ai-for-science

SciHorizon-eLab is an agentic protocol-to-task compiler that compiles scientific protocols into embodied tasks for scalable benchmarking, introducing a benchmark of 300 certified tasks.

This paper introduces a graph neural network method to predict transmembrane protein topology from 3D structures, demonstrating promising results by leveraging atom-level embeddings from AlphaFold without pre-trained weights.

The paper introduces the Transformer-Mamba for Flow Field (TM4FF) framework, a physics-constrained operator learning model that enhances accuracy and robustness in computational fluid dynamics simulations through innovations like Residual Wavelet Mamba and physics-informed loss.

The Large Knowledge Model (LKM) introduces a scientific knowledge infrastructure that transforms research papers into reasoning graphs, creating a Scientific Reasoning Landscape to support scientific search, question answering, and research planning with demonstrated accuracy improvements on benchmarks.

Peerify is a pipeline for automatically verifying peer-review claims against manuscript evidence, using a benchmark of 800 claims from NeurIPS 2024 and ICLR 2024, demonstrating that retrieval-centered verification outperforms entailment baselines.

EnSol is an environment-aware graph neural network that predicts molecular solubility by representing solutes and solvents as graphs and using cross-attention to model interactions, with probabilistic outputs to capture temperature effects and experimental uncertainty. It achieves state-of-the-art performance on benchmark datasets, validated experimentally.

SAIR Foundation announces an initiative to build open-weight AI models for mathematics, inviting community contributions and funding to support responsible AI in science.

TorchCraft is a unified binder design framework that inverts all-atom structure predictors like AlphaFold3 to optimize sequences for minibinders, VHHs, cyclic peptides, and ligand-binding proteins, with experimental validation across multiple formats.

Fraglingo introduces an attachment-aware autoregressive model for molecular design that generates molecules by jointly predicting fragment identity and attachment in a continuous latent space, enhancing property control and flexibility.

Researchers at UCLA, in collaboration with Caltech and NVIDIA, are using neural operators to advance AI-for-science applications in controlling complex quantum systems.

OpenDiscoveryTrace is a public dataset of 558 AI scientific agent trajectories that captures reasoning processes, not just outputs, to enable auditing and evaluation of scientific methodologies across multiple models and domains.

Anima Anandkumar describes a guest post on Terence Tao's blog detailing a method using Physics-Informed Neural Networks to find singular solutions for the Euler equations in fluid dynamics, highlighting its earlier release and broader applications in scientific computing.

This paper proposes a conditional generative latent framework to encode connectome graphs, enabling reconstruction, generation, and functional analysis in reservoir computing, with insights into task-specific structural mechanisms.

Northwestern University announces a new AI4Energy IIN/Trienens Postdoc fellowship for interdisciplinary research combining AI, nanotechnology, and energy science.

The paper introduces Elite-Weighted Supervised Fine-tuning (EW-SFT), a method for goal-directed molecular optimization that uses reward to guide elite selection and updates via the model's native loss, applicable across various generative architectures and tasks.

This paper proposes a multi-agent framework that enables LLM agents to conduct controlled experiments using simulation models for pharmaceutical process design, yielding more specific and actionable recommendations than language-only reasoning.

DeepSCan is a new AI model that learns the rules of cell-surface display and generates novel membrane display modules, several outperforming strong natural sequences, marking a step toward programmable cellular engineering.

This MIT Technology Review article argues that AlphaFold-style deep learning on massive datasets is not the ideal template for accelerating science, and that AI agents capable of reasoning and experimentation will drive future breakthroughs.

This paper investigates how molecular generative models internally organize molecular identity in their latent spaces, revealing piecewise-constant regions and coarse-to-fine boundaries across three architectures.

This paper introduces ADSL-PDE, a domain-specific language that provides a structured search space for auto-designing neural PDE solvers, improving search efficiency and optimization stability by abstracting away low-level implementation details. The evolutionary agent built on this representation achieves over 52% performance improvement within the first ten iterations across PDE benchmarks.