iiinova — where your digital self keeps evolving

Yehlin Cho @yehlincho.bsky.social · 21/05/2026

Website: yehlincho.github.io/absolute-sta... Code: github.com/yehlincho/ab... Paper: biorxiv.org/content/10.6...

yehlincho.github.io

Accurate Protein Stability Prediction for Small Domains Using Mega-Scale Experiments

Yehlin Cho @yehlincho.bsky.social · 21/05/2026

We evaluated our models on de novo designs, nanobodies, and binder ranking tasks, and found they consistently improved discrimination of stable, well-folded proteins.

143 0

Yehlin Cho @yehlincho.bsky.social · 21/05/2026

We analyzed Augmented ESM3ΔG on ~3M sequences from 583 CATH families and found predicted stability increases with organismal optimal growth temperature, outperforming other zero-shot proxies (e.g., ProteinMPNN, AlphaFold2 pLDDT)

100 0

Yehlin Cho @yehlincho.bsky.social · 21/05/2026

We evaluated on a curated S1724 dataset from ThermoMutDB and found that Augmented ESM3ΔG performed best after removing the sigmoid output constraint, but accuracy decreased for highly stable proteins (>10 kcal/mol), highlighting the need for more high-stability training data.

100 0

Yehlin Cho @yehlincho.bsky.social · 21/05/2026

SaProtΔG and ESM3ΔG also performed well in predicting the effects of point mutations, insertions, and deletions in Megascale and MGnify domains.

100 0

Yehlin Cho @yehlincho.bsky.social · 21/05/2026

We evaluated robustness on proteins with high buried nonpolar surface area (NPSA > 3,000 Å²; 2,006 domains, 61% of the test set), where the NPSA–stability correlation drops from 0.60 to 0.36, while SaProtΔG and ESM3ΔG remain strong (r = 0.88→0.80 and 0.88→0.77).

100 0

Yehlin Cho @yehlincho.bsky.social · 21/05/2026

We fine-tuned the base models with LoRA using a single stability head and a sigmoid-corrected output layer. This addresses cDNA display limits, where values outside −1 to 5 kcal/mol are often capped or unreliable, improving predictions of stability extremes.

100 0

Yehlin Cho @yehlincho.bsky.social · 21/05/2026

We fine-tuned sequence- and structure-pretrained models on our large-scale stability data to predict absolute folding stability. The resulting SaProtΔG and ESM3ΔG achieved Spearman r = 0.88 and 0.87, with RMSE = 0.80 kcal mol⁻¹.

100 0

Yehlin Cho @yehlincho.bsky.social · 21/05/2026

We individually purified 20 domains and measured folding stability using GdnHCl chemical denaturation and the linear extrapolation method. These measurements closely matched our high-throughput results (RMSE 1.0 kcal/mol, Pearson r = 0.86)

100 0

Yehlin Cho @yehlincho.bsky.social · 21/05/2026

We measured stability for 1.8M diverse protein domains (60–80 aa) from the MGnify metagenomic database, spanning 200k+ sequence families, and created the MGnify Stability Dataset.

100 0

Yehlin Cho @yehlincho.bsky.social · 21/05/2026

🚀 Excited to share our new work: Absolute Stability Predictor! 📊: forms.gle/4ZnXZSnTBvay... Built the MGnify Stability Dataset (1.8M+ measurements) and developed stability prediction models, together with @grocklin.bsky.social @KotaroTsuboyama, @sokrypton.org and teams.

1296 0

Yehlin Cho @yehlincho.bsky.social · 13/10/2025

Exciting applications are coming soon, with experimental validation in the next version! 📄paper: www.biorxiv.org/content/10.1... 💻code: github.com/yehlincho/Pr...

biorxiv.org

230 0

Yehlin Cho @yehlincho.bsky.social · 13/10/2025

Protein Hunter enables multimer binder design, multi-motif scaffolding, partial redesign, and nucleic acid binder design — offering a general pipeline for protein design that can be applied to any AF3-style models, existing or in development.

120 0

Yehlin Cho @yehlincho.bsky.social · 13/10/2025

Additionally, Protein Hunter supports all-atom molecular binder design. We show in silico success rates for four small molecules, where iterative cycles of Boltz2 and LigandMPNN achieve the highest AF3 success rates.

100 0

Yehlin Cho @yehlincho.bsky.social · 13/10/2025

We also demonstrate the success of the pipeline on cyclic peptides, exemplified with the MDM2 target. Macrocyclic peptide design can be achieved through cyclic positional encodings.

100 0

Yehlin Cho @yehlincho.bsky.social · 13/10/2025

However, diffusion-based models favor α-helical topologies (reflecting training bias), reducing structural diversity. To enhance β-sheet content, we applied a negative helix bias to Pairformer pair features before diffusion, increasing sheet-rich samples.

100 0

Yehlin Cho @yehlincho.bsky.social · 13/10/2025

Repeating this process significantly improves the in silico success rates of AlphaFold3 and the designability of both unconditional and conditional (binder) design tasks.

100 0

Yehlin Cho @yehlincho.bsky.social · 13/10/2025

Protein Hunter: Starting from an all "X" sequence, we find that diffusion-based structure prediction models can hallucinate reasonable looking structures, which can be further improved through iterative sequence design and structure prediction, similar to AF2Cycler and LASErMPNN.

110 0

Yehlin Cho @yehlincho.bsky.social · 13/10/2025

And they do it remarkably well with an all-“X” sequence. ❌😮 AF3-style models treat unknown PDB residues as X tokens and explicitly handle non-canonical amino acids and ligands, enabling folding of undefined sequences while minimizing bias from amino acid specific features.

100 0

Yehlin Cho @yehlincho.bsky.social · 13/10/2025

It actually folds into a structure and binds near the target! We found that AF3-like structure prediction models (Boltz, Chai, AF3) can hallucinate proteins within their diffusion modules.

110 0

Yehlin Cho @yehlincho.bsky.social · 13/10/2025

Have you ever wondered what AF3-like structure prediction models would produce when given a random protein sequence and a target of your choice? Would it form a completely disordered structure that wraps around the target, or would it still fold and bind to it?

120 0

Yehlin Cho @yehlincho.bsky.social · 13/10/2025

Thrilled to announce our new preprint, “Protein Hunter: Exploiting Structure Hallucination within Diffusion for Protein Design,” in collaboration with @Griffin, @GBhardwaj8 and @sokrypton.org 🧬Code and notebooks will be released by the end of this week. 🎧Golden- Kpop Demon Hunters

35316 0

Yehlin Cho @yehlincho.bsky.social · 03/06/2025

🚀 Excited to release BoltzDesign1! ✨ Now with LogMD-based trajectory visualization. 🔗 Demo: rcsb.ai/ff9c2b1ee8 Feedback & collabs welcome! 🙌 🔗: GitHub: github.com/yehlincho/Bo... 🔗: Colab: colab.research.google.com/github/yehli... @sokrypton.org @martinpacesa.bsky.social

15417 0

Yehlin Cho @yehlincho.bsky.social · 08/04/2025

5. BoltzDesign1 can be used to design sequences and structures that AlphaFold3 predicts to bind to metal ions, nucleic acids, and other biomolecules

120 0

Yehlin Cho @yehlincho.bsky.social · 08/04/2025

4. We achieved the best results by setting the Pairformer recycling step to 0 and fixing the initial BoltzDesign1 sequence at the interface while redesigning the remaining surface regions using LigandMPNN.

100 0

Yehlin Cho @yehlincho.bsky.social · 08/04/2025

3. By utilizing only the Pairformer and Confidence module, our method generates highly diverse binders, with high AlphaFold3 success rates, strong cross-model and self-consistency, as demonstrated by benchmarks on four small-molecule targets from the RFDiffusionAA benchmark set.

100 0

Yehlin Cho @yehlincho.bsky.social · 08/04/2025

2. Instead of optimizing single structures, we optimize directly on the distogram, shaping the probability distributions of atomic distances. We show that the distogram effectively captures interactions between proteins and their targets, serving as a proxy for confidence scores

100 0

Yehlin Cho @yehlincho.bsky.social · 08/04/2025

1. We introduce BoltzDesign1, which inverts the Boltz-1 model—an open-source reproduction of AlphaFold3—to enable the design of protein binders for diverse molecular targets without requiring model fine-tuning.

100 0

Yehlin Cho @yehlincho.bsky.social · 08/04/2025

Excited to share our preprint “BoltzDesign1: Inverting All-Atom Structure Prediction Model for Generalized Biomolecular Binder Design” — a collaboration with @martinpacesa.bsky.social, @Zhidian Zhang, @Bruno E. Correia, and @sokrypton.org 🧬 Code will be released in a couple weeks

56117 0