Session J28: New Ways of Seeing with Data Science

Data science and video games

A video game is a microcosm combining tracking capabilities (movement, attack, defense, scoring), marketplace data (microtransactions), and social networking (chat, user-generated content, friend networks). We use machine learning to touch each of these facets, hoping to improve engagement, profitability, and fun! Projects include cheat detection, churn mitigation, marketing optimization, and toxicity detection.   

Modeling complex physical systems with big data and machine-learning

Modeling complex physical systems continues to be a major challenge in many fields of science and technology. In this talk we present a general framework, which advances such endeavor. Specifically, we describe the development of a machine-learning based model blending architecture for statistically combining multiple models for improving the accuracy of an application-specific forecast or prediction. Most importantly, we demonstrate that in addition to parameters to be predicted or forecasted, including additional state parameters which collectively define a situation as machine learning input provides enhanced accuracy for the blended result. ANOVA (analysis of variance) shows that the error of individual models can have a substantial dependence on the situation. The machine-learning architecture effectively reduces such situation dependence error and thus produces more accurate results compared to conventional multi-model ensemble approaches based on simplistic equally or unequally weighted model averaging. The framework is first illustrated in the context of weather forecasting, which is arguably one of the hardest problems in physics, not only because of the complexity and scale of the problem but also given the additional complication that predictions/forecasts have to be made using limited observation at sparse locations. We will also demonstrate that the framework is applicable to other applications.
  

Machine learning for seeing and hearing more

Modern machine learning has had some of its greatest successes in perceptual problems like image and sound understanding. Extracting relevant information from such high dimensional input is frequently the challenge in scientific data understanding as well. I’ll survey some exciting results from the Google Accelerated Science team in the areas of cellular imaging for biomedical research, extracting surprising results from human clinical imaging, and disease staging from auditory signals. Collectively these show the promise of further machine assistance in making sense of scientific data, with a great deal of exciting work still to come.   

Machine Learning in Scanning probe microscopy: accelerating imaging, enhancing resolution and Bayesian methodologies for theory-experiment matching

Scanning probe microscopy (SPM) [1], including a variety of modes in atomic force microscopy and scanning tunneling microscopy, have yielded tremendous insights into the functioning of materials at the atomic and nanoscale, and remain a major tool of nanoscience in fields as diverse as condensed matter physics, catalysis and surface chemistry, and molecular biology. Despite the proliferation of SPMs, most are still wedded to traditional paradigms of limited data acquisition, analysis by hand-crafted simple models, and a decided lack of uncertainty quantification.

In this talk, I will discuss how SPM can be greatly enhanced via careful and tailored use of machine/statistical learning methodologies in every aspect, from data acquisition to real-time analytics to model comparison and selection. I will show that Gaussian process regression and active learning schemes can enable increases in sampling efficiency in large dimensional spaces, enabling new experiments that were previously unfeasible. Moreover, complete information acquisition in conjunction with statistical learning methodologies can enable new SPM techniques with enhanced spectral and spatial resolution that are orders of magnitude faster than existing state of the art techniques [2]. Finally, the use of Bayesian methodologies for model fitting and model selection can enable a rigorous uncertainty-quantified answer regarding the appropriate cantilever dynamics model to employ in analyzing dynamic atomic force microscopy data. These methods that are developed point towards an autonomous future of ‘self-exploring’ SPM systems for physics knowledge generation.

References
[1] “Scanning Probe Microscopy, from Sublime to Ubiquitous,” https://journals.aps.org/prl/scanning-probe-microscopy(2016).
[2] Somnath et al., Nat Commun. 9, 513 (2018)   

Immunotherapy Modeling: Molecular Interaction and Recognition of MHC/peptide/TCR Complexes

Cancer immunotherapy has been among the most promising breakthroughs in oncology, particularly in the case of immune check point inhibitors, however, the effective response rate remains quite low, only about 20-30%. In this talk, I will talk about our recent collaborative work which solves one mystery behind this low response rate with molecular modeling and machine learning techniques. We found that patients with certain HLA genotype (HLA-B44) have consistently higher survival rate, while patients with some other type (HLA-B15) have much poorer survival rates. It’s also shown that patients harboring tumors with very high mutation rates responded disproportionately well to these immune checkpoint inhibitor treatments. Large scale molecular dynamics simulations further reveal that HLA-B15 proteins with poorer therapeutic outcomes had structural appendages (HLA bridges with residues Arg62, Ile66, and Leu163) that closed over the cancer neoantigens with much less flexibility. The same techniques have also been applied to the design and development of vaccines for HIV and T1D, which has been of great interest as well in recent years. With a combined in silico and in vivo approach, we studied the TCR/peptide/HLA interactions from multiple clonotypes specific for a well-defined HIV-1 epitope, and found that effective and ineffective clonotypes bind to the terminal portions of the peptide-HLA through similar salt bridges, but their hydrophobic side-chain packings can be very different, which accounts for the major part of the differences among these clonotypes. Meanwhile, a novel super potent autoantigen has been identified for T1D, which opens new door for potential T1D vaccine. Together with state-of-the-art free energy perturbation calculations for point mutations on antigens, our results clearly indicate a direct structural basis for heterogeneous T cell function.