Gian Marco Visani

check out our new manuscript (led by @GMarcoVisani) structure-based machine learning model for TCR-pMHC complexes, predicting T-cell affinity to peptide-MHC complexes, quantifying T-cell receptor specificity, and designing de-novo immunogenic peptides: arxiv.org/abs/2503.00648

HERMES: Holographic Equivariant neuRal network model for Mutational Effect and Stability prediction 1. HERMES introduces a cutting-edge 3D rotationally equivariant neural network model to predict the effects of mutations on protein stability and binding, significantly improving performance compared to previous models. 2. A key feature of HERMES is its pre-training on atomic 3D structures to predict amino acid identities, which enhances accuracy when fine-tuned to predict mutation effects. 3. HERMES offers substantial improvements in speed (2.75× faster) over its predecessor H-CNN while retaining high accuracy in mutation effect predictions, making it a practical choice for large datasets. 4. The model’s architecture respects permutation anti-symmetry in mutations, ensuring that the effect of reversing a mutation is predicted as the negative of the forward mutation, a feature that most models achieve through data augmentation. 5. When tested on experimental and computational protein structures, HERMES outperformed state-of-the-art models, including Stability-Oracle and ProteinMPNN, in predicting protein stability and binding effects. 6. HERMES was rigorously benchmarked across multiple datasets, showing excellent performance even when using computationally predicted protein structures, such as those from ESMFold, for both training and testing. 7. The paper highlights HERMES' ability to predict not only stability effects but also the binding affinity changes of mutations using the SKEMPI dataset, showing high correlation with experimental data. 8. HERMES is open-source, enabling users to fine-tune the model on their specific datasets for customized mutation effect predictions, potentially accelerating discoveries in protein engineering and disease-related mutation analysis. @TheArmita @GMarcoVisani 💻Code: github.com/StatPhysBio/he… 📜Paper: doi.org/10.1101/2024.0…

Check out HERMES, our equivariant structure-based suit of models to predict stability and fitness effect of mutations, led by @GMarcoVisani arxiv.org/abs/2407.06703 bonus: fine-tune your own protein models. Google Colab: colab.research.google.com/drive/1JQxkXeG… Github: github.com/StatPhysBio/pr…

Holographic-(V)AE: An end-to-end SO(3)-equivariant (variational) autoencoder in Fourier space, Gian Marco Visani, Michael N. Pun, Arman Angaji, and Armita Nourmohammad @TheArmita @GMarcoVisani @UW @uwcse @physicsuw #NeuralNetworks go.aps.org/4cyPTMd

@Oxer22 Great work! Check out our pre-trained residue-level structure embedding if you’d like to try an alternative complementary to PLM embeddings 🙂

New SOTA results on protein function prediction with simple ideas😎! We show simple combination of protein language models and recent protein structure encoders can yield large imporvements. This can be further enhanced by structure-based pre-training. arxiv.org/abs/2303.06275

"Holographic-(V)AE: an end-to-end SO(3)-Equivariant (Variational) Autoencoder in Fourier Space" has been updated Favorable performance on predicting affinity from structures of protein-ligand complexes biorxiv.org/content/10.110…

👀 List of accepted papers at the Machine Learning for Drug Discovery workshop at #ICLR2023 Lots of high quality works! sites.google.com/view/mldd-2023…

Should universities provide chatgpt plus to all of their students to level the playing field?

I think we can call it shut on 'Open' AI: the 98 page paper introducing GPT-4 proudly declares that they're disclosing *nothing* about the contents of their training set.

@biorxiv_bioinfo Any fellow Italian giggling at this? 🤣

PROSTATA: Protein Stability Assessment using Transformers biorxiv.org/cgi/content/sh… #biorxiv_bioinfo

Interested in graduate school in CS? Consider applying to the Tufts PhD #computerscience program! I am looking for some new #phdstudents for Fall 2023 to work in funded #phdpositions. My lab works at the intersection of #mach…lnkd.in/ekJ_EKQv lnkd.in/e2WKTGsB

A post @workshopmlsb presentation thread, for those interested in using Protein Language Models (PLMs) for diversification: the seq2MSA model can diversify arbitrary protein sequences, including sampling functional insertions and deletions (1/5)

In 2020, the AI program #AlphaFold @DeepMind made headlines for predicting protein structures. A study suggests that it could also help us understand how these structures form. Read the commentary by @TheArmita Michael Pun @physicsuw, @GMarcoVisani @uwcse. physics.aps.org/articles/v15/1…

Machine Learning in Comp Bio (MLCB) 2022 schedule is available now: mlcb.github.io Please register (it's free) to receive information about (virtually) watching the talks.

Check out our new preprint (led by Mike Pun @UW) on 3D rotationally equivariant model of protein structure micro-environments, with a holographic CNN (H-CNN): doi.org/10.1101/2022.1…

Happy to share that our work on rotation-equivariant representation learning for spherical and 3D data has been accepted at @workshopmlsb! Excited to come discuss how symmetry-aware DL may help us characterize protein function from structure. Preprint: doi.org/10.1101/2022.0…

Holograms+ SO(3) equivariance+ latent representations + ML= Holographic VAE (H-VAE) excited to share our work (led by @GMarcoVisani) on representation learning while respecting rotational symmetry many applications to dream about doi.org/10.1101/2022.0… #GeometricDeepLearning

@colvieux è vero che @wooljumper ha esagerato, però pensa anche che forse quella era la verità...

@LucaGiangre non ho detto questo. Ma non si può dare del "bravo ragazzo" ad un delinquente morto!