The present and future of deep learning in radiology

European Journal of Radiology, Journal Year: 2019, Volume and Issue: 114, P. 14 - 24

Published: March 2, 2019

Language: Английский

---

Cancer Drug Response Profile scan (CDRscan): A Deep Learning Model That Predicts Drug Effectiveness from Cancer Genomic Signature

Scientific Reports, Journal Year: 2018, Volume and Issue: 8(1)

Published: June 5, 2018

In the era of precision medicine, cancer therapy can be tailored to an individual patient based on genomic profile a tumour. Despite ever-increasing abundance data, linking mutation profiles drug efficacy remains challenge. Herein, we report Cancer Drug Response scan (CDRscan) novel deep learning model that predicts anticancer responsiveness large-scale screening assay data encompassing 787 human cell lines and structural 244 drugs. CDRscan employs two-step convolution architecture, where mutational fingerprints molecular drugs are processed individually, then merged by 'virtual docking', in silico modelling treatment. Analysis goodness-of-fit between observed predicted response revealed high prediction accuracy (R

Language: Английский

---

SMILES-BERT

Published: Sept. 4, 2019

With the rapid progress of AI in both academia and industry, Deep Learning has been widely introduced into various areas drug discovery to accelerate its pace cut R&D costs. Among all problems discovery, molecular property prediction one most important problems. Unlike general applications, scale labeled data is limited prediction. To better solve this problem, methods have started focusing on how utilize tremendous unlabeled improve performance small-scale data. In paper, we propose a semi-supervised model named SMILES-BERT, which consists attention mechanism based Transformer Layer. A large-scale used pre-train through Masked SMILES Recovery task. Then pre-trained could easily be generalized different tasks via fine-tuning. experiments, proposed SMILES-BERT outperforms state-of-the-art three datasets, showing effectiveness our unsupervised pre-training great generalization capability model.

Language: Английский

---

A multilayer multimodal detection and prediction model based on explainable artificial intelligence for Alzheimer’s disease

Scientific Reports, Journal Year: 2021, Volume and Issue: 11(1)

Published: Jan. 29, 2021

Abstract Alzheimer’s disease (AD) is the most common type of dementia. Its diagnosis and progression detection have been intensively studied. Nevertheless, research studies often little effect on clinical practice mainly due to following reasons: (1) Most depend a single modality, especially neuroimaging; (2) are usually studied separately as two independent problems; (3) current concentrate optimizing performance complex machine learning models, while disregarding their explainability. As result, physicians struggle interpret these feel it hard trust them. In this paper, we carefully develop an accurate interpretable AD model. This model provides with decisions along set explanations for every decision. Specifically, integrates 11 modalities 1048 subjects from Disease Neuroimaging Initiative (ADNI) real-world dataset: 294 cognitively normal, 254 stable mild cognitive impairment (MCI), 232 progressive MCI, 268 AD. It actually two-layer random forest (RF) classifier algorithm. first layer, carries out multi-class classification early patients. second applies binary detect possible MCI-to-AD within three years baseline diagnosis. The optimized key markers selected large biological measures. Regarding explainability, provide, each global instance-based RF by using SHapley Additive exPlanations (SHAP) feature attribution framework. addition, implement 22 explainers based decision trees fuzzy rule-based systems provide complementary justifications in layer. Furthermore, represented natural language form help understand predictions. designed achieves cross-validation accuracy 93.95% F1-score 93.94% 87.08% F1-Score 87.09% resulting system not only accurate, but also trustworthy, accountable, medically applicable, thanks provided which broadly consistent other medical literature. proposed can enhance understanding processes providing detailed insights into different risk.

Language: Английский

---

MODE: automated neural network model debugging via state differential analysis and input selection

Published: Oct. 26, 2018

Artificial intelligence models are becoming an integral part of modern computing systems. Just like software inevitably has bugs, have bugs too, leading to poor classification/prediction accuracy. Unlike model cannot be easily fixed by directly modifying models. Existing solutions work providing additional training inputs. However, they limited effectiveness due the lack understanding misbehaviors and hence incapability selecting proper Inspired debugging, we propose a novel debugging technique that works first conducting state differential analysis identify internal features responsible for then performing input selection is similar program in regression testing. Our evaluation results on 29 different 6 applications show our can fix effectively efficiently without introducing new bugs. For simple (e.g., digit recognition), MODE improves test accuracy from 75% 93% average whereas state-of-the-art only improve 85% with 11 times more time. complex object able over 91% minutes few hours, fails bug or even degrades

Language: Английский

---

Deep Learning with Microfluidics for Biotechnology

Trends in biotechnology, Journal Year: 2018, Volume and Issue: 37(3), P. 310 - 324

Published: Oct. 6, 2018

Language: Английский

---

Computational Strategies for Dissecting the High-Dimensional Complexity of Adaptive Immune Repertoires

Frontiers in Immunology, Journal Year: 2018, Volume and Issue: 9

Published: Feb. 21, 2018

The adaptive immune system recognizes antigens via an immense array of antigen-binding antibodies and T-cell receptors, the repertoire. interrogation repertoires is high relevance for understanding response in disease infection (e.g., autoimmunity, cancer, HIV). Adaptive receptor repertoire sequencing (AIRR-seq) has driven quantitative molecular-level profiling thereby revealing high-dimensional complexity sequence landscape. Several methods computational statistical analysis large-scale AIRR-seq data have been developed to resolve order understand dynamics immunity. Here, we review current research on (i) diversity, (ii) clustering network, (iii) phylogenetic (iv) machine learning applied dissect, quantify compare architecture, evolution, specificity repertoires. We summarize outstanding questions immunology propose future directions systems towards coupling with discovery immunotherapeutics, vaccines, immunodiagnostics.

Language: Английский

---

A Berkeley View of Systems Challenges for AI

arXiv (Cornell University), Journal Year: 2017, Volume and Issue: unknown

Published: Jan. 1, 2017

With the increasing commoditization of computer vision, speech recognition and machine translation systems widespread deployment learning-based back-end technologies such as digital advertising intelligent infrastructures, AI (Artificial Intelligence) has moved from research labs to production. These changes have been made possible by unprecedented levels data computation, methodological advances in learning, innovations software architectures, broad accessibility these technologies. The next generation promises accelerate developments increasingly impact our lives via frequent interactions making (often mission-critical) decisions on behalf, often highly personalized contexts. Realizing this promise, however, raises daunting challenges. In particular, we need that make timely safe unpredictable environments, are robust against sophisticated adversaries, can process ever amounts across organizations individuals without compromising confidentiality. challenges will be exacerbated end Moore's Law, which constrain amount store process. paper, propose several open directions systems, security address help unlock AI's potential improve society.

Language: Английский

---

Precision Oncology: The Road Ahead

Trends in Molecular Medicine, Journal Year: 2017, Volume and Issue: 23(10), P. 874 - 898

Published: Sept. 5, 2017

Language: Английский

---

Deep Learning in Pharmacogenomics: From Gene Regulation to Patient Stratification

Pharmacogenomics, Journal Year: 2018, Volume and Issue: 19(7), P. 629 - 650

Published: April 26, 2018

This Perspective provides examples of current and future applications deep learning in pharmacogenomics, including: identification novel regulatory variants located noncoding domains the genome their function as applied to pharmacoepigenomics; patient stratification from medical records; mechanistic prediction drug response, targets interactions. Deep encapsulates a family machine algorithms that has transformed many important subfields artificial intelligence over last decade, demonstrated breakthrough performance improvements on wide range tasks biomedicine. We anticipate future, will be widely used predict personalized response optimize medication selection dosing, using knowledge extracted large complex molecular, epidemiological, clinical demographic datasets.

Language: Английский

---