, BMC-SmolVLM achieves performance comparable to biomedical VLMs with 7–13B parameters. 🤖 BMC-Agent: By leveraging the BIOMEDICA Index, BMC-Agent can retrieve and analyze relevant articles given a biomedical query. We demonstrate that this plug-and-play approach enables LLMs to answer questions based on newly modified medical knowledge—even when the base models initially get the answer wrong. Lastly, as part of our release, we’re introducing the BIOMEDICA Subsets— a collection of pre-filtered datasets created in response to community requests.

Shout out to my stellar first co-authors @minwsun and @jmhb0 for leading this effort, as well as the incredible team of computer scientists, statisticians,  biologists, and clinicians that made this possible: @jnirsch Christopher Polzak, @Zhang_Yu_hui, @cliangyu_, Jeffrey Gu, Ivan Lopez, @AkliluJosiah2, Anita Rau, Austin Katzer, Collin Chiu, @orr_zohar, @XiaohanWang96, Alfred Song, Chia-Chun Chiang, @robtibshirani , and @yeung_levy

@Himanshu_nitrr Valid question! Given the broad space in biomedical research (EHR-derived datasets, scientific-derived datasets, human-aggregated metadata), we specifically compare our work to scalable methods derived from scientific literature!"

@Ale9806_ Thanks for releasing the dataset, just wondering why there's no mention of MedTrinity-25M (arxiv.org/pdf/2408.02900)

Biomedical datasets are often confined to specific domains, missing valuable insights from adjacent fields. To bridge this gap, we present BIOMEDICA: an open-source framework to extract and serialize PMC-OA. 📄Paper: lnkd.in/dUUgA6rR 🌐Website: lnkd.in/dnqZZW4M

@myeong_official :) more on this very very soon!

@Ale9806_ This is a great step forward! How does the integration of different metadata types (captions, figure references, etc.) enhance the overall usefulness of the dataset for training models?

Introducing video differencing, a new task for detecting differences between video frames. Notably, even the most advanced Video LLMs struggle with this challenge, underscoring the long road ahead!

Which LLMs work best for medical queries? 🩺✨ Introducing MedArena 🏥—the first chatbot arena just for clinicians ⚕️! 👩‍⚕️👨‍⚕️ Licensed healthcare professionals can submit medical queries and compare + rank responses from two randomly selected LLMs (such as o1, Gemini 🌟, and Perplexity 🤖) 🏆. 💡 Your participation will be a valuable contribution to an open-source dataset of real-world queries and preferences for medicine 📊🔍. 🔗 Visit medarena.ai to try it out! 🚀 @james_y_zou @ericwu93 #MedTwitter 💬 #MedTech 🧬 #DigitalHealth 🌐 #MedicalAI 🤝

Check out our new work accepted to ICLR 2025. We introduce time-to-event (TTE) pretraining to leverage temporal supervision from longitudinal EHR data and estimate the risk of future events. By scaling to 225M clinical events, we achieve SOTA prognostic performance!

🚀 Introducing Temporal Preference Optimization (TPO) – a video-centric post-training framework that enhances temporal grounding in long-form videos for Video-LMMs! 🎥✨ 🔍 Key Highlights: ✅ Self-improvement via preference learning – Models learn to differentiate well-grounded from inaccurate responses without manual annotations. ✅ Multi-level temporal grounding – Effectively captures both localized segments and comprehensive video sequences. ✅ Efficient and scalable – Utilizes only 10k video QA pairs for post-training and can scale seamlessly to larger datasets. ✅ Proven performance – TPO enhances two state-of-the-art Video-LMMs on LongVideoBench, MLVU, and Video-MME. 🔥 LLaVA-Video-TPO is now the top-performing 7B model on Video-MME, highlighting TPO's potential in advancing temporal reasoning. This work was co-led with the talented undergraduate @rui__li, alongside fantastic collaborators @Zhang_Yu_hui and Zeyu Wang, and advised by @yeung_levy!

Work done at @StanfordAILab

[10/10 @cliangyu_,@jnirsch, Jeffrey Gu, Ivan Lopez, @AkliluJosiah2, Austin Katzer, Collin Chiu, Anita Rau, @XiaohanWang96,@Zhang_Yu_hui, Alfred Song, @robtibshirani,@yeung_levy

[9/10] Shout out to my stellar first co-authors @minwsun and @jmhb0 for leading this effort, as well as the incredible team of computer scientists, statisticians, biologists, and clinicians that made this possible:

[8/10] While our models offer state-of-the-art performance, all evaluations indicate that there is still significant room for improvement. We release all our contributions under a permissive license to facilitate broader use and further development

[7/10] 💡 We demonstrate the utility and accessibility of our resource by training BMC-CLIP, a suite of CLIP-style models continuously pre-trained on our dataset using different training recipes via streaming.

[6/10] 🎯 We demonstrate the utility and accessibility of our resource by training BMC-CLIP, a suite of CLIP-style models continuously pre-trained on our dataset using different training recipes via streaming.

[5/10] 💵 Our archive is hosted in HuggingFace, enabling streaming. Eliminating the need to download 3.9 TB of data locally in order to use BIOMEDICA.