Find open-source science resources

Bi-directional DNA language model based on the Mamba state space architecture, enabling efficient long-range genomic sequence modeling with linear-time complexity and built-in reverse-complement equivariance; achieves strong performance on chromatin accessibility, enhancer, and promoter prediction benchmarks (Stanford & UC Berkeley, 500+ stars)

Transform arXiv papers into Beamer slides using LLMs

Universal scientific research intelligence covering 50+ disciplines, repositioning LLMs as cross-disciplinary generators with human experts as verifiers; 30B model outperforms Claude Opus and GPT on 5 research benchmarks

Open-source AI research assistant for biomedicine — chat to run RNA-seq, drug discovery, clinical analysis, and more; built on OpenClaw and Claude Code with 140 K-Dense scientific skills, real-time dashboard, and RStudio/JupyterLab integration (xjtulyc, 669+ stars, 2026)

Structure-aware protein language model using 3D structural vocabulary (Foldseek) for joint sequence-structure pretraining, achieving SOTA on protein engineering and fitness prediction benchmarks (ICML 2024, Westlake University & Repl)

Scientific foundation model and AI research copilot for idea generation, cross-disciplinary connection discovery, and hypothesis formation; trained with a decoupled reward-comment RL architecture and achieves GPT-4o-competitive novelty/rationale on STEM and social-science idea-generation benchmarks (270+ stars, MIT License, 2026)

Bilingual protein language model translating between protein sequence and structure, finetuned from ProtT5-XL on 17M AlphaFoldDB structures using Foldseek's 3Di structural alphabet, enabling sequence-to-structure prediction, structure-to-sequence inverse folding, and unified protein representation learning (RostLab, 310+ stars)

Multimodal deep learning framework integrating peptide-MHC protein sequence, structure, and biochemical properties to predict class-I immunogenicity for infectious disease epitopes and cancer neoepitopes with cancer-wildtype contrastive learning, enabling personalized vaccine design (Krishnaswamy Lab, Yale University)

First benchmark for automatic video generation from scientific papers (NeurIPS 2025)

RFantibody is a pipeline for structure-based de novo antibody and nanobody design, integrating backbone design with RFdiffusion, sequence design with ProteinMPNN, and structure prediction with RoseTTAFold2. It provides a comprehensive toolset for generating and filtering high-quality antibody designs.

Interactive personal genome analysis toolkit using Claude Code and Python. Parses raw genotyping data from consumer DNA services and analyzes SNPs across 17 categories including health risks, pharmacogenomics, ancestry, and nutrition, with a terminal-style HTML dashboard.

Open-source platform for building, extending, and experimenting with scientific agents, providing modular agent construction tools and standardized evaluation pipelines for accelerating autonomous scientific discovery research (748+ stars, MIT License)

First benchmark evaluating LLMs' ability to rediscover scientific laws through interactive experimentation across 324 tasks in 12 physics domains, featuring memorization-resistant metaphysical shifts of canonical laws (HKUST)

GenBio AI's software stack for the AI-Driven Digital Organism, supporting adaptation and finetuning of multiscale biological foundation models across DNA, RNA, protein, structure, and single-cell tasks with reproducible CLIs and pretrained model zoo (2025)

Flow-matching protein folding model using only general-purpose transformer layers, scaled to 3B parameters and trained on 8.6M+ distilled structures; challenges the reliance on complex domain-specific architectures and supports PyTorch and MLX backends with model sizes from 100M to 3B parameters (985+ stars, MIT License)

Foundation models for genomics and transcriptomics pretrained on 3,000+ human genomes and 850+ diverse species, enabling chromatin accessibility prediction, splice site detection, and promoter classification across multiple model scales (InstaDeep, NVIDIA & TUM, Nature Methods 2023)

Universal pretrained neural network potential with charge and magnetic moment awareness, trained on 1.5M+ Materials Project inorganic structures for charge-informed molecular dynamics and phase diagram prediction (Berkeley, Nature Machine Intelligence 2023 Cover)

Deep learning-based object detection and segmentation for star-convex shapes, widely adopted for cell and nucleus segmentation in fluorescence and electron microscopy via a compact neural network architecture with non-maximum suppression and shape-based post-processing (Nature Methods 2020, 1.2K+ stars)

Euclidean neural networks for arbitrary point transformations enabling E(3)-equivariant deep learning, foundational library for building geometry-aware neural networks in molecular dynamics, materials science, and physics

Rectified Quaternion Flow for efficient protein backbone generation, 37× faster than RFDiffusion with 0.972 designability (ICML 2025)

Physics-informed neural networks

Azure Semantic Kernel multi-agent PPT generation reference

Multi-agent system automatically transforming research papers into interactive AI agents with MCP server generation, tutorial auto-detection, and benchmark extraction (2.2K+ stars, MIT License, 2025)

Self-supervised vision foundation model for generalized structural brain MRI analysis, pretrained on ~49,000 scans from diverse datasets and generalizing across brain age prediction, dementia/MCI classification, IDH mutation detection, glioma survival prediction, time-to-stroke estimation, MR sequence classification, and brain tumor segmentation; outperforms task-specific models especially with limited training data (Mass General Brigham & Harvard Medical School, 129+ stars)

Lightweight supervised slide foundation model with 0.9M parameters pretrained on 24K whole-slide images for pan-cancer morphological classification, achieving competitive performance with much larger self-supervised models (TITAN, GigaPath) while enabling finetuning on consumer-grade GPUs; includes standardized MIL implementations and benchmarking across 15+ classification tasks (Mahmood Lab, Harvard Medical School, 153+ stars)

Evaluating multimodal autonomous agents in realistic scientific workflows across real scientific software environments (KAlgebra, Celestia, Grass GIS, Lean 4, etc.) with VM-based evaluation infrastructure and agent trajectories

Structure prediction and design of proteins with noncanonical amino acids, enabling AI-powered modeling of synthetic biology constructs and expanded genetic code systems (133+ stars, 2025)

Scientific equation discovery with agentic AI, elevating LLMs from equation proposers to autonomous scientists that write code, analyze data, implement equations, and optimize based on experimental feedback; outperforms baselines by 6-35% across four science disciplines with robustness to noise and out-of-domain generalization (GAIR-NLP / SJTU, 49+ stars, Apache 2.0)

Foundation model for joint segmentation, detection, and recognition of biomedical objects across nine imaging modalities, with v2 introducing BoltzFormer architecture for end-to-end 3D inference (Microsoft, Nature Methods 2025)

DNA sequence analysis

Open-source toolkit and benchmark for learning-based theorem proving in Lean, providing programmatic Lean interaction, a 98K+ theorem dataset extracted from 217 Lean projects, and ReProver—the first retrieval-augmented LLM-based theorem prover for Lean—with reproducible training pipelines underpinning much subsequent Lean prover research (Caltech & NVIDIA, NeurIPS 2023 Outstanding Paper, Datasets & Benchmarks)

Discrete diffusion framework for generative protein sequence design over evolutionary-scale databases, supporting unconditional generation, evolutionary-guided conditional design, motif scaffolding, and intrinsically disordered region generation through order-agnostic autoregressive diffusion, enabling sequence-only protein design without structural priors (Microsoft Research, Nature Communications 2024)

ICML 2025 drug discovery generalist using masked discrete diffusion and fragment-based generation with molecular context guidance (NVIDIA)

DeepMind's Olympiad-level geometry theorem prover combining neural language model with symbolic deduction engine, AlphaGeometry2 solves 84% of IMO geometry problems (42/50) at gold-medalist level (Nature 2024)

Standard data-centric AI package for data quality and machine learning, automatically detecting label errors, outliers, and dataset issues to improve scientific dataset reliability and model performance (11K+ stars, MIT License)

Efficient foundation model and benchmark for multi-species genome understanding with context-aware nucleotide representations, improving upon DNABERT for diverse genomic task transfer learning (UIUC MAGICS Lab, 484+ stars)

Fast, modular, and accurate de novo design of protein binders based on the Protenix foundation model, achieving 17-82% nanomolar hit rates across diverse targets with 2-6× improvement over prior methods like AlphaProteo and RFdiffusion (229+ stars, Apache 2.0)

AI-human collaborative research platform where a human researcher works with a team of LLM agents via team and individual meetings to perform scientific research; demonstrated by designing new SARS-CoV-2 nanobodies with wet-lab validation

Deep equivariant generative model predicting ligand-specific protein-ligand complex structures with dynamic receptor conformational flexibility, enabling accurate docking for flexible protein targets

Official implementation of the second-generation fully autonomous scientific discovery system, extending the original with agentic tree search and reduced template dependency to achieve workshop-level accepted papers (6.7K+ stars, 2025)

First fully autonomous open-ended scientific discovery system with official implementation: hypothesis→experiment→writing→review simulation (13.8K+ stars, 2024)

Shanghai AI Lab's deep learning-based global weather forecasting model pushing skillful forecasts beyond 10 days lead, with open-source inference code and pretrained ONNX model weights (arXiv 2023)

Cross-modal self-supervised foundation model for galaxies by Polymathic AI, jointly embedding multi-band galaxy imaging and optical spectra into a shared latent space to enable zero/few-shot redshift estimation, galaxy property prediction, morphology classification, and cross-modal similarity search (MNRAS Letters 2024)

Trainable, memory-efficient PyTorch reproduction and retraining of AlphaFold2 providing new insights into its learning dynamics and out-of-distribution generalization; widely used as the open-source AlphaFold2 backbone underpinning many downstream protein structure prediction and design pipelines (Columbia AlQuraishi Lab & OpenFold Consortium, Nature Methods 2024)

Generative AI framework for inverse design of 3D RNA structure and function using geometric deep learning, learning design rules from 3D structures to capture complex tertiary interactions (pseudoknots, non-canonical base pairs) with expert-level accuracy for designing functional RNAs including aptamers and ribozymes (bioRxiv 2025)

Multimodal whole-slide pathology foundation model jointly pretrained on H&E histology and diagnostic text reports, enabling zero-shot cancer subtyping, biomarker prediction, and multimodal reasoning across diverse cancer types (Mahmood Lab, 341+ stars)

Full spaCy pipeline and models for scientific/biomedical documents, enabling named entity recognition, abbreviation resolution, and UMLS linking for scientific literature mining (1.9K+ stars, Apache 2.0)

Autonomous multi-agent research loop for model architecture discovery that ran 1,773 experiments over 20,000 GPU hours and produced 106 state-of-the-art linear-attention architectures, surpassing human-designed baselines including Mamba2 and DeltaNet (1.1K+ stars, Apache 2.0)

Foundation model jointly trained on single-cell and spatial transcriptomics data, enabling unified representation learning across cellular and tissue spatial contexts for cell type prediction, spatial domain inference, and cross-modal integration (theislab, bioRxiv 2024, 164+ stars)

100M-parameter foundation model pretrained on 50M+ human single-cell transcriptomes covering ~20,000 genes, achieving SOTA on gene expression enhancement, drug response and perturbation prediction (Nature Methods 2024)