📢 #ICML2025 Spotlight from ByteDance Research.🚀 Elucidating the Design Space of Multimodal Protein Language Models. Multimodal protein Language Models (PLMs), such as ESM3 and DPLM-2, often struggle at protein structure modeling. We uncover why and how to fix it. Our designs

🎉 DPLM-2.1 at #ICML2025!. It's happening now!.📍 Location: West Exhibition Hall B2-B3, Poster #W-115 🕓 Time: Tuesday, July 15, 4:30–7:00 p.m. PDT. Come chat with our team about protein modeling and the next generation of diffusion protein language models (DPLM). Learng more.

🎉 DPLM-2.1 at #ICML2025!. It's happening today (7/15). We’ll be presenting our DPLM-2.1 poster. Come chat with me and @zaixiang_zheng about protein modeling and the next generation of diffusion protein language models (DPLM). 📍 Location: West Exhibition Hall B2-B3, Poster.

🎉 Just arrived at #ICML2025!. We’ll be presenting our DPLM-2.1 poster on Tuesday. Come chat with me and @zaixiang_zheng about protein modeling and the next generation of diffusion protein language models (DPLM). 📍 Location: West Exhibition Hall B2-B3, Poster #W-115.🕓 Time:.

Revisit FlashAttention:. I often appreciate revisiting some classic work developed in AI, as it helped me learn many insights. Among them, FlashAttention is a nice example of how fundamental knowledge might drive AI breakthroughs. It optimizes attention by combining:.1.

"It's not hard for Zuck to poach OpenAI talent, not just because he has the money, but because open-source AI is fulfilling the original OpenAI mission.". This is brutally true.

Couldn't summarize the affinity prediction part yet, saving for later. Boltz-2 Paper: Code: Talk:

📌Performance over modalities:.Boltz-2 performs similarly to Boltz-1 on easier modalities. On harder modalities like Antibody and Protein-DNA, Boltz-2 improves with a meaningful amount even though a gap still exists to AlphaFold3.

📌Training speed.Boltz team announced collaborations with NVIDIA. They wrote the first kernels to optimize the speed of Triangle Attention and Multiplicative Layers. These kernels greatly improves the training speed and memory bottleneck.

📌Physical Quality of Poses.Issues in AF3: Chirality, Bond Distances, Stereochemistry, and Steric Clashes.Solution: Feynman-Kac steering inference-time potentials (A sampling trick for diffusion models). The high level idea is:.1. Define function to guide diffusion paths that are

📌Improvements on Capturing Local Dynamics.Boltz-2 MD ensembles generally.obtain stronger correlations with the ground truth simulation and lower errors than Boltz-1, BioEmu and AlphaFlow. A significant gap with other baselines, however, lies in the diversity (Diversity IDDT).

📌Molecular Dynamics.Motivation: Expose Boltz-2 not only to single equilibrium points from crystal structures but also to local fluctuations and global structural ensembles. Data:.- ATLAS: Protein data.- Misato: Contains some Protein-Ligand data.- mdCATH. Training: .- MD data:

📌Controllability.They integrate three new components to achieve finer controls of the model’s predictions:.1. Method conditioning: The type of structure prediction methods such as molecular dynamic. 2. Template conditioning & steering: Integrate structures of similar complexes

📌Iterate from Botlz-1 Feedbacks.They adopted some feedbacks they received from Boltz-1 when designing Boltz-2, such as:.1. Add controllability.2. Model local dynamics.3. Improve physical quality of poses.4. Optimize training speed and memory.

📌Boltz-2 Architecture:.Boltz-2’s architecture consists of: . Trunk: It functions like the transformer for proteins, encoding input sequence into representations. Denoising module: A diffusion model that takes in the representation to predict structures. Confidence module: It

📌Why Affinity Prediction?.Affinity prediction is one of the major bottlenecks in preclinical drug development, including several stages such as Target-discovery, Hit-discovery, Hit-to-lead, and Lead optimization. While we have good structure prediction models, the cost & time

📌 Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction .Boltz-2, developed by MIT researchers, is a new structural biology foundation model. While AlphaFold3 and Boltz-1 excel in structure prediction, they struggle with binding affinity. Boltz-2 aims to predict

📌Notes on Boltz-2. Just watched the video talk led by @GabriCorso.@jeremyWohlwend, and Saro Passaro that introduced Boltz-2, a structural biology foundation models. I summarized some learning notes below.🧵.

I recently discovered that NotebookLM from @Google is an incredible learning tool. I can upload a 40-minute Youtube video presentation and multiple PDFs as context. Normally, another person would need to spend 40 minutes watching the video for an in-depth discussion, but

👏 YC’s AI Startup School with Dr. Fei-Fei Li. Dr. Fei Fei Li, the godmother of AI, shared key moments from her research journey during the AI Startup School hosted by @ycombinator. Below are some stories that I found particularly interesting. 🔹 The creation of ImageNet:.In an