profile

ProteinReach by Dr. Omar

BioEMU and MDS

Published about 1 month ago • 1 min read

I recently used BioEmu for some testing. This could be a complement for a regular MD simulations pipeline but don’t believe it will replace MDS entirely.
​
BioEmu’s model pipeline is quite powerful. It combines one pretraining step with two finetuning steps, drawing on experimentally determined structures, over 200 milliseconds of MD simulation data, and protein stability measurements from in vitro testing.
​
That last piece is super important. Incorporating wet lab data significantly improves the chances of the model capturing somewhat what is happening biologically or at least what one can see in the lab.
​
Here is how I envision this playing out in practice. Use an emulator like BioEmu for initial screening across thousands of proteins or candidates of interest. Identify the most promising ones. Then verify those with full MD runs. That workflow could dramatically reduce the computational cost of in silico experiments while accelerating biotechnological applications.
​
IMPORTANT: The current scope has real limitations. BioEmu does not model protein-ligand or protein-protein interactions, struggles with larger proteins, and does not capture kinetic information. And this is a significant limitation.
​
Nonetheless, deep learning approaches like this one are moving protein modeling toward a place where scale and speed no longer are the biggest bottleneck.

Thank you,

Best,

Omar.

© ProteinReach|2026 SciLearningWorkshops LLC

​
​Unsubscribe · Preferences​

ProteinReach by Dr. Omar

Your Weekly Protein Modeling Briefing

A weekly newsletter delivering the latest protein modeling tools and strategies to streamline your in silico processes, reduce wet lab workload, and save your organization time, effort, and resources.

Read more from ProteinReach by Dr. Omar

In a recent paper with David Young, and team, we laid the foundations for a framework uncovering novel mechanisms for targeting autoimmune diseases and conditions like sepsis. Here are the three main takeaways: -Analysis of 32 placental mammal FCGR1A sequences revealed two critical binding residues, H148 and W149, as the strongest signatures of evolutionary positive selection, both at the FcγRI-IgG binding interface.-Reconstructed ancestral variants (H148P and W149R) subjected to 100 ns MD...

28 days ago • 1 min read

I get asked all the time why MD simulations need an NPT step before production.NPT stands for constant Number of particles, Pressure, and Temperature. And this is one of the most widely used isothermal-isobaric ensembles for MD equilibration. Here is what it does:1-NPT allows the box volume to fluctuate while maintaining constant pressure 2-NPT prevents artificial compression or expansion that would distort the outcome of downstream processes (e.g., binding energy)Please I would highly...

about 1 month ago • 1 min read

I get asked if AI can design a new protein from scratch. The answer used to be sort of, now it's yes.The most widely used approach is via diffusion modelsThe same underlying approach behind image generators, have been adapted to generate protein backbones from scratch. Tools like RFdiffusion take a structure prediction network and repurpose it as a generative model.Starting from noise, iteratively denoise, and you end up with a novel, physically plausible protein backbone. Then a separate...

about 1 month ago • 1 min read