Jorge Bravo Abad

A generative artificial intelligence approach for the discovery of antimicrobial peptides against multidrug-resistant bacteria - Nature Microbiology This study presents a generative artificial intelligence approach for the high-throughput discovery of antimicrobials against multidrug-resistant bacteria.

Antibiotic resistance is outpacing new drug discovery. Yihui Wang and coauthors introduce ProteoGPT + helper models to identify, test, and generate antimicrobial peptides—scanning millions and creating new ones, validated against multidrug-resistant bacteria. www.nature.com/articles/s41...

Zero-Shot Autonomous Microscopy for Scalable and Intelligent Characterization of 2D Materials Characterization of atomic-scale materials traditionally requires human experts with months to years of specialized training. Even for trained human operators, accurate and reliable characterization r...

ATOMIC brings zero-shot autonomy to 2D materials: SAM+LLM control the scope, segment flakes, and classify layers—no training data. 99.7% monolayer accuracy, grain-boundary detection, robust to imaging drift, and generalizes to graphene, MoS2, WSe2, SnSe. pubs.acs.org/doi/10.1021/...

Foundation model for efficient biological discovery in single-molecule time traces - Nature Methods META-SiM brings foundation model power to single-molecule time traces, excelling across diverse analysis tasks. Paired with the web-based META-SiM Projector and entropy mapping, it rapidly reveals hidden molecular behaviors inaccessible by other means.

Li and coauthors present META-SiM, a transformer foundation model for single-molecule fluorescence data. It streamlines analysis, flags rare states, and even uncovered a new splicing intermediate missed before. www.nature.com/articles/s41...

High-Performance and Reliable Probabilistic Ising Machine Based on Simulated Quantum Annealing Simulated quantum annealing enhances probabilistic Ising Machines by enabling faster, more reliable solutions to complex optimization problems using interacting copies of the system guided by a time-d...

Raimondo & coauthors show that simulated quantum annealing makes probabilistic Ising machines faster, more reliable, and robust to hardware variability. A CMOS design demonstrates nanosecond updates and low power—paving the way for scalable optimization hardware. journals.aps.org/prx/abstract...

Heat-rechargeable computation in DNA logic circuits and neural networks - Nature Heat recharges enzyme-free DNA circuits, enabling complex logic operations and neural networks to perform multiple computations, offering a universal energy source for molecular machines and advancing autonomous behaviours in artificial chemical systems.

DNA can compute—but most circuits are single-use. Tianqi Song & Lulu Qian show how a simple heat pulse “recharges” DNA logic and neural networks, enabling 16+ rounds of reusable computation with >200 species. A step toward sustainable molecular computing. www.nature.com/articles/s41...

Fourier-based three-dimensional multistage transformer for aberration correction in multicellular specimens - Nature Methods Adaptive optical vision Fourier transformer (AOViFT) is a machine learning-based framework for accurately inferring aberrations and restoring diffraction-limited performance in diverse biological specimens.

Thayer Alshaabi & coauthors introduce AOViFT: a Fourier-based 3D transformer that corrects aberrations without guide stars or wavefront sensors. Fast, low-cost AO for live imaging in zebrafish embryos and beyond. www.nature.com/articles/s41...

Elana Simon & James Zou introduce interPLM, a sparse autoencoder framework that reveals interpretable features in protein LMs—capturing active sites, motifs, and domains, uncovering missing annotations, and steering sequence generation for more transparent bio-AI. www.nature.com/articles/s41...

Predictive model for the discovery of sinter-resistant supports for metallic nanoparticle catalysts by interpretable machine learning - Nature Catalysis The activity and stability of supported metal catalysts is in large part influenced by their interaction with the support. Now, neural network molecular dynamics simulations are combined with interpretable machine learning to reveal the governing factors of metal–support interactions for Pt nanoparticles on various oxide supports, identifying key features and proposing sinter-resistant supports.

Jiang and coauthors combine neural-network MD with interpretable ML to predict sinter-resistant supports for Pt catalysts. Key features guide screening of 10,000+ oxides, validated at 800 °C with ceria and BaO showing strong stability. www.nature.com/articles/s41...

Machine Learning the Energetics of Electrified Solid-Liquid Interfaces A framework rooted in perturbation theory extends machine-learning interatomic potential approaches, capturing complex dynamics and energetics at electrified solid-liquid interfaces.

Bergmann and coauthors present RAZOR, a response-augmented ML potential that learns how interfacial energies shift with bias. On OH/Cu(100), it captures pH-dependent site switching seen in experiments—bringing first-principles fidelity to ML-speed electrochemistry. journals.aps.org/prl/abstract...

Gunasekaran and coauthors introduce Future-Guided Learning: a predictive coding–inspired approach where a “teacher” model looks ahead to guide a “student.” It boosts seizure prediction and cuts errors in chaotic systems—making forecasts more adaptive and resilient. www.nature.com/articles/s41...

A brain-inspired agentic architecture to improve planning with LLMs - Nature Communications Multi-step planning is a challenge for LLMs. Here, the authors introduce a brain-inspired Modular Agentic Planner that decomposes planning into specialized LLM modules, improving performance across tasks and highlighting the value of cognitive neuroscience for LLM design.

Brain-inspired planning for LLMs: MAP coordinates Monitor, Actor, Predictor, Evaluator, Decomposer, Orchestrator + light tree search. Fewer invalid moves, stronger transfer: 74% ToH (vs ~11% GPT-4), near-perfect CogEval, beats baselines on PlanBench & StrategyQA. www.nature.com/articles/s41...

Solving the many-electron Schrödinger equation with a transformer-based framework - Nature Communications Accurately solving the Schrödinger equation is challenging. Here, authors present QiankunNet, a Transformer-based framework that efficiently captures quantum correlations, achieving high accuracy in complex molecular systems using neural network quantum states.

Honghui Shang and coauthors introduce QiankunNet, a Transformer-based neural quantum state. With autoregressive sampling and physics-informed initialization, it achieves 99.9% of FCI correlation energy and tackles large active spaces like the Fenton reaction. www.nature.com/articles/s41...

Harnessing surrogate models for data-efficient predictive chemistry: descriptors vs. learned hidden representations Predictive chemistry often faces data scarcity, limiting the performance of machine learning (ML) models. This is particularly the case for specialized tasks such as reaction rate or selectivity predi...

Predictive chemistry often struggles with scarce data. Surrogate models can help, but should we use their predicted QM descriptors or hidden embeddings? Chen & Stuyver show that hidden spaces usually win—faster, more robust, and data-efficient. pubs.rsc.org/en/content/a...

YouTube video by Jorge Bravo Abad IA generativa con Física: creando una máquina de Boltzmann que dibuja números

New clip from my latest talk (in Spanish): How we can use concepts from physics to design generative AI models, such as Restricted Boltzmann Machines capable of generating handwritten digits. www.youtube.com/watch?v=wX5r...

Physics-informed deep learning for plasmonic sensing of nanoscale protein dynamics in solution An infrared metasurface combined with physics-informed AI enables sensing of nanoscale protein dynamics in solution.

Chenchen Wu and coauthors combine a graphene–gold plasmonic sensor with a physics-informed CNN to track protein folding directly in water. The hybrid approach resolves sub-10-nm structures and real-time shifts during assembly with >2× the accuracy of standard CNNs. www.science.org/doi/10.1126/...

Jiang and coauthors: mice and AI agents learn to cooperate in remarkably similar ways. Cooperation is encoded in the anterior cingulate cortex in brains, and in specialized units in artificial networks. Biology and AI converge on shared principles. www.science.org/doi/10.1126/...

Conferencias- De átomos a algoritmos: la revolución de la IA en Física y Química YouTube video by FundacionAreces

My new talk (in Spanish) on how Artificial Intelligence is reshaping Physics — from quantum spins to gravitational waves: www.youtube.com/live/CdCmqvv...

Quantum learning advantage on a scalable photonic platform Recent advances in quantum technologies have demonstrated that quantum systems can outperform classical ones in specific tasks, a concept known as quantum advantage. Although previous efforts have foc...

Zheng-Hao Liu and coauthors show a provable quantum learning advantage on a photonic platform. Using entangled photons and Bell measurements, they cut sample complexity by 11.8 orders of magnitude, scaling to 100+ modes and opening new paths for quantum sensing and ML www.science.org/doi/10.1126/...

Robot-assisted mapping of chemical reaction hyperspaces and networks - Nature A low-cost robotic platform using mainly optical detection to quantify yields of products and by-products allows the analysis of multidimensional chemical reaction hyperspaces and networks much faster than is possible by human chemists.

Jia and coauthors unveil a $25k robot that maps chemical reaction hyperspaces. Using UV–Vis spectral unmixing, it scans thousands of conditions, finds smooth yield landscapes, anomalies, and switchable networks—offering a scalable path to discovery and optimization. www.nature.com/articles/s41...

Merging conformational landscapes in a single consensus space with FlexConsensus algorithm - Nature Methods FlexConsensus is a multi-autoencoder-based algorithm for merging different conformational landscapes from cryogenic electron microscopy heterogeneity analysis into a common latent space for the identification of similarities and differences among various methods. This helps in the validation of estimated conformational landscape and provides tools to streamline the heterogeneity workflow.

Cryo-EM reveals how proteins flex but different algorithms often give conflicting maps of motion. Herreros & coauthors present FlexConsensus, a deep learning framework that merges them into a single reliable consensus space, improving trust in protein dynamics analysis www.nature.com/articles/s41...

Efficient protein structure generation with sparse denoising models - Nature Machine Intelligence A small and fast diffusion model is presented, which is able to efficiently generate long protein backbones.

Jendrusch and Korbel present SALAD, a sparse denoising model for protein design. It generates backbones up to 1,000 amino acids, faster and leaner than prior models, and adapts to new tasks with “structure editing”—from motif scaffolding to multi-state design. www.nature.com/articles/s42...

Error-controlled non-additive interaction discovery in machine learning models - Nature Machine Intelligence Diamond, a statistically rigorous method, is capable of finding meaningful feature interactions within machine learning models, making black-box models more interpretable for science and medicine.

Winston Chen and coauthors present Diamond, a framework to uncover non-additive feature interactions in ML models with strict error control. From diabetes progression to gene regulation, it turns black-box outputs into reliable, testable scientific insights. www.nature.com/articles/s42...

Analogue speech recognition based on physical computing - Nature A temporal-signal processor based on two in-materia computing hardware platforms—reconfigurable nonlinear-processing units (RNPUs) and analogue in-memory computing (AIMC)—is used for both feature extr...

Zolfagharinejad and coauthors show that analogue hardware can “hear.” Using nonlinear silicon units for cochlea-like feature extraction and in-memory chips for classification, they achieve near-software speech recognition at millisecond latency and ultra-low energy. www.nature.com/articles/s41...

DeepSeek-R1, now published in Nature, shows how reinforcement learning can turn AI into more than an imitator. By rewarding correct answers, it develops reasoning strategies on its own—achieving breakthroughs in math, coding, and STEM problem-solving. www.nature.com/articles/s41...

Towards compute-efficient Byzantine-robust federated learning with fully homomorphic encryption - Nature Machine Intelligence Lancelot, a compute-efficient federated learning framework using homomorphic encryption to prevent information leakage, is presented, achieving 20 times faster processing speeds through advanced crypt...

Federated learning lets institutions train AI without sharing raw data—but it’s vulnerable to attacks and leaks. Siyang Jiang and coauthors present Lancelot, a homomorphic-encryption framework that secures training while cutting costs. www.nature.com/articles/s42...

Target-specific de novo design of drug candidate molecules with graph-transformer-based generative adversarial networks - Nature Machine Intelligence Inhibiting AKT1 kinase can have potentially positive uses against many types of cancer. To find novel molecules targeting this protein, a graph adversarial network is trained as a generative model.

Atabey Ünlü and coauthors present DrugGEN, a graph-transformer GAN that designs drug-like molecules tailored to specific protein targets. Trained on bioactive datasets, it generated and validated AKT1 inhibitors, showing how AI can directly shape drug discovery. www.nature.com/articles/s42...

SurFF: a foundation model for surface exposure and morphology across intermetallic crystals - Nature Computational Science A foundation machine learning model, SurFF, enables DFT-accurate predictions of surface energies and morphologies in intermetallic catalysts, achieving over 105-fold acceleration for high-throughput m...

Most industrial reactions happen on catalyst surfaces, but predicting which surfaces actually form is costly. Jun Yin and coauthors present SurFF, a foundation model that predicts surface exposure across intermetallic crystals with DFT-level accuracy, 100k× faster. www.nature.com/articles/s43...

A complete photonic integrated neuron for nonlinear all-optical computing - Nature Computational Science This study reports a complete photonic neuron integrated on a silicon-nitride chip, enabling ultrafast all-optical computing with nonlinear multi-kernel convolution for image recognition and motion ge...

A leap in photonic AI: Tao Yan and coauthors demonstrate a complete photonic integrated neuron on a chip. It combines optical convolution and activation, achieving ultrafast, energy-efficient processing for tasks from image recognition to motion generation. www.nature.com/articles/s43...

Real-time raw signal genomic analysis using fully integrated memristor hardware - Nature Computational Science The authors report a memristor-based system that analyzes raw analog signals from a genomic sequencer directly in memory. By bypassing slow data conversion, the system achieves substantial improvement...

Peiyi He and coauthors show how memristor arrays can process raw nanopore sequencing signals directly in memory. A step toward real-time, low-power DNA sequencing—and a compelling use case for analog computing. www.nature.com/articles/s43...

Sampling-enabled scalable manifold learning unveils the discriminative cluster structure of high-dimensional data - Nature Machine Intelligence A sampling-based manifold learning method is proposed to study the cluster structure of high-dimensional data. Its applicability and scalability have been verified in single-cell data analysis and ano...

Dehua Peng and coauthors advance manifold learning with scalable algorithms that boost stability, accuracy, and interpretability—opening new possibilities for analyzing high-dimensional data in biology, physics, and AI. www.nature.com/articles/s42...