Narbe Mardirossian

Last week we announced a big leap forward for universal potency models with EMMI Predict: TerraBind. TerraBind delivers best-in-class potency prediction while running 26x faster and reducing inference cost by 96% relative to standard industry potency models. A key driver of this performance is our collaboration with @nvidia and the integration of the NVIDIA cuEquivariance library. By centering its computational core on triangle attention and multiplication, TerraBind's architecture is uniquely positioned to fully leverage cuEquivariance's GPU-optimized kernels, providing a 3-4x speedup compared to standard implementations. With EMMI Predict: TerraBind at the center of our Generate-Predict-Select workflow, this combination of proprietary data, purpose-built models, and NVIDIA-accelerated hardware is transforming how we discover and optimize small-molecule therapeutics. We're excited to keep pushing the boundaries of predictive chemistry with partners like NVIDIA as we work to bring better medicines to patients faster. Read our blog terraytx.com/insights/emmi-… Read our arxiv paper arxiv.org/abs/2602.07735

NEW: Terray has taken a contrarian AI approach, building a non-diffusion-based model for binding affinity predictions They shared some H2H results against Boltz-2, a popular open-source model, showing 20% more accuracy + 26x faster to run: endpoints.news/terray-says-it…

Most AI drug discovery tools can only evaluate thousands of molecules at a time. Powered by Nvidia, Terray's new model TerraBind can screen billions 26x faster and with better accuracy. Faster screening = faster cures 🚀🧬 Proud investor in @Terray_Tx terraytx.com/insights/emmi-…

Today we announced EMMI Predict: TerraBind, our universal potency model that dramatically expands Terray’s ability to move molecules to the clinic faster. Far outpacing publicly available models, TerraBind is a cornerstone of Terray’s end-to-end workflow for small molecule discovery and development. Read our blog: terraytx.com/insights/emmi-…

Awesome new results and great use of NVIDIA accelerated computing libraries for life sciences!

Incredible website: lakesofyosemite.com

@kooshiar Did you try Gemini?

Prompt: a vertical poster with tall empty sky so I can write on it an image of san francisco from above, with most famous parts seen in the picture: golden gate bridge, salesforce tower, transamerica buildig etc picture should be beutiful with happy colors and photo realistic sky should be clear and tall

Every model I tried in last 2 years failed at this until GPT5 👇

How does this have traction? What gains does anyone think this gets over GPUs to get 20x over H100s, or do they assume the audience isn't sophisticated enough to ask that question with any rigor and just sees nonsense like "burned the transformer into the chip"?

BREAKING — Senator Lee has just withdrawn his public lands sell-off provision! Thanks to your efforts, the American people — and our public lands — have won! 🏔️🏔️🏔️

Definitely a cool dataset! An Allegro-Tensormol Behaving physically and optimizing off RdKit after like 30k rows of training. After a little validation, happy to OS for anyone who like minimalist, scalable, charged UFFs.

We're Hiring. ats.rippling.com/terraytx/jobs/…

@CorinWagen @olexandr @AriWagen @RowanSci Add the closest local minimum from each docked pose to your conformer ensemble. Quick fix for now.

@olexandr @AriWagen @RowanSci hahaha yeah that always points to a flaw in the conformer search. we're working on it...

Though it has its flaws, docking is pretty magical—algorithms predict the binding sites of ligands and (attempt to) score them—all within minutes. Today, I'm excited to share that @RowanSci now supports protein–ligand docking ... and we use NNPs to detect strained poses 🧵

@owl_posting 😂 maybe they meant to write: "Validated (through binding assays) hundreds of binders that are either > 10uM or <= 5 atoms/residues different from a clinical or FDA-approved compound"; there it's fixed 😬

on god ?

I'm thrilled to announce a new preprint describing collaborative work with @prof_ajay_jain and Ann Cleves Jain, "Deep-Learning Based Docking Methods: Fair Comparisons to Conventional Docking Workflows". arxiv.org/abs/2412.02889

@AriWagen @CorinWagen @bernadsz @OrbMaterials @OpenCatalyst From a quick glance, the GFN2-xTB results don't look right. What's being shown on the main page? MAEs in kcal/mol? WTMAD-2? How can the error for GFN2-xTB be 9.54 for Inter-NCIs but none of the individual datasets are anywhere close to that value?

@CorinWagen @bernadsz @OrbMaterials @OpenCatalyst r2SCAN-3C or B3LYP would be good! I've gone ahead and added the first few reference rows to ariwagen.com/nnp-arena.html NCI performance is interesting to look at here.

Orb-v2 and Orb-d3-v2 (@OrbMaterials) and OMat24 (@OpenCatalyst) GMTKN55 benchmarks are live on NNP Arena: ariwagen.com/nnp-arena Necessary disclaimer: these models were trained on periodic systems, not molecular systems. Working to add new benchmarks soon!

Happy several team members got to be in Times Square for this.

I think @Terray_Tx is on the right path - key to success for #AI in drug discovery is to create robust training data with a high throughput platform that generates high dimensional data. Similar to what @srikosuri is doing with @OctantBio & @DaphneKoller is doing w/ @insitro.

❗️EXCLUSIVE❗️ Terray Therapeutics has raised a $120 million Series B round, aiming to get its first drug candidate into the clinic while advancing in its AI-fueled approach to making molecules. endpts.com/terray-therape…

Terray Therapeutics raises $120M Series B for its AI ways of designing molecules labn.org/los-angeles-bi… #biotech #news

A tiny chip, almost too small to see in the photo, could be the key to improving the speed, cost, and success rate of drug development. We got a firsthand look when @Terray_Tx's Jacob Berlin visited to share his vision last month. This nickel-sized chip (or microarray) is generating biological datasets 50x larger than public sources. Huge congrats to Terray on their $120M Series B funding round led by Bedford Ridge Capital and NVentures! As @endpointsnews reports, Terray is using this data to "train AI models to discover and design small molecules in a new way... placing AI's ideas at the center of its R&D process."