DeepGuide architecture
DeepGuide uses a CAE to derive a reduced-dimensionality representation of the underlying distribution of sgRNA sequences in the whole genome. The autoencoder is composed of an encoder (6 layers) and a decoder (6 layers). The objective of the unsupervised training is to infer the internal weight so the input layer to the encoder is as close as possible to the output layer of the decoder. The CAE encoder has two Conv1D layers of 20 filters and 40 filters, respectively, one MaxPooling1D layer, one AveragePooling1D layer, and two BatchNormalization layers (see Supplementary Table 2 for the order). A rectified linear activation function (ReLU) is used as activation and the Glorot uniform initializer is used to initialize the convolutional filters. The layer regularizer for the encoder is L2 with a value of 10E−4. The decoder has the same structure as the encoder but uses UpSampling1D instead of MaxPooling, and UpSampling1D instead of AveragePooling1D. The layer regularizer in the decoder is again L2 with a value of 10E−4. The loss function for training is the binary cross-entropy, and Adam is the optimizer with a learning rate of 10E−3. A batch size of 64 and 200 epochs are used for training (no early stopping).
The encoder in the second network has the same structure as the encoder in the CAE (see Supplementary Table 3). The initial configuration of the network downstream of the encoder uses one flatten layer, three fully connected layers (fc8, fc9, fc10) of 80 neurons, 40 neurons, and 40 neurons, respectively. The feature map for layer pool6 is 7 × 40, which is 280 dimensional. The feature map for the first fully connected layer (fc8) is 280 × 80 = 22400 dimensional. The feature map for the second and third fully connected layers (fc9 and fc10) are 80 × 40 = 3200 and 40 × 40 = 1600 dimensional, respectively. Layer mult11 is a multiplication layer that combines sequence and nucleosome occupancy features. ReLU is the activation and Glorot uniform initializer is used to initialize the convolutional filters. The second network is trained for 150 epochs using back-propagation; if the value of loss function does not improve for 15 consecutive epochs the training is terminated.
The third fully connected network is used to provide DeepGuide with nucleosome occupancy data. The nucleosome occupancy for each sgRNA is a floating-point value in [0,1]. The third network uses one fully connected layer with 40 units to expand the one-dimensional nucleosome occupancy value to a 40-dimensional vector, to match the dimensionality of the output layer of the second network. Sequence and nucleosome data are merged by performing an element-wise multiplication between the output layer of the second network and the output layer of the third network. When DeepGuide is used in “classification mode” (i.e., binary output) the activation function is a sigmoid; when DeepGuide is used in “regression mode” (i.e., CS output), the activation function is linear.
Note that following the ablation analysis, only two fully connected layers (and no multiplication layer) are used for Cas12a; similarly, only one fully connected layer connected to the multiplication layer is used for Cas9.
DeepGuide training and pre-training
For the pre-training step of the CAE all k-mers from the Y. lipolytica genome were extracted using a sliding window of 1 bp. For Cas9 the input length was 28 bp, which includes the length of each possible spacer (20 bp), plus 3 bp for a PAM sequence, and 2 bp upstream and downstream for context. For Cas12a, 32-mers were used to account for the 25 bp spacer, a 4 bp PAM, 1 bp of context upstream of the PAM, and 2 bp of context downstream of the spacer (see Fig. 4b). These unlabeled sgRNA datasets contained over 20 million k-mers each. sgRNA sequences were converted into a numerical representation using one-hot encoding, that is, each sgRNA was converted into a 4 × n dimensional binary matrix where n is the length of the guide.
The training data to DeepGuide consisted of sgRNA sequences, their nucleosome occupancy score, and their CS values. sgRNA sequences were one-hot encoded, while nucleosome occupancy data were processed as explained in the “Nucleosome occupancy analysis” subsection below. CS scores were produced as explained in the “CS analysis” subsection also provided below.
When the pre-training concluded, the internal weights of the CAE were used to initialize the encoder in the second network. The second network was trained via back-propagation using either ~45,000 sgRNAs for Cas9 or ~58,000 sgRNA for Cas12a, each with their associated CS value. In all, 60% of these guides were used for training, 20% for validation, and 20% for testing. The training step not only allowed the inference of the weights for the fully connected layers downstream of the encoder but also fine-tuned the weights of the encoder. As explained in the section “Ablation analysis of DeepGuide” (main text) the pre-training step helped the supervised learning to converge faster and improved the prediction performance.
Supplementary Fig. 6 illustrates the loss curve for training and validation of the CNN without pre-training and with pre-training as a function on the number of training epochs. Observe that in the CNN without pre-training the difference between training and validation loss function starts increasing after about 20 epochs. In contrast, for the CNN with pre-training, the training and validation curves of the loss function are overlapping after about 30 epochs. This indicates that the pre-training prevents the network from overfitting and helps the network to generalize better.
sgRNA library design
Custom Matlab scripts were used to design an LbCas12a sgRNA library with ~8-fold coverage of all protein-coding sequences annotated in the Y. lipolytica PO1f parent strain genome, CLIB89 [https://www.ncbi.nlm.nih.gov/assembly/GCA_001761485.1]26. A list of 25 nucleotide (nt) sgRNAs with a TTTV (V = A/G/C) PAM were identified in both the top and the bottom strand of the coding sequence of each gene (CDS). A second list containing all possible 25nt sgRNAs with a TTTN PAM from the top and bottom strands of all 6 chromosomes in Y. lipolytica was also generated and used to test for sgRNA uniqueness. The uniqueness test was carried out by comparing the first 14nt of each sgRNA in the first list to the first 14nt of every sgRNA in the second list. If a sequence occurred more than once, the sgRNA was identified as non-unique and excluded from consideration. The sgRNAs that passed the test for uniqueness were then picked in an unbiased manner, with even representation from the top and bottom strands when possible, starting from the 5’ end of the CDS. Six-hundred and fifty-one sgRNAs of random sequence confirmed to not target in the genome were also designed using a similar methodology but with more stringent criteria for uniqueness (i.e., first 10 nt were not found anywhere in the genome). A detailed procedure of sgRNA design for both Cas9 and Cas12a is provided in ref. 42 and additional data on the Cas9 guide design criteria are provided in ref. 8. Briefly, for Cas9 sgRNAs the first version of sgRNA Designer27 was used to identify the top predicted guides for every CDS, these guides were filtered for uniqueness, and the top six unique guides were selected.
Microbial strains and culturing
The parent yeast strain used in this study was Y. lipolytica PO1f with genotype MatA, leu2-270, ura3-302, xpr2-322, axp-2. The PO1f Cas9 and the PO1f Cas12a strains were constructed by integrating UAS1B8-TEF(136)-Cas9-CYCT and UAS1B8-TEF(136)-LbCpf1-CYCT expression cassettes into the A08 locus43. The PO1f Cas9 ku70 and PO1f Cas12a ku70 strains were constructed by disrupting KU70 using CRISPR-Cas9 as previously described23. All strains used in this study are listed in Supplementary Table 6. All plasmid construction and propagation were conducted in E. coli TOP10. Cultures were conducted in Luria-Bertani (LB) broth with 100 mg L−1 ampicillin at 37 °C in 14 mL polypropylene tubes, at 225 r.p.m. Plasmids were isolated from E. coli cultures using the Zymo Research Plasmid Miniprep Kit.
Plasmid construction
All plasmids and primers used in this work are listed in Supplementary Tables 7 and 8. To create the LbCas12a sgRNA expression plasmid (pLbCas12ayl), we first added a second direct repeat sequence at the 5’ of the polyT terminator in pCpf1_yl (see ref. 44). This was done to ensure that library sgRNAs could end in one or more thymine residues without being construed as part of the terminator. To make this change, pCpf1_yl was first linearized by digestion with SpeI. Subsequently, primers ExtraDR-F and ExtraDR-R were annealed and this double-stranded fragment was used to circularize the vector (NEBuilder® HiFi DNA Assembly) For integrating LbCas12a, pHR_A08_LbCas12a was constructed by digesting pHR_A08_hrGFP (Addgene #84615) with BssHII and NheI, and the LbCas12a fragment was inserted using the New England BioLab (NEB) NEBuilder® HiFi DNA Assembly Master Mix. The LbCas12a fragment was amplified along with the necessary overlaps by PCR using Cpf1-Int-F and Cpf1-Int-R primers from pLbCas12ayl. Successful cloning of the entire fragment was confirmed with sequencing primers A08-Seq-F, A08-Seq-R, Tef-Seq-F, Lb1-R, Lb2-F, Lb3-F, Lb4-F, and Lb5-F. To create the Cas12a sgRNA genome-wide library expression plasmid (pLbCas12ayl-GW) the UAS1B8-TEF- LbCas12a-CYC1 fragment was removed from pLbCas12ayl with the use of XmaI and HindIII restriction enzymes. Subsequently, the primers BRIDGE-F and BRIDGE-R were used to circularize the vector, and the M13 forward primer was used to ensure the correct assembly of the construct.
To conduct the validation experiments of predicted CS values by DeepGuide, four genes with easily screenable phenotypes were selected and 10 sgRNAs (five highly active and five with poor activity) targeting each of these genes for Cas9 and Cas12a were selected and cloned for individual disruption experiments. All 40 Cas9 sgRNAs with required overlaps for cloning were purchased from a commercial vendor (IDT-DNA) as single-stranded primers and assembled into pCRISPRyl (Addgene #70007) after linearizing the vector with AvrII, using NEBuilder® HiFi DNA Assembly. In a similar manner, the 40 Cas12a sgRNAs with necessary overlaps were cloned into pLbCas12ayl, after linearizing the vector with SpeI. These primers are also included in Supplementary Table 8.
sgRNA library cloning
The LbCas12a library targeting the protein-coding genes in PO1f was ordered as an oligonucleotide pool from Agilent Technologies Inc. and cloned in-house using the Agilent SureVector CRISPR Library Cloning Kit (Part Number G7556A). The backbone vector (pLbCas12ayl-GW) was first linearized by PCR using the primers InversePCR-F and InversePCR-R, DpnI digested, cleaned up using Beckman AMPure XP SPRI beads, and transformed into E.coli TOP10 cells to verify minimal contamination from the circularized plasmid. Library oligos were amplified by PCR using the primers OLS-F and OLS-R for 15 cycles as per vendor instructions using Q5 high fidelity polymerase and cleaned up using the AMPure XP beads. The linearized backbone and the amplicons were combined in 4 replicate reactions of sgRNA library cloning that were carried out as per vendor instructions and pooled prior to bead cleanup. Two amplification bottles containing 1 L of LB media and 3 g of library-grade low gelling agarose were prepared, autoclaved, and cooled to 37 °C. Eighteen replicate transformations of the cloned library were conducted using Agilent’s ElectroTen-Blue cells (Catalog #200159) via electroporation (0.2 cm cuvette, 2.5 kV, 1 pulse). Cells were recovered and with a 1 hr outgrowth in SOC media at 37 °C (2% tryptone, 0.5% yeast extract, 10 mM NaCl, 2.5 mM KCl, 10 mM MgCl2, 10 mM MgSO4, and 20 mM glucose.) The transformed E. coli cells were then inoculated into two amplification bottles and grown for 2 days until colonies were visibly suspended in the matrix. Colonies were recovered by centrifugation and subject to a second amplification step by inoculating an 800 mL LB culture. After 4 hr, the cells were collected, and the pooled plasmid library was isolated using the ZymoPURE II Plasmid Gigaprep Kit (Catalog #D4202) yielding ~2.4 mg of plasmid DNA containing the Cas12a sgRNA library. The library was subject to a NextSeq run to test for fold coverage of individual sgRNA and skew.
Yeast transformation and screening
Transformation of Y. lipolytica with the sgRNA plasmid library was done using a previously described method with slight modifications8. Briefly, 3 mL of YPD was inoculated with a single colony of the strain of interest and grown in a 14 mL tube at 30 °C with shaking at 200 RPM for 22-24 hours (final OD ~30). Cells were pelleted by centrifugation (6,300 g) and washed with 1.2 mL of transformation buffer (0.1 M LiAc, 10 mM Tris (pH=8.0), 1 mM EDTA). To these resuspended cells, 36 µL of ssDNA mix (8 mg/mL Salmon Sperm DNA, 10 mM Tris (pH = 8.0), 1 mM EDTA), 180 µL of β-mercaptoethanol mix (5% β-mercaptoethanol, 95% triacetin), and 8 µg of plasmid library DNA were added, mixed via pipetting, and incubated for 30 min at room temperature. After incubation, 1800 µL of PEG mix (70% w/v PEG (3350 MW)) was added and mixed via pipetting, and the mixture was incubated at room temperature for an additional 30 min. Cells were then heat shocked for 25 min at 37 °C, washed with 25 mL of sterile milliQ H2O, and used to inoculate 50 mL of SD-leu media for screening experiments. Dilutions of the transformation (0.01% and 0.001%) were plated on solid SD-leu media to calculate transformation efficiency. Three biological replicates of each transformation were performed for each condition. Transformation efficiency for each replicate is presented in Supplementary Table 9. Details of the Cas9 library are provided in ref. 8.
Screening experiments were conducted in 50 mL of liquid media in a 250 mL baffled flask (220 rpm shaking, 30 °C). Cells first reached confluency after 2 days of growth (OD600 ~12), at which time 200 µL (which includes a sufficient number of cells for approximately 500-fold library coverage) was used to inoculate 25 mL of fresh media. The cells were again subcultured upon reaching confluency at day 4 for the growth screen, and the experiment was halted after 6 days of growth. At each time point (i.e., days 2, 4, and 6), 1 mL of culture was removed and treated with DNase I (New England Biolabs; 4 and 25 µL of DNaseI buffer) for 1 h at 30 °C to remove any extracellular DNA. Cells were isolated by centrifugation at 4500 × g and the resulting cell pellets were stored at −80 °C for future analysis.
Library isolation and sequencing
Growth screen samples were thawed and resuspended in 400 µL sterile, milliQ H2O. Each cell suspension was split into two, 200 µL samples, and plasmids from each sample were isolated using a Zymo Yeast Miniprep Kit (Zymo Research). Splitting into separate samples here was done to accommodate the capacity of the Yeast Miniprep Kit. The split samples from a single pellet were then pooled, and plasmid copy number was quantified using quantitative PCR with qPCR-GW-F and qPCR-GW-R and SsoAdvanced Universal SYBR Green Supermix (Biorad). Each pooled sample was confirmed to contain at least 107 plasmids.
To prepare samples for next-generation sequencing, isolated plasmids were subjected to PCR using forward (ILU1-F, ILU2-F, ILU3-F, ILU4-F) and reverse primers (ILU(1-12)-R) containing all necessary barcodes and adapters for next-generation sequencing using the Illumina platform (Supplementary Table 10). Schematics of the amplicons from the Cas9 and Cas12a experiments submitted for NGS are pictured in Supplementary Fig. 7. At least 0.2 ng of plasmids (approximately 3 × 107 plasmid molecules) were used as templates, and PCR reactions were amplified for 16 cycles and not allowed to proceed to completion to avoid amplification bias. PCR product was purified using SPRI beads and tested on the bioanalyzer to ensure the correct length. Samples were pooled in equimolar amounts and submitted for sequencing on a NextSeq 500 at the UCR IIGB core facility.
Generating sgRNA read counts from raw reads
Next-generation sequencing reads were processed using the Galaxy platform45. First, read quality was assessed using FastQC v0.11.8. The reads were then demultiplexed using Cutadapt v1.16.6, trimmed using Trimmomatic v0.38, and mapped to each sgRNA using a combination of Bowtie 2 v2.4.2, and custom MATLAB scripts for counting bowtie alignments and naïve exact matching. Parameters used for each method are provided in Supplementary Table 11 and MATLAB scripts are provided as part of the GitHub link found below in the section “Data availability”. Supplementary Table 12 provides further information correlating the NCBI SRA file names to the information needed for demultiplexing the readsets. Analysis of the CRISPR-Cas12a growth screens revealed that five sgRNAs were not present in the sequencing data. A pairwise comparison between normalized read abundances for biological replicates was done to verify consistency, see Supplementary Fig. 2 and Supplementary Table 1.
CS analysis
The CS associated with each guide was determined by taking the log2 of the ratio of normalized read counts of the control condition to the normalized read counts of the treatment condition. The control condition was taken as the normalized read counts at the end of the growth screen in a strain without Cas12a or Cas9. The treatment condition included constitutively expressed Cas9 or Cas12a with disrupted KU70. Normalized counts were taken as the total number of reads for a given sgRNA divided by the total reads for the corresponding sample. If no reads were identified for a given sgRNA, a pseudo-count of one was added to the read count to facilitate subsequent calculations. In all cases, normalized read counts for each biological replicate were averaged together to produce an average normalized read count and associated standard deviation for each sgRNA. All normalized read counts and CS values are provided in Supplementary Data 3 and 4.
Nucleosome occupancy analysis
To account for genomic features, specifically nucleosome occupancy, we determined an average normalized occupancy score (ranging from 0 to 1) for every target locus using previously published MNase-Seq coverage data46 (Supplementary Data 5). Per base nucleosome occupancy scores were summed up for each sgRNA, averaged, and normalized to a value between 0 and 1 by taking its ratio to the highest averaged value. This information was integrated into DeepGuide via a separate FCCN, the first step of which was to convert the one-dimensional occupancy data into an 80-dimensional real vector using a fully connected layer with 80 neurons. Using element-wise multiplication, the output of this layer was combined with the output of the last fully connected layer of the CS-predicting CNN to generate CS predictions that account for guide sequence, genomic context, and nucleosome occupancy.
Validation of predicted sgRNA for Cas9 and Cas12a
Four genes with easily screenable phenotypes, including MEF1, CAN1, MGA1, and RAS2 were selected for the validation of predicted sgRNA CS values (Supplementary Fig. 3). Gene sequences and the per base nucleosome occupancy of these genes were provided as input to the DeepGuide algorithm. As output DeepGuide predicted a CS value for each sgRNA of a given gene. sgRNAs were sorted from best to worst based on the predicted CS value from sequence-only (for Cas12a) and sequence plus nucleosome occupancy (for Cas9). The top 5 and bottom 5 sgRNA from the list were tested for editing efficiency.
To screen for RAS2 and MGA1 gene disruption, cultures with CRISPR plasmids growing in SD-Leu were diluted and plated in triplicate on YPD to obtain greater than 50 colonies on each plate. After two days of growth at 30 °C, the number of smooth colonies was counted and expressed as a fraction of the total colonies on the plate. For disruption of the CAN1 gene, cultures were similarly diluted and plated on YPD to obtain single colonies. Thirty colonies in triplicate were then randomly selected and streaked on SD-leu agar media supplemented with 50 mg L−1 of L-canavanine. Colonies that grew on SD with canavanine were identified as positive for CAN1 disruption. To screen for MFE1, cultures were similarly plated, and 30 colonies from each transformation were randomly selected and streaked on SD-Oleic acid and dotted on YPD. Growth on YPD but not on SD-Oleic acid indicated MFE1 disruption. Screening of MFE1 was done on agar plates containing SD media supplemented with oleic acid as the sole carbon source (SD oleic acid; 0.67% Difco yeast nitrogen base without amino acids, 0.079% CSM (Sunrise Science, San Diego, CA), 2% agar 0.4% (v/v) Tween 20, and 0.3% (v/v) oleic acid).
Reporting summary
Further information on research design is available in the Nature Research Reporting Summary linked to this article.
