Shukla Group

There is no amount of lead that is safe in drinking water? Yet, off-the-shelf lead diagnostics in drinking water are not sensitive enough by orders of magnitude. Brenda Wang, @ShuklaGroup @LucksLab @akarimlab addressed this issue with synthetic biology. nature.com/articles/s4146…

How are cell-free systems accelerating biological design? We worked with @UWproteindesign to computationally design soluble MHC proteins. This holds promise to enable new immunotherapies. pnas.org/doi/10.1073/pn…

Excited to share our latest work to close out 2025: Dynamic ASK1 proximity networks uncover SCF-dependent and noncanonical roles in ABA and drought adaptation🌱 With the Walley lab, we map how ASK1/SKP1 dynamically rewires protein networks during stress. biorxiv.org/content/10.648…

Exciting holiday present for the Gupta lab! 🎉 🎊 Our paper "Molecular mechanism of substrate transport by human peroxisomal ABCD3" is now published in Proceedings of the National Academy of Sciences!! 🎉 🎊 @PNASNews pnas.org/doi/10.1073/pn…

Now published in @NatureComms. Congratulations to all the authors! nature.com/articles/s4146…

Really happy to have taken part as a Binding Illini in @ShuklaGroup!

1/ 🔬 Thrilled to share our new preprint: ESMDynamic: Fast and Accurate Prediction of Protein Dynamic Contact Maps from Single Sequences 📄 Now on bioRxiv: biorxiv.org/content/10.110…

ESMDynamic: Fast and Accurate Prediction of Protein Dynamic Contact Maps from Single Sequences 1. A novel study introduces ESMDynamic, a deep learning model that predicts dynamic residue-residue contact probability maps directly from protein sequences. This innovation leapfrogs traditional methods by capturing the conformational dynamics of proteins without needing multiple sequence alignments or extensive computational resources. 2. ESMDynamic is built on the ESMFold architecture and trained on contact fluctuations from experimental structure ensembles and molecular dynamics (MD) simulations. This unique training approach enables the model to capture diverse modes of structural variability, making it highly accurate and efficient in predicting dynamic contacts. 3. The model’s performance was benchmarked on two large-scale MD datasets (mdCATH and ATLAS), where it matched or outperformed state-of-the-art ensemble prediction models like AlphaFlow, ESMFlow, and BioEmu. Notably, ESMDynamic achieved this while offering significantly faster inference times, making it a powerful tool for rapid dynamic property predictions. 4. ESMDynamic’s ability to generalize to unseen systems was demonstrated through applications on the ASCT2 and SWEET2b transporters, a de novo troponin C design, and the HIV-1 protease homodimer. The model successfully recovered experimentally validated dynamic contacts, showcasing its broad applicability in protein engineering and functional analysis. 5. An automated pipeline using ESMDynamic predictions was presented for selecting collective variables for Markov State Model construction. This pipeline produced high-quality kinetic models from unbiased MD simulations of SWEET2b, highlighting the model’s potential to streamline complex analyses and lower the barrier to incorporating dynamics into protein modeling pipelines. 6. The study also revealed ESMDynamic’s emergent ability to predict dynamic protein interfaces, as demonstrated in the HIV-1 protease homodimer. Despite not being explicitly trained on multimeric proteins, the model accurately captured dynamic contacts at the dimer interface, suggesting broader applicability in multimeric systems. 7. The authors acknowledge support from the National Science Foundation Early CAREER Award (NSF MCB-1845606) and the National Institutes of Health (Award No. R35GM142745). The study was led by Diwakar Shukla and his team at the University of Illinois at Urbana-Champaign, with contributions from Diego E. Kleiman, Jiangyan Feng, and Zhengyuan Xue. 📜Paper: biorxiv.org/content/10.110… #ProteinDynamics #DeepLearning #StructuralBiology #ProteinEngineering #MolecularDynamics #Bioinformatics

How can we use biological datasets to accelerate design for synthetic biology applications? We developed an ML-driven, multi-objective framework—powered by augmented directional tokens—to engineer proteins for specificity and selectivity. w/@ShuklaGroup biorxiv.org/content/10.110…

With a new five-year, $15M award, the @NSF has renewed its support of the @Molecule_Maker Lab Institute, which develops AI tools for discovery and synthesis of molecules for uses in medicine, energy, industry and more. chbe.illinois.edu/news/stories/n…

Two for the Price of One: Integrating Large Language Models to Learn Biophysical Interactions 1. This study presents a framework to combine large language models (LLMs) to learn biophysical interactions, particularly focusing on binding affinity and specificity between molecular complexes. 2. The authors highlight the limitations of current LLMs, which are typically unimodal and lack explicit representations of intermolecular interactions, especially for tasks like protein-ligand, protein-peptide, and protein-nucleic acid interactions. 3. They propose the idea of "composing" representations from distinct LLMs to effectively capture complex biological interactions. This is achieved by integrating language models trained on different data modalities via concatenation, contrastive learning, attention mechanisms, and composition. 4. Contrastive learning is emphasized as a fast and effective technique to align embeddings from different models by maximizing the similarity between interacting pairs while minimizing similarity with non-interacting pairs. 5. Attention mechanisms, particularly cross-attention, are highlighted for their ability to dynamically focus on relevant features of interacting molecules, improving the prediction of binding affinities and specificities. 6. The most promising approach described is "composition," which involves merging language models via joint training tasks. This method retains the representational power of individual models while improving the overall performance on tasks that require multimodal information. 7. The authors demonstrate the effectiveness of these approaches by comparing various multimodal embedding strategies on peptide-protein binding prediction tasks, showing that cross-attention and composition outperform simpler methods like concatenation. 8. They also explore interpretability and efficiency challenges associated with large models, suggesting that parameter-efficient fine-tuning (PEFT) and sparse autoencoders could improve transparency and training efficiency. 9. The framework offers new possibilities for developing interaction-aware LLMs applicable to diverse biological tasks, including drug discovery, protein-ligand docking, and biomolecular interaction prediction. 10. Future work will focus on refining composition methods to improve interpretability and computational efficiency, while also expanding their applicability to broader biochemical interactions. @diwakarshukla @TannerDean19 @ShuklaGroup 📜Paper: arxiv.org/abs/2503.21017 #LargeLanguageModels #ProteinLigandInteractions #DeepLearning #Bioinformatics #ContrastiveLearning #CrossAttention #MachineLearning #DrugDiscovery

Congratulations to professors Guironnet, Rogers and Shukla who have been named James W. Westwater professorial scholars in recognition of their excellence in teaching and research! Learn more: chbe.illinois.edu/news/stories/G…

A cannabinoid bitopic, biased and peripheral agonist analgesic with limited tolerance and tetrad in Nature. Work led by @Vipin36590264 with @robgereau @DrorLab kaavya Krishna Kumar @WashUClinPharm @WashUanesthesia A short thread #citeas" target="_blank" rel="nofollow noopener">nature.com/articles/s4158…

Love the beautiful illustration🥰! Our work published @ScienceAdvances , featured by Daily Texan News! 私達の最新の研究成果がローカル大学新聞にて取り上げられました🤠🤘🌱🦠 "UT researchers discover communication between growth, immunity pathways in plants" thedailytexan.com/2025/02/19/ut-…

Congratulations to CHBE Professor Ying Diao, one of three Illinois researchers who have received Presidential Early Career Awards for Scientists and Engineers.

.@shuklagroup & @diwakarshukla and colleagues report the transport mechanism of an alanine-alanine peptide through the PepTSo transporter by MD simulation nature.com/articles/s4200…

Congratulations to Joseph Clark, undergraduate researcher in our lab and his mentor, Xuenan Mi for publication of their paper on peptide language model, PeptideESM! @MCB_Illinois @diwakarshukla

Congratulations to the newest doctor from our group, Dr. Xuenan Mi! She will join @JNJNews as a research scientist in January! @ChBEIllinois @diwakarshukla

@agr_chemistry @SusannaBar2 @nchembio Thank you for writing this commentary!

congratulations to @SusannaBar2 and the team @ShuklaGroup. read the original @nchembio study here: nature.com/articles/s4158…

It was our pleasure (@SwansonGroup_UT) to highlight the latest on lasso peptide folding and substrate–cyclase tolerance from the Shukla, Burk, and Mitchell teams at @ChemistryUIUC @ChBEIllinois @lassogen: nature.com/articles/s4158…