Science 12.08.2022_downmagaz.net

How Musk’s monster rocket could Vocal simplification led to Protecting biodiversity with tools transform space science p. 702 speech complexity pp. 706 & 760 from the insurance sector p. 714 $15 12 AUGUST 2022 science.org GUIDANCE SYSTEM Death’s-head moths correct course based on an internal “compass” p. 764

CALL FOR PAPERS Plant Phenomics is a Science Partner Journal published in affiliation with the State Key Laboratory of Crop Genetics & Germplasm Enhancement, Nanjing Agricultural University (NAU) and distributed by the American Association for the Advancement of Science (AAAS). Plant Phenomics publishes novel research that advances both in eld and indoor plant phenotyping, with focus on data acquisition systems, data management, data interpretation into structural or functional traits, integration into process based or machine learning based models, and connects phenomics to applications and other research domains. Submit your research to Plant Phenomics today! Learn more: spj.sciencemag.org/plantphenomics The Science Partner Journals (SPJ) program was established by the American Association for the Advancement of Science (AAAS), the non-pro t publisher of the Science family of journals. The SPJ program features high-quality, online-only, Open-Access publications produced in collaboration with international research institutions, foundations, funders, and societies. Through these collaborations, AAAS furthers its mission to communicate science broadly and for the bene t of all people by providing top-tier international research organizations with the technology, visibility, and publishing expertise that AAAS is uniquely positioned to o er as the world’s largest general science membership society. Learn more at spj.sciencemag.org @SPJournals @SPJournals ARTICLE PROCESSING CHARGES WAIVED UNTIL 2022

Produced by the Science Advertorial Chong Tang, deputy director of research and development, BGI Tech. Innovative data visualization for gene-related research A biomedical engineering Ph.D. student at the University of California, Davis, Yongheng related diseases. “I don’t need to go through all the calculations,” she explains. “Dr. PHOTO: PROVIDED BY BGI GENOMICS Wang had a problem: He wanted to identify the top 10 targets from thousands of genes Tom is the bridge. With just a couple of clicks, I can see the graphs, which are easy affected by a drug he developed. The process to mine and extract information would to read.” No programming skills are needed for analyzing these genetic sequences, normally take a huge investment of time and effort, switching back and forth between she adds. “I really like this very visual Venn diagram comparing the different gene different programming languages and software packages. Wang was delighted to expressions in samples.” learn of a solution: an innovative data visualization system called “Dr. Tom” from BGI Genomics, a leading global provider of genomic sequencing and proteomics services, This visualization feature is designed to be interactive so it is easier to bounce headquartered in Shenzhen, China. ideas around the team, Tang says. He spent 5 years designing, testing, and fine- tuning Dr. Tom to serve his community, and now Dr. Tom’s web-based interface is Developed by a team of expert scientists and bioinformaticians at BGI Genomics, Dr. inspiring innovative research and reducing researchers’ headaches. Tom is a powerful tool for analysis, visualization, and interpretation of many types of data, including RNA sequencing (RNA-Seq), long noncoding RNA-Seq, microRNA- “The BGI culture is a culture of scientists and for scientists.” Tang says. The Seq, whole-genome bisulfite sequencing, single-cell RNA-Seq, and proteomics data. organization considers it so vital to advance scientific discovery that it is gifting this software to the world. Having an easy learning curve, Dr. Tom is also versatile, and most importantly, allows researchers without backgrounds in bioinformatics to quickly, efficiently, Making a mark and reliably create meaningful visualizations. “The intuitive interface allows us to select genes of interest and rank differentially expressed genes based on their fold Relying on cutting-edge sequencing and bioinformatics technology, BGI changes,” says Wang. “Dr. Tom covers all the practical functionalities. I think a high Genomics has provided scientists and researchers in 100 countries and regions school student could do it.” with integrated solutions across a broad range of applications spanning basic life sciences research, clinical research in human health, and agriculture and Designed by scientists for scientists biodiversity preservation and sustainability. Dr. Tom is a project fueled by the passion of Chong Tang, deputy director of research Since its launch in 2018, Dr. Tom has enchanted more than 15,000 users from over and development at BGI Tech. A biochemist and molecular biologist with expertise in 20 countries and regions. “More than 60 published papers in the field of disease software engineering, Tang’s extensive experience helped him uncover the specific treatment, developmental regulation, immunity, and environmental adaptation have pain point experienced by his research colleagues around the world: Given the shown Dr. Tom to be a valuable and important tool in addition to any institution’s tsunami of 'omics data, there is a critical need for a flexible, easy, and turnkey data own internal data curation and analysis efforts,” says Tang. analysis program. “I designed it in such a way that scientists will like it,” he says. BGI Genomics has partnered with academic institutions to offer workshops on Dr. Tom is designed to be flexible and simple for users with limited computing leveraging this system effectively, and will be expanding these opportunities to knowledge, notes Margot Maurer, research associate in the Wolfson Centre for more countries and regions in the future. Age-Related Diseases at King’s College London. Maurer uses Dr. Tom to analyze the transcriptomics differences in sensory ganglia to help her investigate RNA and age- Sponsored by



























































































RESEARCH | REVIEW Fig. 3. Deep-learning applications for seismological tasks. The size of each circle is scaled to the number of published papers for each application and color-coded according to the average size of training data. Horizontal axes represent the average percentage of observational and recorded data versus synthetic training data. Fig. 4. Hierarchical distribution of studies in our database. (A and B) Studies are distributed according to (A) the seismological applications and (B) used DNNs. Fine tuning of trained models, especially those developments of specialized and hybrid DNN quential nature of seismic data makes RNNs trained with synthetic data, is popular in seis- architectures composed of a mixture of neural- appealing for many seismological tasks. More mology, and we expect that meta-learning will layer types. RNNs are popular in tasks in recently, neural networks with attention mech- be a growth area in seismology. which temporal relations in data play a crucial anisms, dilated CNNs, and physics-informed role in the modeling. Network architectures DNNs are gaining momentum. CNNs dominate seismological models across with contracting encoders and expanding de- many sectors (Fig. 4B) because of their effec- coders such as U-Net and Autoencoders are Explainable deep learning tiveness for automatic feature extraction and highly suitable for modeling seismic data across sparse representation learning of seismic data. many applications. GANs are another popular Despite the widespread application of deep Some characteristics of seismic data such as approach for seismic data processing. Among learning in seismology and its breakthrough their limited bandwidth, sequential nature, RNNs, long short-term memory (LSTM) units performance in multiple areas, the black-box or data structure (for example, 3D and 4D are the most popular network type. The se- nature remains a source of skepticism. Accept- imaging in active seismology) have motivated ing indecipherable models on their own terms Mousavi and Beroza, Science 377, eabm4470 (2022) 12 August 2022 6 of 11

RESEARCH | REVIEW can contribute to scientific goals of the com- kernels to build an interpretable model of physics-inspired neural network architectures munity (139), but presenting the properties anisotropic spatial dependency in seismic- have been developed for seismic FWI (111, 145). of these models in understandable terms to ity patterns. The parameters of these kernels Sun et al. (111) used a recurrent architecture humans (interpretability) and revealing the are reparameterized by DNNs and learned to model time-dependent dynamics of seismic underlying reasons for their output decision from training data by using an imitation wave propagation. To enforce physical rules (explainability) (140) would allay such skep- learning approach. Such an interpretable for wave propagation in the model, they set ticism and lead to wider adoption and new model can help uncover patterns [and hopefully up LSTM-like cells—including some physical insights. causal relationships (148)] in data. operators, which take the wavefield at a past instant as input—estimated the shot record at Currently, most deep-learning approaches Incorporating geophysics into DNNs the current instant as output, and saved the in seismology concern building a model based modeled wavefield in the memory cell for cal- on known input-output pairs to predict the Deep-learning approaches that incorporate culating the next time step. In these models, outputs corresponding to previously unseen physical laws have gained momentum in the the trainable parameters of the DNN can be inputs without interpretability or explainabil- machine learning community (149) and a viewed as a reparameterization of an Earth ity. In most of these cases, it is not possible to growing number of implementations in seis- model. The forward propagation of informa- explain why specific outcomes were obtained mology. The main idea is to integrate data and tion through this representation resembles from an algorithmic point of view nor what mathematical physics (domain knowledge) the forward propagation of a seismic wave- else these models can tell us. models, even if only partially understood. The field through a heterogeneous medium. Thus, key objectives are to provide interpretability training such a network and updating its Visualizing the learned kernels at differ- and explainability for a deep-learning method, weights by using DNN back-propagation and ent layers of a CNN is the main approach to regularize them in the presence of missing based on the misfit between the simulated used in some of the seismological studies or noisy data, and to enhance their out-of- and observed seismic data amounts to back- (93, 107, 141, 142) to understand something distribution generalization by restricting the propagation of the residual wavefield and about the representations a deep learning solution space to physically plausible solutions. under certain assumptions is equivalent to the model has learned and how it drives the out- optimization process in conventional gradient- put; however, the extracted multiscale feature Domain knowledge can be incorporated into based FWI. This can result in a more interpret- representations from seismic data are too the training data, the hypothesis space, and the able model in which the latent representation complex for direct visualization to provide training procedure of a deep-learning model is physically meaningful. interpretable and understandable insights. (Fig. 5). Among these, incorporating the do- main knowledge into training data through Physical constraints can be introduced into Although it may not be possible to elucidate data generation and augmentation by using the training process either by incorporating how a deep-learning model works, gaining numerical simulations is the most commonly them into the loss function (implicit approach), useful information about learned models is used approach in seismology for active seismic by explicit inclusion of a numerical model- possible by using interpretation tools such as data processing (47, 150), wave simulation (68), ing into the training loop, or both. The first backward propagation techniques, saliency and velocity estimation (91). These models are approach is used more recently in seismology maps, or heatmaps to reveal relevant patterns usually fine tuned through some examples of and primarily for forward problems such as in the input based on feature importance or field data. simulation of the pressure wavefield (68) or relevance scores (143, 144). These interpret- estimation of first arrival times (77). In these able representations of the input can be used The hypothesis space can be constrained studies, a form of the wave equation (68) or along with domain knowledge to explain the with domain knowledge through design of Eikonal equation (77) is used to penalize the output. specialized neural network architectures based on the characteristics of the problem. Such Mousavi et al. (12), for example, used atten- tion mechanisms in their detection and picker Fig. 5. Explainable DNN. The domain knowledge, physics, can be incorporated into the training data through DNN. They used a hierarchy of attention mod- (A) data simulation, (B) the hypothesis space by designing specialized neural network architectures, ules to gain insight into the interaction of task- and (C) the training procedure of a deep-learning model by enforcing consistency with governing equations. specific decoder branches and where and on The architecture design based on domain knowledge provides some transparency and intepretability of what the DNN focused. Doing so gave a degree the model. The interpretable representation of the input data along with some domain knowledge can be used of interpretability to their results. Attention to explain the output. was used in (54) to build an interpretable deep- learning model for seismic facies analysis. They used 3D spatial-spectral attention maps to reveal the relations between the seismic spec- tral response (input) and the geological forma- tions (output). Liao et al. (22) used a backward propagation technique to compare different deep-learning models for phase-picking on the basis of the learned representations from data. In addition to explaining the outcomes, designing an interpretable deep-learning model is possible in which the model itself and its components became explainable by using domain knowledge (111, 145). In this approach, the network architecture is de- signed to provide insight by promoting or restricting specific attributes of the data being modeled (146, 147). Zhu et al. (87), for ex- ample, used a mixture of Gaussian diffusion Mousavi and Beroza, Science 377, eabm4470 (2022) 12 August 2022 7 of 11