Search

Drug discovery with NVIDIA and cutting-edge AI: creating pharmaceutical synthesis enzymes from scratch

Translated title. Original: エヌビディアと最先端AIによる創薬：ゼロから医薬品合成酵素を生み出す

A method named Proteina-Complexa performs joint sequence-structure design in a continuous latent space with inference-time search for de novo protein binders. In a large benchmark with 1M proteins across 127 targets via multiplexed phage display, it surpasses open methods, delivering picomolar binders and >50% hit rates.

Vibe genes? Natural(ish) language comes to protein, genome design software - Ars Technica

Researchers built a natural-language-driven protein design tool that creates proteins from user prompts, with Hugging Face offering a similar DNA-level tool; both rely on a two-step process of generating a compatible shape from constraints and then selecting a sequence to ensure proper folding, though success rates vary.

Open AI model predicts how RNA folds, trained on 21 million examples | WATTBA

Translated title. Original: Open AI-model voorspelt hoe RNA zich vouwt, getraind op 21 miljoen voorbeelden | WATTBA

Researchers present structRFM, an AI model that learns RNA folding from sequence and pairing information, trained on 21 million sequences. It outperforms prior methods, extends to 3D structure, and is openly available for validation and further development.

Google DeepMind’s new AlphaFold can model a much larger slice of biological life | MIT Technology Review

AlphaFold 3 can predict structures and interactions involving DNA, RNA, proteins, and drug-like molecules, expanding its potential use in drug discovery. Its performance varies by task, and researchers say some predictions remain inaccurate.

AI hallucinations led to a discovery. Polish researchers describe a new “impossibly tangled” family of proteins | Science in Poland

Translated title. Original: Halucynacje AI doprowadziły do odkrycia. Polacy opisali nową „niemożliwie zasupłaną” rodzinę białek | Nauka w Polsce

Researchers at the University of Warsaw confirmed an AI-predicted protein knot that existing theories considered too complex to form. Found in proteins from gut bacteria, the eight-crossing knot’s formation remains unexplained.

Deep learning model predicts rice protein interactions from sequence data

A deep learning model called RCNNricePPI predicts rice protein-protein interactions from sequence data, using a Siamese network and word2vec embeddings; it shows strong computational metrics and is available as a web resource, intended to guide experimental validation rather than replace lab work.

MegaRobo Launches Poreva Open Protein Design Contest | TechUpscale

MegaRobo announces Poreva Open, a global AI protein design contest evaluating 10,000+ computer-generated designs through standardized, double-blind wet-lab testing, with data transparency and potential investment for top teams; it runs in five challenges focused on cancer- and metabolism-related targets, with registration in late 2026 and results in 2027.

[2611.13348] ProTEUS: A mutation-resolved world model for closed-loop protein evolution

ProTEUS combines mutation-effect prediction, fast physics-constrained structure modeling, and experimental feedback to guide protein evolution. In two campaigns, it produced aldehyde dehydrogenase and levansucrase variants with 10 times the activity of wild type.

RFD4-Proteina

RFD4-Proteina is a physics-informed, deep-learning framework for in silico protein design that uses a shared latent space to model amino acids, sugars, DNA, and RNA. It combines intrinsic recurrent refinement and extrinsic constraint-driven search, delivering higher design throughput and success than competing methods, with experimental validations including binding, catalysis, and a de novo enzyme demonstration.

RFD4-Proteina

RFD4-Proteina is a physics-informed, shared latent-space model for in silico protein design, using intrinsic and extrinsic reasoning to generate physically plausible biomolecules and achieve higher design throughput and pass rates across diverse benchmarks, including proteins, small molecules, enzymes, and DNA binding, with validation on binding, catalysis, and de novo enzyme design.