Interpretable Machine Learning Approach Research Articles

A statistical and machine learning approach to the study of astrochemistry.

In order to obtain a good understanding of astrochemistry, it is crucial to better understand the key parameters that govern grain-surface chemistry. For many chemical networks, these crucial parameters are the binding energies of the species. However, there exists much disagreement regarding these values in the literature. In this work, a Bayesian inference approach is taken to estimate these values. It is found that this is difficult to do in the absence of enough data. The Massive Optimised Parameter Estimation and Data (MOPED) compression algorithm is then used to help determine which species should be prioritised for future detections in order to better constrain the values of binding energies. Finally, an interpretable machine learning approach is taken in order to better understand the non-linear relationship between binding energies and the final abundances of specific species of interest.

Identifying Descriptors for Promoted Rhodium-Based Catalysts for Higher Alcohol Synthesis via Machine Learning.

Rhodium-based catalysts offer remarkable selectivities toward higher alcohols, specifically ethanol, via syngas conversion. However, the addition of metal promoters is required to increase reactivity, augmenting the complexity of the system. Herein, we present an interpretable machine learning (ML) approach to predict and rationalize the performance of Rh-Mn-P/SiO2 catalysts (P = 19 promoters) using the open-source dataset on Rh-catalyzed higher alcohol synthesis (HAS) from Pacific Northwest National Laboratory (PNNL). A random forest model trained on this dataset comprising 19 alkali, transition, post-transition metals, and metalloid promoters, using catalytic descriptors and reaction conditions, predicts the higher alcohols space-time yield (STYHA) with an accuracy of R 2 = 0.76. The promoter's cohesive energy and alloy formation energy with Rh are revealed as significant descriptors during posterior feature-importance analysis. Their interplay is captured as a dimensionless property, coined promoter affinity index (PAI), which exhibits volcano correlations for space-time yield. Based on this descriptor, we develop guidelines for the rational selection of promoters in designing improved Rh-Mn-P/SiO2 catalysts. This study highlights ML as a tool for computational screening and performance prediction of unseen catalysts and simultaneously draws insights into the property-performance relations of complex catalytic systems.

An interpretable machine learning approach based on DNN, SVR, Extra Tree, and XGBoost models for predicting daily pan evaporation

Evaporation is an important hydrological process in the water cycle, especially for water bodies. Machine Learning (ML) models have become accurate and powerful tools in predicting pan evaporation. Meanwhile, the “black-box” character and the consistency with the physical process can decrease the practical implication of ML models. To overcome such limitations, we attempt to develop an interpretable based-ML framework to predict daily pan evaporation using Extra Tree, XGBoost, SVR, and Deep Neural Network (DNN) ML models using hourly climate datasets. To that end, we integrated and employed the Shapely Additive explanations (SHAP), Sobol-based sensitivity analysis, and Local Interpretable Model-agnostic Explanations (LIME) to evaluate the interpretability of the models in predicting daily pan evaporation, at Sidi Mohammed Ben Abdellah (SMBA) weather station, in Morocco. The validation results of the models showed that the developed models are accurate in reproducing the daily pan evaporation with NSE ranging from 0.76 to 0.83 during the validation phase. Furthermore, the interpretability results of the ML models showed that the air temperature (Ta), solar radiation (Rs), followed by relative humidity (H) are the most important climate variables with inflection points of the Ta_median, Ta_mean, Rs_sum, H_mean, and w_std are 17.42 °C, 17.65 °C, 3.8 kw.m−2, 69.59%, and 1.25 m s−1, sequentially. Overall, the interpretability of the models showed a good consistency of the ML models with the real hydro-climatic process of evaporation in a semi-arid environment. Hence, the proposed methodology is powerful in enhancing the reliability and transparency of the developed models for predicting daily pan evaporation. Finally, the proposed approach is new insights to reduce the ‘‘Black-Box’’ character of ML models in hydrological studies.

In-Depth Understanding of Pedestrian–Vehicle Near-Crash Events at Signalized Intersections: An Interpretable Machine Learning Approach

This study used a pedestrian-involved near-crash database and adopted an interpretable machine learning framework using SHapley Additive exPlanations (SHAP) to understand the factors associated with critical pedestrian-involved near-crash events. The results indicate that pedestrians with a relatively higher walking speed are more likely to be involved in critical near-crash events. Furthermore, critical pedestrian-involved near-crash events are highly associated with vehicles with driving speeds of less than 10 mph. A higher pedestrian volume is highly associated with critical near-crash events with left-turn vehicles. It is possible that a higher pedestrian volume increases the occurrence of jaywalking behavior or encourages more pedestrians to step into the crosswalk when they should not. By contrast, a higher pedestrian volume is highly associated with non-critical near-crash events with right-turn vehicles. Right-turn vehicles often expect that there will be pedestrians crossing, and a higher volume of pedestrian traffic increases a driver’s awareness and caution while turning. The study also found that a longer signal cycle is highly associated with critical near-crash events when the pedestrian volume is low, while a relatively short signal cycle length is highly associated with critical near-crash events when the pedestrian volume is high. During non-peak hours, pedestrians have less tolerance for a relatively longer signal cycle. Moreover, a relatively shorter signal cycle length at peak hours will limit the number of pedestrians that can cross during a cycle and encourage the possibility of pedestrians jaywalking or stepping onto the crosswalk when they should not.

Understanding the effect of contextual factors and decision making on team performance in Twenty20 cricket: an interpretable machine learning approach

Understanding the effect of contextual factors and decision making on team performance in Twenty20 cricket: an interpretable machine learning approach

Toward Mental Effort Measurement Using Electrodermal Activity Features.

The ability to monitor mental effort during a task using a wearable sensor may improve productivity for both work and study. The use of the electrodermal activity (EDA) signal for tracking mental effort is an emerging area of research. Through analysis of over 92 h of data collected with the Empatica E4 on a single participant across 91 different activities, we report on the efficacy of using EDA features getting at signal intensity, signal dispersion, and peak intensity for prediction of the participant’s self-reported mental effort. We implemented the logistic regression algorithm as an interpretable machine learning approach and found that features related to signal intensity and peak intensity were most useful for the prediction of whether the participant was in a self-reported high mental effort state; increased signal and peak intensity were indicative of high mental effort. When cross-validated by activity moderate predictive efficacy was achieved (AUC = 0.63, F1 = 0.63, precision = 0.64, recall = 0.63) which was significantly stronger than using the model bias alone. Predicting mental effort using physiological data is a complex problem, and our findings add to research from other contexts showing that EDA may be a promising physiological indicator to use for sensor-based self-monitoring of mental effort throughout the day. Integration of other physiological features related to heart rate, respiration, and circulation may be necessary to obtain more accurate predictions.

Machine Learning Reveals a General Understanding of Printability in Formulations Based on Rheology Additives.

Hydrogel ink formulations based on rheology additives are becoming increasingly popular as they enable 3‐dimensional (3D) printing of non‐printable but biologically relevant materials. Despite the widespread use, a generalized understanding of how these hydrogel formulations become printable is still missing, mainly due to their variety and diversity. Employing an interpretable machine learning approach allows the authors to explain the process of rendering printability through bulk rheological indices, with no bias toward the composition of formulations and the type of rheology additives. Based on an extensive library of rheological data and printability scores for 180 different formulations, 13 critical rheological measures that describe the printability of hydrogel formulations, are identified. Using advanced statistical methods, it is demonstrated that even though unique criteria to predict printability on a global scale are highly unlikely, the accretive and collaborative nature of rheological measures provides a qualitative and physically interpretable guideline for designing new printable materials.

Towards expert–machine collaborations for technology valuation: An interpretable machine learning approach

Although technology valuation has benefited considerably from recent advances in machine learning technology, the results of prior studies in this field are of limited use in practice because they rely solely on black box models whose internal mechanisms are hidden. We develop an analytical framework for successful expert–machine collaborations for technology valuation using interpretable machine learning that makes a model's behaviors and predictions understandable to humans. First, a technological characteristics–economic value matrix is constructed using patent and technology transaction databases. Second, machine learning models are developed to examine the nonlinear and complex relationships between the technological characteristics and economic value of technologies. Third, the performance of the machine learning models is assessed using quantitative metrics. Finally, the SHapley Additive exPlanation method is applied to the best-performing model to explain which technological characteristics influence the economic value of technologies. By these means, we investigate the importance of the features of technological characteristics (and their interactions) in technology valuation and offer theoretical and practical implications of the analysis results. A case study of the technologies registered in the Office of Technology Licensing at Stanford University confirms that our framework is a useful complementary tool for technology valuation.

Interpretable Machine Learning Approach for Identifying the Tip Sharpness in Atomic Force Microscopy

Atomic force microscopy (AFM) is routinely used with indentation techniques to characterize the plastic deformation of materials. The accurate quantification of the features associated with the indent, which is used to quantify the hardness and indentation deformation mechanisms, depends on the sharpness of the AFM tip used for imaging. However, identifying the tip-sharpness of an atomic force microscope requires non-trivial measurements. Here, using machine learning, we develop a model to predict the tip sharpness of the AFM cantilever directly from the indent images. Further, we employ explainable machine learning models, such as integrated gradients and gradient shap, to interpret the features learned by the model. Altogether, we show that machine learning approaches can accelerate experiments by providing non-trivial information about the instrument performance, thereby enabling researchers to perform better quality experiments.

Physically interpretable machine learning for nuclear masses

We present an approach to modeling the ground-state mass of atomic nuclei based directly on a probabilistic neural network constrained by relevant physics. Our physically interpretable machine learning (PIML) approach incorporates knowledge of physics by using a physically motivated feature space in addition to a soft physics constraint that is implemented as a penalty to the loss function. We train our PIML model on a random set of approximately 20% of the atomic mass evaluation (AME) and predict the remaining 80%. The success of our methodology is exhibited by a ${\ensuremath{\sigma}}_{\mathrm{rms}}\ensuremath{\approx}186$ keV match to data for the training set and ${\ensuremath{\sigma}}_{\mathrm{rms}}\ensuremath{\approx}316$ keV for the entire AME with $Z\ensuremath{\ge}20$. We show that our general methodology can be interpreted using feature importance.

Interpretable machine learning approach to analyze the effects of landscape and meteorological factors on mosquito occurrences in Seoul, South Korea

Mosquitoes are the underlying cause of various public health and economic problems. In this study, patterns of mosquito occurrence were analyzed based on landscape and meteorological factors in the metropolitan city of Seoul. We evaluated the influence of environmental factors on mosquito occurrence through the interpretation of prediction models with a machine learning algorithm. Through hierarchical cluster analysis, the study areas were classified into waterside and non-waterside areas, according to the landscape patterns. The mosquito occurrence was higher in the waterside area, and mosquito abundance was negatively affected by rainfall at the waterside. The mosquito occurrence was predicted in each cluster area based on the landscape and cumulative meteorological variables using a random forest algorithm. Both models exhibited good performance (both accuracy and AUROC > 0.8) in predicting the level of mosquito occurrence. The embedded relationship between the mosquito occurrence and the environmental factors in the models was explained using the Shapley additive explanation method. According to the variable importance and the partial dependence plots for each model, the waterside area was more influenced by the meteorological and land cover variables than the non-waterside area. Therefore, mosquito control strategies should consider the effects of landscape and meteorological conditions, including the temperature, rainfall, and the landscape heterogeneity. The present findings can contribute to the development of mosquito forecasting systems in metropolitan cities for the promotion of public health.

The Importance of Being Ernest, Ekundayo, or Eswari: An Interpretable Machine Learning Approach to Name-Based Ethnicity Classification

Name-based ethnicity classification is the task of predicting ethnicity from a name. Ethnicity classification can be a key tool for assessing the fairness of algorithms, demographic studies, and political analysis. While previous state-of-the-art approaches for this task rely on complex neural networks that are difficult to understand, troubleshoot, and tune, we provide an interpretable and intuitive solution that outperforms these (overly) complicated models at a fraction of the computational cost. Using our technique, we can analyze patterns in name-ethnicity databases to show connections between ethnic groups in terms of their overlap of names. We provide techniques to generalize under domain shift, leveraging ‘indistinguishables,’ which are names common to multiple ethnic groups. We provide an application of our method to the estimation of how many political donations for each political party were provided by individuals from various ethnic groups in 2020 leading up to the presidential election.

First person – Gaëlle Letort

ABSTRACT First Person is a series of interviews with the first authors of a selection of papers published in Journal of Cell Science, helping early-career researchers promote themselves alongside their papers. Gaëlle Letort is first author on ‘ An interpretable and versatile machine learning approach for oocyte phenotyping’, published in JCS. Gaëlle conducted the research described in this article while a post-doc in Marie-Emilie Terret and Marie-Hélène Verlhac’s lab at Collège de France, Paris. She is now a Research Engineer (CNRS) in the lab of Department of Developmental and Stem Cell Biology at Institut Pasteur, Paris, investigating mathematical modelling and image analysis applied to biology.

Potential Determinants for Radiation-Induced Lymphopenia in Patients With Breast Cancer Using Interpretable Machine Learning Approach.

Radiation-induced lymphopenia is known for its survival significance in patients with breast cancer treated with radiation therapy. This study aimed to evaluate the impact of radiotherapy on lymphocytes by applying machine learning strategies. We used Extreme Gradient Boosting (XGboost) to predict the event of lymphopenia (grade≥1) and conduced an independent validation. Then, we induced feature attribution analysis (Shapley additive explanation, SHAP) in explaining the XGboost models to explore the directional contribution of each feature to lymphopenia. Finally, we implemented the proof-of-concept clinical validation. The results showed that the XGboost models had rigorous generalization performances (accuracies 0.764 and ROC-AUC 0.841, respectively) in the independent cohort. The baseline lymphocyte counts are the most protective feature (SHAP = 5.226, direction of SHAP = -0.964). Baseline platelets and monocytes also played important protective roles. The usage of taxane only chemotherapy was less risk on lymphopenia than the combination of anthracycline and taxane. By the contribution analysis of dose, we identified that firstly lymphocytes were sensitive to a radiation dose less than 4Gy; secondly the irradiation volume was more important in promoting lymphopenia than the irradiation dose; thirdly the irradiation dose promoted the event of lymphopenia when the irradiation volume was fixed. Overall, our findings paved the way to clarifying the radiation dose volume effect. To avoid radiation-induced lymphopenia, irradiation volume should be kept to a minimum during the planning process, as long as the target coverage is not compromised.

An interpretable machine learning approach to identify mechanism of action of antibiotics

As antibiotic resistance is becoming a major public health problem worldwide, one of the approaches for novel antibiotic discovery is re-purposing drugs available on the market for treating antibiotic resistant bacteria. The main economic advantage of this approach is that since these drugs have already passed all the safety tests, it vastly reduces the overall cost of clinical trials. Recently, several machine learning approaches have been developed for predicting promising antibiotics by training on bioactivity data collected on a set of small molecules. However, these methods report hundreds/thousands of bioactive molecules, and it remains unclear which of these molecules possess a novel mechanism of action. While the cost of high-throughput bioactivity testing has dropped dramatically in recent years, determining the mechanism of action of small molecules remains a costly and time-consuming step, and therefore computational methods for prioritizing molecules with novel mechanisms of action are needed. The existing approaches for predicting bioactivity of small molecules are based on uninterpretable machine learning, and therefore are not capable of determining known mechanism of action of small molecules and prioritizing novel mechanisms. We introduce InterPred, an interpretable technique for predicting bioactivity of small molecules and their mechanism of action. InterPred has the same accuracy as the state of the art in bioactivity prediction, and it enables assigning chemical moieties that are responsible for bioactivity. After analyzing bioactivity data of several thousand molecules against bacterial and fungal pathogens available from Community for Open Antimicrobial Drug Discovery and a US Food and Drug Association-approved drug library, InterPred identified five known links between moieties and mechanism of action.

An interpretable and versatile machine learning approach for oocyte phenotyping.

ABSTRACTMeiotic maturation is a crucial step of oocyte formation, allowing its potential fertilization and embryo development. Elucidating this process is important for both fundamental research and assisted reproductive technology. However, few computational tools based on non-invasive measurements are available to characterize oocyte meiotic maturation. Here, we develop a computational framework to phenotype oocytes based on images acquired in transmitted light. We trained neural networks to segment the contour of oocytes and their zona pellucida using oocytes from diverse species. We defined a comprehensive set of morphological features to describe an oocyte. These steps were implemented in an open-source Fiji plugin. We present a feature-based machine learning pipeline to recognize oocyte populations and determine morphological differences between them. We first demonstrate its potential to screen oocytes from different strains and automatically identify their morphological characteristics. Its second application is to predict and characterize the maturation potential of oocytes. We identify the texture of the zona pellucida and cytoplasmic particle size as features to assess mouse oocyte maturation potential and tested whether these features were applicable to the developmental potential of human oocytes. This article has an associated First Person interview with the first author of the paper.

Onset of heat transfer deterioration caused by pseudo-boiling in CO2 laminar boundary layers

In propulsion and power systems, operating pressures are increasing to improve cycle efficiency, often reaching conditions that approach or exceed the thermodynamic critical pressure of the working fluid. A known phenomenon that occurs at supercritical pressures is supercritical heat transfer deterioration (HTD). Although this phenomenon is well studied, there is still debate about the underlying physics. This paper studies pseudo boiling, the supercritical liquid–gas phase transition, as a possible mechanism that causes HTD, by means of computational fluid dynamics (CFD), boundary layer correlations, and thermodynamic analysis. To this end, we focus on forced laminar convection of supercritical CO2 over a flat plate, excluding alternative factors such as buoyancy or turbulence. In order to accurately represent the strongly non-linear near-critical fluid properties, we use tiny neural networks (TNN) in an interpretable machine learning approach. For use in CFD, TNN combine reference quality accuracy from look-up tables with the memory footprint of equations of state and curve fits. We demonstrate that pseudo boiling theory accurately indicates pressure and temperature regions susceptible to HTD. The simulations show the distinctive boiling curve, with a peak critical heatflux and a Leidenfrost minimum, demonstrating that pseudo boiling is, indeed, a sufficient physical mechanism to cause this effect, in absence of turbulence or boyancy. During HTD, we observe the formation of a supercritical gas film at the wall, with low density, low thermal conductivity, a high compressibility factor, and low viscosity. The boundary of the gaseous film is characterized by a peak in heat capacity, indicative of pseudo boiling, and a sudden transition to a dense supercritical liquid state. This suggests that the subcritical boiling crisis does have a direct representation at supercritical pressures through pseudo boiling, and that pseudo film boiling is a sufficient mechanism to cause HTD.

Regional and Single Nucleotide Correction of Sequence Bias in Chromatin Accessibility Data

Chromatin accessibility assays directly measure which regions of DNA are actively regulated in the genome. Genomic chromatin accessibility data provide the gold standard of spatial and temporal resolution of regulatory landscapes in various cell types and conditions. Accessibility assays depend on enzymatic digestion of DNA and these enzymes are biased with respect to the sequences they recognize and cut. Sequence biases can lead to incorrect conclusions about the nature of the chromatin landscape within an experiment. Previous work from our group has shown that we can use observed k‐mer cut frequencies and k‐mer representation in the genome to correct the majority of this bias. However, these k‐mer scaling approaches did not effectively correct ATAC‐seq data, which assesses chromatin accessibility using the Tn5 transposase enzyme. Here, we find that the oligomeric state of Tn5 and wide sequence bias window precludes effective k‐mer scaling to account for Tn5 sequence bias. Tn5 sequence specificity is influenced by sequence content up to ten bases away from the Tn5 insertion site in both directions. In comparison, Benzonase, Cyanase, MNase, and DNase are primarily influenced by the sequence within 3 bases of the cleavage site. We could not effectively apply k‐mer scaling to ATAC‐seq data because the majority of insertion sites happen at unique 20‐mers in the genome and the method requires many genomic instances per k‐mer. To overcome this limitation, we applied an interpretable machine learning approach, prediction rules ensemble, to correct Tn5‐digested naked DNA. Application of a prediction rules ensemble effectively corrects local (single nucleotide resolution) sequence bias as well as sequence biases influenced by regional GC content. Correction of Tn5 sequence biases is important as the field continues to generate ATAC‐seq data at accelerating rates and use these data to accurately infer regulatory element activity.

An interpretable machine learning approach for evaluating the feature importance affecting lost workdays at construction sites

This study investigates the intensity of accidents by considering the lost workdays based on real data and derived the feature importance quantitatively. Occupational injuries lead to lost workdays for construction workers. The term “lost workdays” refers to a period in which workers cannot perform regular jobs because of severe physical or mental damage from occupational injuries. Many studies focusing on type analysis, accident prediction, and risk management of construction accidents, combined with the intensity and frequency of accidents, have been conducted to prevent and minimize accidents. However, the existing method for measuring the intensity of occupational accidents/injuries is a qualitative method that reflects experts’ opinions, mainly using brainstorming, the Delphi method, and the analytic hierarchy process. Thus, we propose a framework that combines traditional analysis and interpretable machine learning approaches. The random forest model was trained with 11,223 injured worker samples, feature importance and frequency analysis with the chi-squared test, and local interpretable model-agnostic data. According to the results of this study, optimal safety management is possible if appropriate resources are allocated. If the risk of accidents is measured with objective and quantitative data, the prevention and analysis of accidents can be more reliable, and resources can be more efficiently allocated.

A robust and interpretable machine learning approach using multimodal biological data to predict future pathological tau accumulation

The early stages of Alzheimer’s disease (AD) involve interactions between multiple pathophysiological processes. Although these processes are well studied, we still lack robust tools to predict individualised trajectories of disease progression. Here, we employ a robust and interpretable machine learning approach to combine multimodal biological data and predict future pathological tau accumulation. In particular, we use machine learning to quantify interactions between key pathological markers (β-amyloid, medial temporal lobe atrophy, tau and APOE 4) at mildly impaired and asymptomatic stages of AD. Using baseline non-tau markers we derive a prognostic index that: (a) stratifies patients based on future pathological tau accumulation, (b) predicts individualised regional future rate of tau accumulation, and (c) translates predictions from deep phenotyping patient cohorts to cognitively normal individuals. Our results propose a robust approach for fine scale stratification and prognostication with translation impact for clinical trial design targeting the earliest stages of AD.