NSF Invests Nearly $32M to Accelerate Novel AI-Driven Approaches in Protein Design Aug. 8, 2025 -- The U.S. National Science Foundation Directorate for Technology, Innovation and Partnerships (NSF TIP) announced an inaugural investment

NSF Invests Nearly $32M to Accelerate Novel AI-Driven Approaches in Protein Design
ow.ly/NJzp50WCc00 #NSF #AIwire

Worth a watch:

Head of Signal, Meredith Whittaker, on so-called "agentic AI" and the difference between how it's described in the marketing and what access and control it would actually require to work as advertised.

Benchmarking protein structure predictors to assist machine learning-guided peptide discovery Machine learning models provide an informed and efficient strategy to create novel peptide and protein sequences with the desired profiles. Nevertheless, they are primarily trained on sequences where ...

If that helps, we compared AF2 predicted structures of a small set of peptides against their PDB structures under different experimental methods (NMR solutions, X-ray, SSNMR, EM) as ground truth. pubs.rsc.org/en/content/a...

Think coronavirus spikes have run out of surprises? Think again.

Our latest preprint dives into the highly unusual spikes of marine mammal coronaviruses.

www.biorxiv.org/content/10.1...

This #cryoEM study was led by @viralfusion.bsky.social, with key contributions from an amazing team.

Derguini, F.; Plisson, F.; Massiot, G. 𝘗𝘳𝘦𝘱𝘢𝘳𝘢𝘵𝘪𝘰𝘯 𝘰𝘧 𝘵𝘢𝘨𝘪𝘵𝘪𝘯𝘪𝘯 𝘊 𝘢𝘯𝘥 𝘍 𝘥𝘦𝘳𝘪𝘷𝘢𝘵𝘪𝘷𝘦𝘴 𝘢𝘴 𝘢𝘯𝘵𝘪-𝘤𝘢𝘯𝘤𝘦𝘳 𝘢𝘨𝘦𝘯𝘵𝘴. FR 2941697 A1 20100806 2010.
Europe PMC europepmc.org/article/PAT/...
Google Patents
patents.google.com/patent/FR294...

As an engineer at heart, I wondered how we would scale up if we found a non-toxic, bioactive analogue. That led us to 𝗯𝗶𝗼𝗰𝗮𝘁𝗮𝗹𝘆𝘀𝗶𝘀 (think immobilised enzymes) - screening lipases and esterases to reach, in 1 step, a key intermediate: Taginitol C. This work was patented.

Over three years, we synthesised dozens of Taginitin C analogues, eventually streamlining the synthesis from 10 to 6 steps 𝘷𝘪𝘢 judicious protection/deprotection strategies.

The molecule was notoriously sensitive to acids, bases, nucleophiles, and UV light. Resources were limited, and we were performing multi-step syntheses on just 𝟱 𝗺𝗴 𝘀𝗰𝗮𝗹𝗲𝘀, in 𝘁𝗶𝗻𝘆 𝗿𝗼𝘂𝗻𝗱-𝗯𝗼𝘁𝘁𝗼𝗺 𝗳𝗹𝗮𝘀𝗸𝘀, working in darkened fume hoods.

Together with various teammates (MSc students) we worked on the hit-lead optimisation of Tagitinin C - a toxic germanocrolide inhibiting the ubiquitin-proteasome pathway.

Eighteen years ago, I joined one of my first drug development R&D programs at the 𝘐𝘯𝘴𝘵𝘪𝘵𝘶𝘵 𝘥𝘦𝘴 𝘚𝘤𝘪𝘦𝘯𝘤𝘦𝘴 𝘦𝘵 𝘛𝘦𝘤𝘩𝘯𝘰𝘭𝘰𝘨𝘪𝘦𝘴 𝘥𝘶 𝘔𝘦𝘥𝘪𝘤𝘢𝘮𝘦𝘯𝘵 𝘥𝘦 𝘛𝘰𝘶𝘭𝘰𝘶𝘴𝘦 - a public-private partnership between the @cnrs.fr and Pierre Fabre Laboratories.

Tagitinin C, a Sesquiterpene Lactone, and Derivatives as Proteasome Inhibitors Tagitinin C, a germacranolide, isolated from Tithonia diversifolia was shown to have an interesting level of activity on the proteasome pathway. It is however a particularly unstable molecule, sensit...

Never too late to publish, I am proud to see this work finally published: Tagitinin C, a Sesquiterpene Lactone, and Derivatives as Proteasome Inhibitors doi.org/10.1002/ejoc...

Investigating the determinants of performance in machine learning for protein fitness prediction Machine learning (ML) has revolutionized protein biology, solving long-standing problems in protein folding, scaffold generation and function design tasks. A range of architectures have shown success ...

Choosing ML architectures for protein engineering is often challenging. Our “new” updated preprint provides a rational framework to match ML models to protein fitness tasks, showing landscape ruggedness influences prediction accuracy. Mahakaran dana Adam et al www.biorxiv.org/content/10.1...

Google Colab

Run BioEmu in Colab - just click "Runtime → Run all"! Our notebook uses ColabFold to generate MSAs, BioEmu to predict trajectories, and Foldseek to cluster conformations.
Thanks @jjimenezluna.bsky.social for the help!
🌐 colab.research.google.com/github/sokry...
📄 www.biorxiv.org/content/10.1...

Protein function often depends on protein dynamics. To design proteins that function like natural ones, how do we predict their dynamics?

@hkws.bsky.social and I are thrilled to share the first big, experimental datasets on protein dynamics and our new model: Dyna-1!

🧵

PDB101: Irina Bezsonova Gallery PDB-101: Training, Outreach, and Education portal of RCSB PDB

You can download my protein structure-inspired artwork from pdb webpage:

pdb101.rcsb.org/sci-art/bezs...

@rcsbpdb.bsky.social
@pdbeurope.bsky.social
#sciart

YouTube is the world's 2nd-largest search engine. So why aren't more conference keynotes and presentations there? 🤔

INTRODUCTION

Modern-Day Oracles or Bullshit Machines?

Jevin West (@jevinwest.bsky.social) and I have spent the last eight months developing the course on large language models (LLMs) that we think every college freshman needs to take.

thebullshitmachines.com

Fig. 1 Inverse Folding Molecular Dynamics (IF-MD) protocol. (A) Free energy landscape of the process of protein-protein dissociation in the case of SARS-CoV-2 RBD (red) in complex with the nanobody H11 (blue) visiting different metastable states along the unbinding mechanism along two collective variables. Two different unbinding trajectories from the bound state are shown in black and red. (B) IF-MD protocol to sample protein sequence space by constraining k ensemble averaged observables q target k and using Bayesian Optimization to propose new sequences. (C) Specific example of IF-MD constrained to a target the unbinding rate constant, and using Infrequent Metadynamics to sample the prior ensemble. Unbinding example trajectories sampled by Infrequent Metadynamics are shon in blue, along the RMSD from the cryo-EM structure of H11 bound to SARS-CoV-2 RBD as a function of time.

Short thread about this interesting preprint that explores antibody design by combining MD with inverse folding and active learning. It's a bit rough around the edges but it introduces a cool idea I hope is further fleshed out www.biorxiv.org/content/10.1...

Are protein language models the universal key?
doi.org/10.1016/j.sb...

Brilliant and thoughtful piece.

What if all the world's biggest problems have the same solution? YouTube video by Veritasium

Brilliant video on the development of protein structure prediction featuring #AF2 and #Rosetta series
youtu.be/P_fHJIYENdI?...

Microproteins: emerging roles as antibiotics Recent advances in computational prediction and experimental techniques have detected previously unknown microproteins, particularly in the human microbiome. These small proteins, produced by diverse ...

(1/4)In our new paper in @cp-trendsgenetics.bsky.social @cellpress.bsky.social, we highlight some fascinating tiny proteins—microproteins—which appear to be widespread in nature. www.cell.com/trends/genet...

Real-World Applications and Experiences of AI/ML Deployment for Drug Discovery OR SEARCH CITATIONS

Great editorial from @evotec.bsky.social using AI+ML tools to assist drug discovery campaigns in the small and middle chemical spaces
pubs.acs.org/doi/10.1021/...

Explainable deep learning and virtual evolution identifies antimicrobial peptides with activity against multidrug-resistant human pathogens - Nature Microbiology An AI-based learning model is applied to low-abundance human oral bacteria and identifies antimicrobial peptides with efficacy against multidrug-resistant bacterial pathogens.

Beilun Wang and team recently reported the development of antimicrobial peptides combining an ensemble of DL predictors with EvoGradient, modifying candidate sequences strategically from evolutionary information, improving their antimicrobial activity and selectivity.
www.nature.com/articles/s41...

Machine learning-guided discovery and design of non-hemolytic peptides - Scientific Reports Scientific Reports - Machine learning-guided discovery and design of non-hemolytic peptides

We previously used that strategy to set guidelines for discovering and designing non-hemolytic peptides:
www.nature.com/articles/s41...

The process mimics the hypothesis-driven process of a medicinal chemist, especially with Deep Mutational Scanning. It also merits explaining the results from your predictive model(s) by comparing mutants.

Another strategy starts from sequences identified with the best predictive scores and applies a generative algorithm to create analogues (via random mutagenesis or directed evolution), followed by predictions.

Leveraging large language models for peptide antibiotic design Large language models (LLMs) have significantly impacted various domains of our society, including recent applications in complex fields such as biolo…

A common strategy for identifying sequences with the desired characteristics is to combine a generative algorithm with multiple models predicting properties and functions. This narrows the search in the vast peptide fitness space.
www.sciencedirect.com/science/arti...

Accelerating the Discovery and Design of Antimicrobial Peptides with Artificial Intelligence Peptides modulate many processes of human physiology targeting ion channels, protein receptors, or enzymes. They represent valuable starting points for the development of new biologics against communi...

Generative models have created peptide sequences that share antimicrobial characteristics (e.g., amphipathic character, charged residues, helical constraint).
link.springer.com/protocol/10....

These models are routinely used to discover novel AMPs from biodiverse extant and extinct species. Check out the great work done by @delafuentelab.bsky.social team with their APEX models.

Closer to my experience in #CompBiology - the development of AI algorithms to discover and design antimicrobial peptides against #AMR, where predictive ML models are employed to predict the antimicrobial nature (classification) or activity (regression), primarily from sequences.