Xu Shilu
At the end of July, DeepMind said AlphaFold predicted more than 200 million protein structures.
·“ Some people are saying that AlphaFold is open source, our country does not have to do it, this is a big mistake. First, their so-called open source is only the surface structure open source, the underlying technology is not open source..." Ma Jianpeng, dean of the Institute of Complex Systems and Multi-scale Research of Fudan University, said in an interview.
In Ma Jianpeng's view, AlphaFold is another core technology of the card neck.
Almost the entire protein universe was predicted
On July 28, 2022, the British company DeepMind said that AlphaFold has predicted almost all protein structures in the world, and in just one year, the data in their protein structure database has expanded from 2 million to more than 200 million.
AlphaFold is also an intelligent artificial system built by DeepMind after AlphaGo (Alpha Dog), which is mainly used to predict protein structure. So, what are the basic principles of AlphaFold operation?
It uses machine learning algorithms, configures deep learning neural networks, and is trained in the Protein Three-Dimensional Structure Database (PDB) and hundreds of thousands of experimentally determined protein structures and sequences in other databases.
After training, AlphaFold first looks for the relevant gene sequence in the database when faced with a new gene sequence, and then after a lot of calculation and comparison, predicts the protein 3D structure corresponding to the new gene sequence, and its prediction results have considerable accuracy.
Back a year ago, in July 2021, early in AlphaFold's release of the protein database and open source code, DeepMind used AlphaFold to predict 98.5% of human protein structure, and the results were published in the journal Nature.
In the eyes of many experts, AlphaFold is just a "glimpse" of the future, bringing biology into a digital age, and it will change the current state of biomedical research around the world.
Currently, DeepMind is partnering with the Drugs for Neglected Disease Initiative (DNDI) to advance their findings. At the same time, DeepMind also made structural predictions of organisms identified by the World Health Organization (WHO) as a high priority study, helping scientists further study persistent diseases such as leprosy and schistosomiasis.
DeepMind founder Demis Hassabis. Visual China Infographic
In response, Demis Hassabis, founder and CEO of DeepMind, lamented, "AlphaFold has had an incredible impact on some of our biggest global challenges. We hope this expanded database will help countless scientists do their important work and open up entirely new avenues for scientific discovery. ”
AlphaFold has facilitated the scientific research work of scholars in the field of life sciences around the world, and mainland scholars are committed to exploring "AlphaFold that belongs to China". But how is it done, is there its own original technology?
How to build a Chinese AlphaFold?
Domestic attention to AlphaFold is the last two years.
At the end of 2020, AccutarBio announced the completion of nearly $100 million in financing to accelerate the development of innovative drugs by using computational models trained on massive amounts of data through deep learning and physical modeling to replace biological and chemical experiments.
Dr. Fan Jie, founder of AccutarBio, said that their vision is to use AI to comprehensively improve the speed and efficiency of innovative drug research and development, launch drug products with global patents, and strive to use the power of disruptive technologies to continuously change the drug discovery industry.
Teams working on cross-border integration of AI are not unique. In December 2021, Beijing Shenshi Technology launched a protein structure prediction tool, Uni-Fold, and open sourced training code and inference code for scientists and entrepreneurs from all walks of life to test and use.
Because AlphaFold only open source model inference code, and does not open source training code, researchers can only use it according to the process designed by AlphaFold, and cannot adjust its training process to migrate to more application scenarios.
Ou Weinan, an academician of the Chinese Academy of Sciences and a professor at Peking University, said, "Although DeepMind open sourced the inference code, the training technology of the model is the core competitiveness."
Professor Ma Jianpeng of Fudan University.
Similarly, in December 2021, Professor Ma Jianpeng's team at the Institute of Complex Systems Multiscale Studies at Fudan University, in collaboration with the Shanghai Artificial Intelligence Laboratory, published an article in briefings in Bioinformatics entitled "OPUS-Rota4: agradient-based protein side-chain modeling framework assisted by." deeplearning-based predictors" paper, which briefly describes the results of the algorithms they developed.
The research team developed the OPUS series algorithm with independent intellectual property rights, which can be used to predict the three-dimensional structure of protein backbones and side chains, and it is worth mentioning that the protein side chain prediction algorithm is OPUS-Rota4 algorithm, which is more accurate than AlphaFold.
Specifically, the researchers used AlphaFold to obtain the predicted structures of 15 proteins and used different methods to remodel their side chains based on the predicted main chain structure. The results show that OPUS-Rota4 results are significantly better than other sidechain modeling methods and are closer to native conformation than the sidechains predicted by AlphaFold2.
Performance of different sidechain modeling methods on CASP14-AF2(15).
Not long ago, another domestic AI pharmaceutical company Huashen Zhiyao announced that it has made an important progress in the field of ai and life science combination, and developed a new technology in protein structure prediction - OmegaFold.
In general, research on the use of AI to promote the development of life sciences is blossoming everywhere in China. Major enterprise and university teams are constantly advancing the cross-border integration of AI and biopharmaceuticals, and they are trying to catch up with DeepMind's AlphaFold. Obviously, the source of innovation in this field is not in our hands, how to maintain the advantage?
"We have to start with the underlying technology"
Protein structure prediction is no longer a new field, scientists have been doing it for more than fifty years, but it was not until the advent of AlphaFold that research in this field showed breakthrough results.
In the eyes of many industry experts, its emergence depends to some extent on individual scientists to exert their creativity. However, the mainland also has certain advantages in the development of this field, which is computer system engineering. The engineering of scientific problems, AlphaFold is also one of the most typical examples.
At present, the mainland has not been able to completely surpass alphaFold's core technology, and it may take some time to achieve a truly "0 to 1" breakthrough.
Ma Jianpeng said in an interview, "Some people are saying that AlphaFold is open source, and our country does not need to do it, which is a big mistake." First, their so-called open source is only open source of the surface structure, and the underlying technology is not open source. Second, you take it and you can only do their stuff, you can't improve it, just like an aero engine you can buy, can you go further? ”
Ma Jianpeng believes that we must start from the underlying technology and master the core algorithm.
Resources
1. Benchmark AlphaFold2! Deep Potential Technology released the Uni-Fold protein structure prediction tool and open sourced the training code. Deep Potential Technology.
https://mp.weixin.qq.com/s/Zhn3HJpLnznLM1bpu6D1ew
2. OPUS-Rota4: a gradient-based protein side-chain modeling framework assisted by deep learning-based predictors. Briefings in Bioinformatics.
https://academic.oup.com/bib/article/23/1/bbab529/6461160
3. Surpassing Google's "AlphaFold2" to provide a weapon for new drug research and development: Fudan Complex System Multi-scale Research Institute team publishes a new protein sidechain prediction results. Fudan University.
https://mp.weixin.qq.com/s/idWh_IO66Nhry4LfeITmIA
https://www.sohu.com/a/510336879_629135
4. The world's first! Huashen Zhiyao completes the last piece of the puzzle of single sequence protein structure prediction. Huashen Wisdom Medicine.
https://mp.weixin.qq.com/s/DY3I13k_9QsIpCmyvQBlPg
5. AlphaFold reveals the structure of the protein universe. DeepMind.
https://www.deepmind.com/blog/alphafold-reveals-the-structure-of-the-protein-universe
6. Dialogue with president of Fudan Institute of Complex Systems Multiscale Research Institute: Why protein structure prediction is a breakthrough. Sina Technology.
https://finance.sina.com.cn/tech/2021-11-01/doc-iktzqtyu4671024.shtml
7. Finally, an answer to the question: AI — what is it good for?. VOX.
https://www.vox.com/future-perfect/2022/8/3/23288843/deepmind-alphafold-artificial-intelligence-biology-drugs-medicine-demis-hassabis
Editor-in-Charge: Wu Yuewei Photo Editor: Zhang Tongze
Proofreader: Shi Gong