Machine Transfer Learning Research Articles

A generalizable Cas9/sgRNA prediction model using machine transfer learning with small high-quality datasets

The CRISPR/Cas9 nuclease from Streptococcus pyogenes (SpCas9) can be used with single guide RNAs (sgRNAs) as a sequence-specific antimicrobial agent and as a genome-engineering tool. However, current bacterial sgRNA activity models struggle with accurate predictions and do not generalize well, possibly because the underlying datasets used to train the models do not accurately measure SpCas9/sgRNA activity and cannot distinguish on-target cleavage from toxicity. Here, we solve this problem by using a two-plasmid positive selection system to generate high-quality data that more accurately reports on SpCas9/sgRNA cleavage and that separates activity from toxicity. We develop a machine learning architecture (crisprHAL) that can be trained on existing datasets, that shows marked improvements in sgRNA activity prediction accuracy when transfer learning is used with small amounts of high-quality data, and that can generalize predictions to different bacteria. The crisprHAL model recapitulates known SpCas9/sgRNA-target DNA interactions and provides a pathway to a generalizable sgRNA bacterial activity prediction tool that will enable accurate antimicrobial and genome engineering applications.

Quantitative detection of α1-acid glycoprotein (AGP) level in blood plasma using SERS and CNN transfer learning approach

Surface-enhanced Raman spectroscopy (SERS) is a highly sensitive tool in medical diagnostics and bioanalysis fields, aimed at the qualitative detection of relevant biomolecules. However, quantitative SERS analysis of complex (bio)samples is a more challenging and, in many cases, almost impossible task, requiring functional SERS substrates or advanced spectral data analysis. In this work, we propose the combination of a functional SERS substrate, capable of trapping target biomolecules, with CNN transfer learning for quantitative detection of the relevant α1-acid glycoprotein (AGP, also known as orosomucoid) in human serum. As a SERS substrate, the plasmonic gold grating was functionalized with boronic acid moieties to entrap target AGP. The functionality of the substrate was tested on two model solutions: a solution containing saccharides as competing molecules and human serum with added AGP, which is close to real samples. The convolution neural network (CNN) was previously trained on a huge number of (bio)samples. Then CNN transfer learning was used to quantify AGP concentration in model samples, as well as in human serum. Developed strategy is able to identify the alarming increase of AGP concentration in an express and medically decentralized way, on short time and under lack of spectral data. Generally, the proposed combination of SERS and machine transfer learning could be expanded to a range of alternative cases, where the collection of real samples is restricted and can be substituted by the measurements of similar model systems, without loss of analysis reliability.

Accurate temperature prediction with small absorption spectral data enabled by transfer machine learning

It is of great interest to explore the possibility of applying machine learning methods for tunable diode laser absorption spectroscopy (TDLAS). Conventional supervised machine learning could be used for accurate determination of gas properties such as temperature or concentration in TDLAS. However, it becomes quite challenging when there is only a small amount of measured data. In this work, we propose a transfer machine learning (TML) model for accurate temperature prediction from a small amount of measured data. In the experiment, a two-line thermometer is developed by exploiting H2O absorption lines centered at 1392 and 1371 nm. Based on the limited data of measured laser transmission and an easily obtained large amount of calculated absorption spectra, the proposed model attempts to reduce the feature difference between these two distinct types of datasets and then leverages the large labeled calculated data to build an accurate predictor for the unlabeled measured data. The TML method can achieve a temperature prediction with a mean absolute error of 0.02-0.77 K over the temperature range of 288-338 K and a Pearson correlation coefficient of -0.084 to 0.981.