Journal: Nucleic Acids Research

Article Title: Programmable cross-ribosome-binding sites to fine-tune the dynamic range of transcription factor-based biosensor

doi: 10.1093/nar/gkaa786

Figure Lengend Snippet: Workflow of CLM-RDR development. First, the dynamic range of biosensors and the sequences of their related cRBSs were analyzed to establish an RBS design principle ( Step 1 ). Based on this principle, a cRBSs library was designed and synthesized ( Step 2 ) using DNA microarray. Subsequently, the library was divided into five sub-libraries (I–V) based on the fluorescence intensity of sfGFP measured by FACS ( Step 3 ). Finally, to predict the dynamic range of biosensors with the given cRBSs, NGS and CNN model were employed to analyze the sequences of cRBSs in sub-libraries I–V and establish the CLM-RDR, respectively ( Step 4 ). RBSn (NNNAGNNN), RBSs of cdaR ; RBSm (NNGGAGNN), and RBSs of sfgfp ; N = A, T, C, G.

Article Snippet: Nielsen and Voigt used a deep learning based convolutional neural network (CNN) containing 42 364 plasmid DNA sequences datasets from Addgene to predict the lab-of-origin of a DNA sequence, and achieved 70% prediction accuracy and rapid analyses of DNA sequence information to guide the attribution process and understand the measures ( ).

Techniques: Synthesized, Microarray, Fluorescence

Journal: bioRxiv

Article Title: Fine-tuning biosensor dynamic range based on rational design of cross-ribosome-binding sites in bacteria

doi: 10.1101/2020.01.27.922302

Figure Lengend Snippet: CLM-RDR verification for three genetically encoded biosensors. Sixteen cRBSs were randomly selected for biosensor modification and comparison of the observed and predicted dynamic ranges. The CLM-RDR performed well in predicting the dynamic ranges of ( a ) glucarate biosensor, ( b ) arabinose biosensor, and ( c ) glycolate biosensor. I–V represent the five sub-libraries of cRBSs. The black diagonal denotes y = x. ( d) Structure of P araB -based arabinose sensor. Pc represents the constitutive promoter that controls transcription of the regulatory protein AraC. P araB is an inducible promoter containing the AraC-binding DNA sequence. Blunt-end arrows denote repression. ( e ) Structure of P glcD -based glycolate sensor. P ffs indicates the constitutive promoter that controls transcription of the regulatory protein GlcC. P glcD is a constitutive promoter that controls the transcription of the reporter sfGFP. In the absence of glycolate, GlcC remained as a non-functional regulatory protein, whereas in the presence of glycolate, the regulatory protein GlcC and glycolate bound to the activator GlcC-glycolate, which in turn bound to the upstream activation site (UAS) of the promoter PglcD, thus enhancing transcription and expression of sfgfp . Pointed arrows indicate activation. ( f ) Detailed illustration of 16 cRBSs and three biosensors. Solid circle: glucarate biosensor; solid triangle: arabinose biosensor; solid diamond: glycolate biosensor.

Article Snippet: Nielsen and Voigt used a deep learning based convolutional neural network (CNN) containing 42,364 plasmid DNA sequences datasets from Addgene to predict the lab-of-origin of a DNA sequence, and achieved 70% prediction accuracy and rapid analyses of DNA sequence information to guide the attribution process and understand the measures .

Techniques: Modification, Binding Assay, Sequencing, Functional Assay, Activation Assay, Expressing

Journal: Nature Communications

Article Title: Deep learning to predict the lab-of-origin of engineered DNA

doi: 10.1038/s41467-018-05378-z

Figure Lengend Snippet: Plasmid dataset and machine learning approach. a Machine learning plasmid attribution strategy. b Plasmid publication dates across the dataset. c Depositing labs ranked by how many plasmids they have submitted in the dataset. A minimum cutoff of nine plasmids was enforced for training (dashed line). d Plasmids ranked by how much associated DNA sequence (in bp) has been submitted. DNA sequence information is categorized as Partial Depositor (purple), Full Depositor (green), Partial Repository (red), and Full Repository (blue). The summed DNA sequence length across all four categories is also shown (black). Plasmid order not maintained between the five curves

Article Snippet: A convolutional neural network was trained on the Addgene plasmid dataset that contained 42,364 engineered DNA sequences from 2230 labs as of February 2016.

Techniques: Plasmid Preparation, Sequencing

Journal: Nature Communications

Article Title: Deep learning to predict the lab-of-origin of engineered DNA

doi: 10.1038/s41467-018-05378-z

Figure Lengend Snippet: Convolutional neural network accuracy. a Convolutional neural network (CNN) architecture. DNA sequences are converted to 16,048 × 4 matrices, where the identity of each nucleotide is converted to a one-hot vector. This input is scanned by 128 convolutional filters ( f 1 – f 128 ) each with a width, w , of 12 nucleotide positions. Per-position nucleotide filter weights are converted to Boltzmann factors and visualized using a heatmap (Methods). The maximum activation for each filter, max (f k ) , across the entire input sequence is taken. Activations are fed through two fully connected layers, which generates neural activity predictions for each lab, A (Name), before behind converted to probabilities using the softmax function, P (Name). The lab prediction is taken as the highest softmax probability. Batch normalization layers are not shown. b Training accuracy (gray) and validation accuracy (black) per epoch for the chosen architecture. Cross-validation accuracy was computed after training (dashed line). c Output prediction rank of the actual lab-of-origin for plasmids in the cross-validation set. d Neural network softmax probabilities (left panel) for a Christopher Voigt lab plasmid ( pVRa38_1322 ) and a Pamela Silver lab plasmid ( pPS1622 ). Labs with the highest probabilities are labeled. Normalized distribution of pre-softmax neuron activity (“Activity”, right panel) for the plasmids of interest. Arrows highlight the activity for labeled labs at left. The vertical dashed lines mark the origin. e Normalized distribution of activity for 10 4 random DNA sequences with length 3685 nt. f P value distributions for random DNA sequences for lengths 8000, 3685, and 1000 nt (from left to right). Empirical data (solid lines) and fits to P ( A > x ) = 1−exp(−exp(− λ ( x − μ ))) (dashed lines) are shown. g Distribution fit steepness parameter ( λ ) versus plasmid length with a trend of λ = 0.59–6.2 × 10 −6 x (dashed line). h Distribution fit offset parameter ( μ ) versus plasmid length with a trend of μ = 7.5–3.4 × 10 −4 x (dashed line)

Article Snippet: A convolutional neural network was trained on the Addgene plasmid dataset that contained 42,364 engineered DNA sequences from 2230 labs as of February 2016.

Techniques: Plasmid Preparation, Activation Assay, Sequencing, Activity Assay, Labeling