Over 100 Groundbreaking AI for Science Papers to Watch in 2025 – A Rapid Overview

AI for Science 2025 Highlights

In the past year AI has moved from a set of isolated tools to a systematic, reusable research paradigm. HyperAI’s HyperAI‑Super‑Neuron project curated and classified more than one hundred influential papers that demonstrate how deep learning, diffusion models, reinforcement learning and large language models are reshaping scientific discovery across biomedicine, materials chemistry, climate science, astronomy and more.

Biomedicine & Healthcare

Advancing Protein Ensemble Predictions Across the Order‑Disorder Continuum – a generative model that predicts protein ensembles for intrinsically disordered proteins. (Peptone et al., bioRxiv 2025.10.18.680935v1) NOBLE: Neural Operator with Biologically‑informed Latent Embeddings to capture experimental variability in neuronal models. Achieves a 4 200‑fold speedup over traditional simulators. (NeurIPS 2025, MIT, Caltech, University of Alberta)

PLACER : A graph neural network that resolves protein conformational heterogeneity on the atomic scale. (PNAS, David Baker lab )

Squidiff : Diffusion‑based simulation of multi‑scenario transcriptomics to accelerate precision and spatial medicine. (Nature Methods)

Deep generative classification of blood cell morphology : A diffusion model that outperforms clinical experts in leukemia detection. (Nature) Ctrl‑DNA: Constrained reinforcement learning for cell‑specific cis‑regulatory element design, enabling targeted gene expression control. (NeurIPS 2025, University of Toronto, Changping Lab)

BoltzGen : Universal binder design framework that achieves nanomolar affinity for 66 % of targets. (MIT & collaborators, bioRxiv )

FusionProt : Fusion of sequence and structural information for unified protein representation, reaching SOTA on multiple tasks. (bioRxiv, Technion‑Israel Institute of Technology & Meta AI)

MorphDiff : Transcriptome‑guided diffusion model that predicts cellular morphology changes under perturbations, speeding up phenotype‑driven drug discovery. (Nature Communications)

AlphaPPIMI : Deep learning framework for predicting protein‑protein‑interaction modulators, surpassing previous methods. (Journal of Cheminformatics)

scSiameseClu : Siamese clustering framework for unsupervised single‑cell RNA‑seq analysis, achieving SOTA performance on seven benchmark datasets. (IJCAI 2025)

Multi‑metal binding site predictor : Graph‑based neural network that identifies metal‑binding residues across protein sequences. (bioRxiv, Hong Kong University of Science & Technology)

AlphaFold‑Metainference : Extends AlphaFold to predict structural ensembles of disordered proteins, improving ensemble accuracy. (Nature Communications, Cambridge University)

Prot42 : Protein language model that generates high‑affinity binders for long sequences (up to 8 k residues). (arXiv, Inception AI, Cerebras)

Medical GraphRAG : Graph‑retrieval‑augmented generation framework that sets new SOTA on 11 medical QA datasets. (ACL 2025, Oxford & Carnegie Mellon)

Healthcare Agent : Large language model that elicits proactive, relevant medical consultations, outperforming closed‑source models such as GPT‑4. (Nature Artificial Intelligence)

NeuralCohort : Cohort‑aware neural representation learning that improves hospital stay time prediction by 16.3 % using 8 million EHR records. (ICML 2025, NUS & Zhejiang University)

Visual and Domain Knowledge for Professional‑level Graph‑of‑Thought Medical Reasoning : Multimodal, multilingual foundation model achieving SOTA on 14 disease benchmarks. (Nature Portfolio)

MedFound : 176‑billion‑parameter open‑source medical language model that matches expert performance on disease diagnosis tasks. (Nature Medicine, Beijing University of Posts & Telecommunications & Peking University)

Materials Chemistry & Engineering

PET‑MAD : Lightweight universal interatomic potential trained on fewer than 100 k crystal structures, matching specialist model accuracy. (Nature Communications, EPFL)

ChemOntology : Explicit chemical ontology‑based method that halves computational cost for reaction‑path searches. (ACS Catalysis, Hokkaido University)

MOF‑ChemUnity : Structured, extensible knowledge graph for metal‑organic frameworks that integrates 9 874 papers and 15 000 crystal structures. (ACS Publications) CGformer: Transformer‑enhanced crystal graph network that adds global attention to CGCNN, improving material property prediction. (Matter, Shanghai Jiao‑Tong University)

FASTSOLV : Data‑driven solubility predictor that works at arbitrary temperatures with 50× faster inference. (Nature Communications, MIT)

UNIMATE : Unified model for mechanical metamaterial generation, property prediction and condition verification. (ICML 2025, Virginia Tech & Meta AI)

DreaMS : Self‑supervised learning from 200 million tandem mass spectra, creating the largest public mass‑spectrometry dataset (GeMS). (Nature Biotechnology, Czech Academy of Sciences)

Retrieval‑Retro : Retrieval‑based inorganic retrosynthesis planning system that dramatically speeds up synthesis route generation. (NeurIPS 2024, Korea Institute of Science & Technology)

MatterGen : Diffusion model that generates inorganic crystal structures conditioned on target space groups, enabling rapid materials discovery. (Nature, Microsoft)

Unified differentiable learning of electric response : Equivariant ML framework that learns both potential energy surfaces and electric response functions for a wide range of materials. (Nature Communications, Harvard & Bosch)

Climate & Weather Modeling

ERDM : Elucidated Rolling Diffusion Model that jointly designs progressive noise schedules and time‑loss weighting, achieving state‑of‑the‑art medium‑range forecasts. (NeurIPS 2025, NVIDIA & UC San Diego) OmniCast: Masked latent diffusion model that produces accurate 4‑month seasonal forecasts in under 2 minutes, eliminating error accumulation of autoregressive models. (NeurIPS 2025, UCLA & Argonne National Lab)

FCN‑3 : Geometric probabilistic weather forecasting system that combines spherical signal processing with hidden‑Markov ensembles, delivering 15‑day forecasts in ~1 minute on a single GPU. (NeurIPS 2025, NVIDIA & Berkeley Lab)

End‑to‑end data‑driven weather prediction : First fully data‑driven system that matches traditional numerical models while being tens of times faster. (Nature, Cambridge, DeepMind, UC Berkeley)

Hyperlocal Extreme Rainfall Forecasts in Mumbai : Transfer‑learning based CNN downscaling approach that predicts heavy rain events days in advance using 36 weather stations. (SSRN, IIT Bombay & University of Maryland)

ACE2 Seasonal Forecast Model : Machine‑learning weather model trained on reanalysis data that produces a 4‑month seasonal forecast in ~2 minutes. (npj Climate & Atmospheric Science, UK Met Office & Allen Institute for AI)

Astronomy & Astrophysics

Quasars acting as Strong Lenses Found in DESI DR1 : Convolutional neural network that identifies seven high‑quality strong‑lens quasar candidates from 810 000 spectra. (arXiv, Stanford, SLAC, Peking University, INAF‑Bologna, UCL, UC Berkeley)

AION‑1 : Omnimodal foundation model for astronomical sciences trained on 200 million objects, supporting image, spectra and time‑series tasks. (NeurIPS 2025, UC Berkeley, Cambridge, Oxford)

Each entry includes a brief description of the methodology, the core scientific claim, and a direct citation to the original pre‑print or journal article, allowing readers to locate the full source material instantly.