Preview the new results, new developments, and new perspectives of AI for Science——
* DeepMind’s latest research FunSearch is published in Nature
* Google launches healthcare industry model MedLM
* Jingtai Technology sprints to the Hong Kong Stock Exchange, AI+ robot empowers AI for Science
* GHDDI cooperates with Microsoft Research Center for Scientific Intelligence
* Open source AI tools for seismological processing and analysis
* my country’s first self-developed earth system model announced as open source
* Baidu Feiyuan Propeller team builds protein-small molecule docking conformation prediction model HelixDock
* Domestic research team discloses carbon emission prediction method and system based on hybrid machine learning
* Apple chip "exclusive customized version" machine learning framework open source
See below for more details~
Entreprise's news
DeepMind’s latest research FunSearch is published in “Nature”
Google DeepMind's latest research, FunSearch, is a way to search for new solutions in mathematics and computer science. FunSearch works by pairing a pre-trained large model (LLM), which aims to provide creative solutions in the form of computer code, with an automatic "evaluator" that is responsible for preventing hallucinations and incorrect idea. Through iteration back and forth between these two components, the initial solution "evolves" into new knowledge. FunSearch discovered a new solution to the upper bound set problem, a long-standing problem in mathematics and represents the first discovery of a challenging open problem in science or mathematics using large models. Paper address: http://nature.com/articles/s41586-023-06924-6
Google launches healthcare industry model MedLM
Recently, Google announced the launch of a new set of healthcare-specific artificial intelligence models, MedLM, designed to help clinicians and researchers conduct complex research, summarize doctor-patient interactions, etc. The move marks Google’s latest attempt to monetize artificial intelligence tools for the healthcare industry and is an important milestone in the digital transformation of the healthcare industry. First, MedLM can help clinicians and researchers conduct complex research and data analysis to improve the accuracy and efficiency of medical diagnosis. Secondly, MedLM can summarize doctor-patient interactions and provide doctors with better patient management and service experience. In addition, MedLM can provide better data management and analysis tools for healthcare institutions to improve the efficiency of utilization of medical resources.
Jingtai Technology sprints to the Hong Kong Stock Exchange, AI+ robot empowers AI for Science
QuantumPharm Inc. (Jingtai Technology) officially submitted a prospectus to the Hong Kong Stock Exchange last month and plans to list on the main board under the 18C rules. The 18C rules are mainly aimed at specialized technology companies and have higher requirements for the technological attributes of the industry, involving new generation information technology, advanced hardware and software, advanced materials, new energy and energy conservation and environmental protection, new food and agricultural technology and other industry fields. Jingtai Technology is one of the few drug and materials science R&D companies in the world that simultaneously possesses first-principles computing based on quantum physics, advanced artificial intelligence technology and automated wet laboratory capabilities. It is also one of the few quantum physics + AI + automation driven companies in the world. One of the drug and materials science research and development platforms.
GHDDI partners with Microsoft Research Center for Scientific Intelligence
Recently, the Global Health Drug Discovery Institute (GHDDI) and Microsoft Research AI4Science announced a collaboration. The two parties will jointly develop generative artificial intelligence and basic large models in the field of global health infectious diseases. Technology, focusing on implementation and transformation, accelerating the research and development of innovative drugs. Previously, the two parties have successfully designed a variety of small molecule inhibitors with new structures in the research on key target proteins of Mycobacterium tuberculosis and coronavirus.
BIO Geometry and Zhipu AI jointly build a natural language-life language multi-modal large model
Beijing Biogeometry Biotechnology Co., Ltd. and Beijing Zhipu Huazhang Technology Co., Ltd. recently announced a strategic cooperation to jointly build a multi-modal model of natural language-life language. This model is expected to enhance the usefulness of generative artificial intelligence platforms in the fields of life sciences and medical research.
Tools and resources
AI tools for seismological processing and analysis open source
Open source tools for seismological processing and analysis currently include: phase picking, polarization, and dispersion extraction. The tool has open sourced the 100Hz model in China. Some models are trained based on the CSNCD data set. The picking model of the four seismic phases of PgSgPnSn has the highest accuracy. Visit address: https://gitee.com/cangyeone/seismological-ai-tools
my country's first self-developed earth system model announced as open source
Recently, the Institute of Atmospheric Physics of the Chinese Academy of Sciences released my country's first "complete" earth system numerical model with independent intellectual property rights and announced the release of its source code. This model contains a complete climate system and ecological environment system, integrating 8 sub-system models such as atmospheric circulation and ocean circulation. It is also the core software of the "Earth System Numerical Simulation Device", a major national scientific and technological infrastructure, with a total of approximately 2.7 million lines. The program code is called the "Earth Laboratory".
Baidu Feiyuan Propeller Team builds protein-small molecule docking conformation prediction model HelixDock
The Baidu Flying Paddle Propeller team built a protein-small molecule docking conformation prediction model HelixDock by building large-scale simulation data sets and upgrading geometry-based neural networks, which greatly improved the accuracy of conformation prediction. For more results, see the HelixDock article: https://arxiv.org/abs/2310.13913
Flying propeller access address: https://paddlehelix.baidu.com/
Domestic research team discloses carbon emission prediction method and system based on hybrid machine learning
A domestic research team disclosed a carbon emission prediction method and system based on hybrid machine learning. The data set is processed through the target combination model to obtain the carbon emission prediction results. Among them, the target combination model realizes the calculation of single variables through target calculation weights. The optimal weighted combination of time series forecasting and multi-variable driving factor models takes into account the advantages of each model and improves the accuracy of carbon emission forecasting. Visit address: https://cprs.patentstar.com.cn/Search/Detail?ANE=9HFF9IBA9GDC5BCA8GBA9FHE9AHA8BCA9DFB9CFF9GFF7BDA
Apple chip "exclusive customized version" machine learning framework open source
MLX is a machine learning framework designed specifically for Apple chips (click for detailed explanation) . It aims to support efficient training and deployment of models on Apple chips while ensuring user-friendliness. Its design concept is simple, referring to frameworks such as NumPy, PyTorch, Jax and ArrayFire, and includes key functions such as lazy computation and dynamic graph construction. Access address: https://github.com/ml-explore/mlx/tree/main/examples
Scientific research results DANTE: Oriented to large-scale optoelectronic intelligent computing
Training large-scale optoelectronic neural networks with dual-neuron optical-artificial learning
* Source : Nature Communications
* Field : neural network, optoelectronic intelligence
* Author : Fang Lu’s research group, Department of Electronic Engineering, Tsinghua University
The research team proposed an optical-artificial dual-neuron learning architecture (DuAl-Neuron opTical-artificial lEarning, DANTE) for large-scale optoelectronic intelligent computing. Among them, optical neurons accurately model the light field calculation process, artificial neurons use lightweight mapping functions to establish jump connections to assist gradient propagation, and global artificial neurons and local optical neurons are iteratively optimized using an alternating learning mechanism to ensure learning effectiveness. At the same time, it greatly reduces the spatio-temporal complexity of training, making it possible to train larger and deeper optoelectronic neural networks.
Read the original text: https://www.nature.com/articles/s41467-023-42984-y
Convolutional neural network framework PtyNet: Synchrotron radiation massive data processing
An efficient ptychography reconstruction strategy through fine-tuning of large pre-trained deep learning model
* Source : iScience
* Field : data mining, convolutional neural network
* Author : Chinese Academy of Sciences team
The research team developed a convolutional neural network framework called PtyNet to recover accurate projections of objects from X-ray Ptychography experimental data. With the support of powerful computing clusters, PtyNet can quickly obtain data from synchrotron radiation sources for training and quickly reconstruct images of users' experimental data.
Read the original article: https://doi.org/10.1016/j.isci.2023.108420
Prediction of multiple conformations via sequence clustering and AlphaFold2
Predicting multiple conformations via sequence clustering and AlphaFold2
* Source : Nature
* Field : Bioinformatics
* Author : A research team from Brandeis University and the Howard Hughes Medical Institute, Harvard University, and the University of Cambridge
The research team clustered multiple sequence alignments (MSAs) by sequence similarity, allowing AF2 to sample alternating states of known metamorphic proteins with high confidence. At the same time, the researchers used the AF-Cluster method to study the evolutionary distribution of the predicted structures of the morphing protein KaiB5 and found that the predictions of both conformations were distributed in clusters of the KaiB family.
Read the original article:
https://www.nature.com/articles/s41586-023-06832-9
ProRefiner: Inverse protein folding design model
ProRefiner: an entropy-based refining strategy for inverse protein folding with global graph attention
* Source : Nature Communications
* Field : biological genes, deep learning
* Author : Research team from The Chinese University of Hong Kong, Zhijiang Laboratory, Huawei Noah's Ark Laboratory and Nanjing Medical University
The research team introduced ProRefiner, a memory-efficient global graph attention model that can make full use of denoising context, and demonstrated the applicability of ProRefiner in redesigning transposon-related transposase B (TnpB) properties, 6 out of 20 proposed variants showed improved gene editing activity.
Read the original text: https://www.nature.com/articles/s41467-023-43166-6
KPGT: Self-supervised learning framework
A knowledge-guided pre-training framework for improving molecular representation learning
* Source : Nature Communications
* Field : biomolecules, drug discovery
* Author : Tsinghua University, West Lake University and Zhijiang Laboratory research team
The research team proposed Knowledge-guided Pre-training of Graph Transformer (KPGT), a self-supervised learning framework that provides improved, generalizable and Robust molecular property predictions. The KPGT framework integrates a graph Transformer specifically designed for molecular graphs and a knowledge-guided pre-training strategy to fully capture the structural and semantic knowledge of molecules. Read the original text: https://www.nature.com/articles/s41467-023-43214-1
Activity review
CoRL conference ends, best paper and best system paper announced
The 2023 Conference on Robot Learning (CoRL) was held in Atlanta, USA last month. According to official data, 199 papers from 25 countries were selected for CoRL this year. Popular topics include manipulation, reinforcement learning, etc.
Among them, the best paper award is "Distilled Feature Fields Enable Few-Shot Language-Guided Manipulation"
* 作者：William Shen, Ge Yang, Alan Yu, Jensen Wong, Leslie Pack Kaelbling, Phillip Isola
* Institution : MIT CSAIL, IAIFI
Read the original text: https://openreview.net/forum?id=Rb0nGIt_kh5
For details of other awards, please see the official website: https://www.corl2023.org/awards
NASSMA 2022 AI4Science seminar sharing
The seminar was jointly organized by NASSMA, Mohammed VI University of Technology, Google Deepmind and other institutions. Currently, the seminar’s speech PPT and live replay are online.
The above is all the content to be shared in this issue of "Science AI Weekly"~
If you have the latest research results about AI for Science, first-hand information about companies, etc., please leave a message to "reveal the news".