iiinova — where your digital self keeps evolving

Wei-Tse Hsu @weitse-hsu.bsky.social · 06/08/2026

More importantly, MDArena is designed to be easily extended and tested with new models and agents, and we are already preparing many more tasks drawn from ongoing research for future releases. Please check out the repo and stay tuned for our next release! 🚀 github.com/weitse-hsu/M...

github.com

GitHub - weitse-hsu/MDArena: MDArena is an open benchmark for evaluating AI agents on molecular dynamics (MD) workflows.

MDArena is an open benchmark for evaluating AI agents on molecular dynamics (MD) workflows. - weitse-hsu/MDArena

000

Wei-Tse Hsu @weitse-hsu.bsky.social · 06/08/2026

We evaluated 6 model–harness configurations, including GPT-5.5, Gemini 3.5 Flash, and Gemini 3.1 Pro. The strongest solved about half of the tasks. With several major frontier-model releases in recent weeks, we are now testing GPT-5.6, Claude Opus-5, and more, with updated results coming soon. 🚀

100

Wei-Tse Hsu @weitse-hsu.bsky.social · 06/08/2026

🧬 In the current release, MDArena contains 50 tasks across 29 molecular systems, spanning system preparation, parameterization, free-energy methods, enhanced sampling, and analysis. Each task runs in a reproducible container and is graded for both scientific correctness and workflow quality.

110

Reposted by Wei-Tse Hsu

Open Force Field @openforcefield.org · 05/08/2026

It's exciting to see real users testing agentic workflows that make use of our models and tools. Thanks to the Biggin lab for publicizing these results and showing the work that still needs to be done!

041

Reposted by Wei-Tse Hsu

Phil Biggin @philbiggin.bsky.social · 05/08/2026

Latest preprint where we pooled our sim experiences to create "MDArena" for AI agents trying to automate MD simulation and analysis. Great work led by Nithishwer Mouroug Anand and @weitse-hsu.bsky.social and thanks to all lab members! @oxfordbiochemistry.bsky.social arxiv.org/abs/2608.02642

arxiv.org

MDArena: Evaluating Coding Agents on Realistic Molecular Dynamics Workflows

Accelerating scientific discovery is among the most consequential applications of AI, and computational biomolecular simulation stands out as a particularly promising target within this broader effort...

0156

Reposted by Wei-Tse Hsu

Cole Group @colegroupncl.bsky.social · 27/01/2026

Now out in JACS! 🎉 : "Computing Solvation Free Energies of Small Molecules with Experimental Accuracy"! It's been a pleasure to collaborate on this with Harry Moore (@jhmchem.bsky.social) & Gábor Csányi pubs.acs.org/doi/10.1021/...

1298

Reposted by Wei-Tse Hsu

Stephanie Wankowicz @stephanieaw.bsky.social · 21/01/2026

New Preprint!! We show that binding entropy can be quantitatively predicted from crystallographic ensemble models, accounting for both protein conformational entropy and solvent entropy! www.biorxiv.org/content/10.6...

13914

Wei-Tse Hsu @weitse-hsu.bsky.social · 20/01/2026

🔎 SI highlights: - AEV-PLIG beats Boltz-2 in 4 target classes in the FEP benchmark (loses 1, ties 6); both are competitive with FEP+ in some cases. - ipLDDT & ligand pLDDT are also effective filters; pTM, PAE, PDE are not - Boltz confidence seems to generalize better than its structure module (5/6)

110

Wei-Tse Hsu @weitse-hsu.bsky.social · 20/01/2026

❓ Are co-folding predictions good enough to train scoring functions? 👉 Yes — with careful filtering. We see no performance difference b/w models trained on: - experimental structures - corresponding co-folding predictions This holds across AEV-PLIG, EHIGN, and RF-Score. (4/6)

100

Wei-Tse Hsu @weitse-hsu.bsky.social · 20/01/2026

❓ When can we trust a co-folding prediction? 👉 From reproducing HiQBind with Boltz-1x, a few simple heuristics are recommended high-quality cofolding augmentation: 1️⃣ single-chain systems 2️⃣ Boltz confidence > 0.9 3️⃣ train–test similarity > 60% (3/6)

110

Wei-Tse Hsu @weitse-hsu.bsky.social · 20/01/2026

❓ How much can data augmentation actually improve scoring? 👉 Short answer: only if the added data are high-quality. Adding BindingNet v1 clearly improved performance, but v2 did not—despite being 10x larger—due to its substantially lower quality. Quality beats quantity. (2/6)

110

Wei-Tse Hsu @weitse-hsu.bsky.social · 20/01/2026

📢 Can AI-Predicted Complexes Teach Machine Learning to Compute Drug Binding Affinity? In our recent JCIM work, we tested whether co-folding models can be used for data augmentation for training ML-based scoring functions (SFs). We asked 3 simple but critical questions. 👇 (1/6)

161