MolBioMike

I get the same thing. Boltzgen esp gives me 95% side binding VH designs. Minimal loop engagement, they’re easy to predict, but I know that’s not really going to work very well. I’ve actually created a filter that removes all antibody like sequences from my outputs because they all have great metrics but terrible binding modes. Can’t wait to hear about your results.

Sending designs to the lab for 5 projects so far Some highly curated sets from standard pipelines and some mosaic optimized sequences I have noticed a lot of interesting subtleties to the different models outputs Eg MPNNs have a serious propensity to make strange polyXmers Wt polyGLA and solu polyKE I’ve worked on changing the aa output percents to try and resemble reality Lots of supposed binders binding w/ framework instead of cdrs in the cofolded models Suggests something isn’t right to me Models with agreement build my confidence somewhat but there is still a disconnect here imo I think we are in the ~MSDOS era of this whole thing CLI based, slow to run, minimally capable, technical requirements exceeding normie hardware, etc. Generation is getting better for sure but molecular fitness is sort of an afterthought it seems Standard workflows for most people seem to be generate a shit ton of sequences and then trash most of them due to these oddities Better a priori seems important as optimizing enzymes or binders is possible but de novo is not one shot Tooling like Antifold and Mosaic seem important to the process Lots of opportunities here in this post

If you’re looking for a taq expression vector for your home lab for 20$!!! Look no further.

@karsten_kreis @ManifoldBio @viva_biotech @novonordisk @Cambridge_Uni @DukeU @LMU_Muenchen @DidiKieran @Oxer22 @ZhonglinJC @tomasgeffner @sacdallago @ArashVahdat @PierceOgdenJ @KrisDeibler @hollfelderlab @khmelinskaia @SeoulNatlUni I’m curious if you think this could be used with the Openfold3 weights instead of the AF2 weights.

@ManifoldBio @viva_biotech @novonordisk @Cambridge_Uni @DukeU @LMU_Muenchen @DidiKieran @Oxer22 @ZhonglinJC @tomasgeffner @sacdallago @ArashVahdat @PierceOgdenJ @KrisDeibler @hollfelderlab @khmelinskaia @SeoulNatlUni 🔸 Please find all details, links to papers, model weights and code, and link to the Teddymer dataset on our project page. ⭐️ Project page: research.nvidia.com/labs/genair/pr… (20/n)

📢📢 Proteina-Complexa 📢📢 Atomistic Binder Design with Generative Pretraining and Test-Time Compute + Experimental Validation at Scale ⭐️ Project page (research.nvidia.com/labs/genair/pr…) for: 📜 Method paper (ICLR 2026 Oral) 🧬 Wet lab paper 🛠️ Code & models 📁 Data 🧵 Thread (1/n)

Proteina-Complexa code base 👇

Getting these cool binding modes with these VH like domains where the IDR folds into the beta barrel and a long loop engaged too. ipSAE = 0.864

@jucap_ @paperperday @bio_rat68 Did you just print the linear fragment of dna for expression?

@MolBioMike @paperperday @bio_rat68 We do not need to purify the de novo proteins in this screen - we express them in cell-free and they produce a fluorescence signal only upon interaction with their target

CF2H: a cell-free two-hybrid platform for rapid protein binder screening "a method for detecting protein–protein interactions by fusing bait and prey proteins to a dimerization-activated DNA-binding domain that triggers transcription upon interaction." nature.com/articles/s4146…

Hit me with your favorite .pdb or .cif viewer on iOS

With all the how to guides on ML binder design I think I should write a book about how NOT to design a binder. This design is confidently incorrect. Green is RXB1 truncated target, and Orange is the heavily evolved VL with a very strange binding mode. pLDDT = 0.914 and ipSAE = 0.916

@design_proteins I see some RBX1 binders in there. 👀 Would be sweet to have something like this for mobile.

Building out an interactive plot to quickly check binder designs and check data distributions

Stoked to attend my first Boston Protein Design and Modeling club meeting. Learned about the new rival of Halo tag. Keep an eye out for LUCI-tag. Labeling a protein near you. Good job Jody Mou, I enjoyed your talk.

@jakublala Jakub, I'm curious if you think this method could work for a mini binder. I wanted to find a protein binder for RBX1 that uses the natural binding contacts and then reduces the binding partner to the most stable minimal length protein. What do you think?

k_on, and k_off are going to change dramatically. I bet there will be some impact on k_cat independent of bonding kinetics as well, kind of like a smaller set of pliers being too small to do a job. But this is an exciting approach and the findings will challenge a lot of assumptions.

Excited to share yet another case study on using BAGEL 🥯👨🏻‍🍳 for protein design – this time, we are miniaturizing enzymes! 🧬✂️ Paper 📄 doi.org/10.64898/2026.… Data 💾 zenodo.org/records/188541… Code 👨‍💻 github.com/softnanolab/ba…

One of my more promising designs for RBX1. We'll see how this pans out. I'm finding plenty of implausible binders with pretty high ipSAE. I'm not so sure it's the best metric to optimize for.

back to cranking out proteins! Here's a really nice de novo binder to RBX1 for the proteinbase.com/competitions/g…

I'm glad you noticed that! The bump in the model's confidence comes from constraining the local MSA to a smaller subset of known pdb structures and their relatives, compared to just letting the inference model search the entire sequence similarity space. More confidence, less distraction

@MolBioMike Thanks for posting Mike. That jump in iPTM after you switched MSA database is really significant. So what value is more trustworthy?

Ten Days of Plumbing open.substack.com/pub/mikeminson…

@ClaesBackman I hope you called it "Reviewer 2"

I just published a free Claude Skill that generates feedback on your own academic papers. Link in the comments below.

@LeoCK_Wan CSO. CFO maybe not enterprise scale, but definitely small scale.

Openclaw is all the hype. What are the most impactful business operations it can take over?

On the beach drinking coconuts and folding proteins from my phone!