Designing viral diagnostics with model-based optimization

Abstract

Harnessing genomic data and predictive models will provide activity-informed diagnostic assays for thousands of viruses and offer rapid design for novel ones. Here we develop and extensively validate new algorithms that design nucleic acid assays having maximal predicted detection activity over a virus’s full genomic diversity with stringent specificity. Focusing on CRISPR-Cas13a detection, we test a library of ~ 19,000 guide-target pairs and construct a convolutional neural network that predicts Cas13a detection activity better than other techniques. We link our methods by building ADAPT, an end-to-end system that automatically leverages the latest viral genome data. We designed optimal species-specific assays for the 1,933 vertebrate-infecting viral species within 2 hours for most species and 24 hours for all but 3. ADAPT’s designs are sensitive and specific down to the lineage-level for the range of taxa we tested, including ones that pose challenges involving genomic diversity and specificity. They also exhibit significantly higher fluorescence and lower limits of detection, across a virus’s full spectrum of genomic diversity, than designs from standard techniques. ADAPT is available in an accessible software package and can be applied to other detection technologies to enhance critically-needed viral diagnostic and surveillance efforts.

Deep learning and CRISPR-Cas13d ortholog discovery for optimized RNA targeting

“…279 LwaCas13a guides for KRAS, PPIB and MALAT1 were analyzed and the Pearson correlation coefficient for each positional nucleotide with guide efficiency is shown. C. Correlation of each nucleotide with guide efficiency at each guide position in the LwaCas13a ADAPT dataset (Metsky et al, 2021). 85 perfect match guides from LwaCas13a ADAPT data were analyzed and the Pearson correlation coefficient for each positional nucleotide with guide efficiency is shown.…”

“…Further analysis of motifs in this region revealed a distinct preference for cytosine at position 21 as part of a GW 1-4 C 21 or C 21 W 0-2 G motif with a narrow preference for 50-60% GC content in this window. Analysis of available Cas13a datasets(Abudayyeh et al, 2017;Metsky et al, 2021) did not indicate a similar motif, suggesting that it may be unique to Cas13d.Future work should incorporate more accurate structural prediction for long RNAs, including mRNAs and lincRNAs as those are developed further. We evaluated RNA secondary structure prediction algorithms and found that the LinearFold implementation of the contrafold model(Huang et al, 2019) performed best when compared to Vienna(Lorenz et al, 2011) and Eternafold (Wayment-Steele et al, 2020) in the context of our model (Figure…”

Sensitive and multiplexed RNA detection with Cas13 droplets and kinetic barcoding

“…2A). For this reason, the selection of "good" crRNAs that support efficient Cas13 activity is critical for bulk Cas13-based molecular diagnostics 15 , though how different crRNAs affect the activity of Cas13 is not well understood 16 . We therefore tested crRNAs 11 and 12 in our droplet assay and compared it to the activity of crRNA 4.…”

Equipment-free detection of SARS-CoV-2 and Variants of Concern using Cas13

“…We used a computational design technique (see Methods; manuscript in prep.) to generate crRNAs that maximize predicted SNP discrimination at position 417, and used ADAPT 36 to help in designing RPA primers for these crRNAs. After assay optimization, we demonstrated that our 417 assays can differentiate between the three genotypes with high accuracy (Fig.…”

“…The CRISPR RNAs (crRNAs) for SNP discrimination were designed using a generative sequence design algorithm 55 . This approach uses ADAPT's predictive model to predict the activity of candidate crRNA sequences against on-target and off-target sequences 36 . These predictions of candidate crRNA activity guide the generative algorithm's optimization process, in which it seeks to design crRNA probes that have maximal predicted on-target activity and minimal predicted offtarget activity.…”