google scholar 👉 x scholar wenhaochai.com/pages/x-schol…

#SpliceAI2 splice variant effect benchmarks. Considerably improved performance in 9 evaluations spanning cryptic splicing, sQTLs, MPRAs, population genetics, plasma proteins, and rare disease. 🧬✂️🧵

NSF supports the Presidential SI Challenge, inspiring America’s students and educators to lead the next wave of innovation. By empowering K–12 schools to build SI readiness, we’re investing in the nation’s future STEM talent and scientific leadership. Learn more at SI.gov.

Why the big dropoff in trust?

BREAKING: Nobel laureate David Baker’s team just published TWO Nature papers in one day! AI-designed proteases from scratch: 🧬 Cysteine proteases: 19% success rate, functional in mammalian cells. 🧬 Metalloproteases: 36% success rate, with over 10 BILLION-fold catalytic acceleration vs. the uncatalyzed reaction. From predicting protein structures to designing catalytic function. 📄 Nature papers: nature.com/articles/s41586-0… nature.com/articles/s41586-0…

PeTFlow: Flow-Matching Based All-Atom Peptide Design Through Side-Chain State Tokenization 1. PeTFlow frames pocket-conditioned peptide binder design as a coupled generation problem over backbone geometry, amino-acid identity, and explicit all-atom side-chain packing—using discrete “residue-state tokens” to unify sequence and conformation in one representation. 2. Core idea: a VQ-VAE learns a 1024-token vocabulary where each token encodes a chemically compatible pairing of residue type and side-chain conformational state, avoiding residue-dependent torsion parameterizations and masks while still enabling explicit atom-level outputs. 3. Architecture has three modules: (i) side-chain state tokenizer (VQ-VAE), (ii) Backbone–VQ Generator that jointly samples peptide residue frames and residue-state tokens under pocket conditioning, and (iii) a VQ-conditioned side-chain generator that decodes full side-chain coordinates while preserving the generated backbone. 4. The Backbone–VQ Generator uses geometry-aware Invariant Point Attention (IPA) to predict clean residue frames (rotation + translation) and a categorical distribution over VQ tokens, enabling simultaneous modeling of backbone conformation and side-chain state selection in a flow-matching generative framework. 5. Training strategy is staged and data-scaled: the VQ-VAE is trained on CATH 4.4.0 monomers to cover diverse side-chain conformations; the generators are pretrained on 67,535 peptide-like protein fragment complexes, then fine-tuned on 10,184 peptide–protein complex structures for binder design. 6. Tokenizer quality: on the CATH monomer test set, the learned codebook shows 99.5% utilization (1019/1024 tokens) with side-chain dihedral reconstruction MSE of 0.0925, supporting the claim that discrete tokens retain meaningful conformational information. 7. Benchmarking on 221 test complexes (40 candidates per target) compares against PepFlow, PepGLAD, and general protein design frameworks RFdiffusion and BoltzGen. PeTFlow improves native-likeness metrics vs peptide-specific baselines and remains competitive against larger generalist models. 8. Reported aggregate results highlight better structural/interface recovery: vs PepFlow, PeTFlow reduces Cα RMSD from 4.437 Å to 3.213 Å and increases DockQ from 0.379 to 0.424; it also improves amino-acid recovery (AAR), indicating tighter coupling between sequence choice and interface geometry. 9. Additional analyses: performance advantages are clearest for short/medium peptides; PeTFlow performs strongly on loop binders and remains favorable on strand/helical cases. It also avoids amino-acid composition collapse seen in some baselines and yields favorable Rosetta interface energy distributions (with BoltzGen slightly lower ΔG in their comparison). 📜Paper: biorxiv.org/content/10.64898… #ComputationalBiology #ProteinDesign #PeptideDesign #GenerativeModels #FlowMatching #VQVAE #StructuralBiology #DrugDiscovery #Bioinformatics

What is the highest-paying job that requires the least amount of skill or effort?