Molecular diagnostic technologies have transformed the detection and surveillance of infectious pathogens. The performance of these assays depends critically on the diagnostic oligonucleotide sequences that they employ to bind to targeted nucleic acids. Current techniques for designing such diagnostic oligonucleotides, including PCR primers and clustered regularly interspaced short palindromic repeats (CRISPR) guide RNAs, select sequences directly from those in nature or from a simple function of natural sequences, such as a consensus, according to machine-learned discriminative models or thermodynamic heuristics^1–6. Thus, outputs of these techniques are identical (or nearly identical) to the input, natural sequences. However, natural diagnostic guide sequences may not perform the best in every application. Artificial diagnostic sequences that differ from any natural sequence by several mismatches may achieve superior diagnostic activity than their natural counterparts. Machine learning methods have been applied to generate new, artificial biological sequences such as proteins^7–12 and DNA regulatory elements^13–15, but, to the best of our knowledge, such techniques have not been applied to design diagnostic sequences.

We hypothesized that harnessing a generative design approach to construct artificial diagnostic oligonucleotide sequences may advance performance on key detection tasks. For instance, if our goal were to detect all forms of a genetically diverse pathogen (‘multitarget detection’) and there are several polymorphisms in an otherwise efficacious target region, an artificial oligonucleotide sequence could more sensitively detect all combinations of polymorphisms than any sequence that fully matches one particular combination (Fig. 1a). This would enable us to achieve improved diagnostic sensitivity. Alternatively, if our goal were to distinguish between two highly similar targets (‘variant identification’), even those that differ by a single-nucleotide polymorphism (SNP), an artificial sequence with optimally positioned mismatches could achieve greater specificity than a sequence identical to one target (Fig. 2a). This would enable us to more accurately identify mutation(s) in a pathogen or discriminate between pathogen lineages. Previous work^16,17 has described heuristic rules to introduce handcrafted specificity-enhancing mismatches into diagnostic oligonucleotide sequences, but these strategies are bespoke and are limited by the degree of variation they can distinguish.

In this study, we focus on designing guides for use with CRISPR–Cas13a-based detection assays; specifically, we design artificial 28 nt CRISPR guide RNA spacer sequences. Cas13a-based assays have increasingly received attention because of their minimal equipment requirements and ease of use in low-resource settings^18,19. We previously described Activity-informed Design with All-inclusive Patrolling of Targets (ADAPT)⁶, which designs diagnostic guides by using a convolutional neural network (CNN) as a discriminative model to select among natural sequences (or consensuses of sequence clusters). Since the approach selects among a predetermined list of sequences, ADAPT cannot design guides that are several mismatches or more from natural sequences, it cannot design guides for variant identification down to single-nucleotide resolution, and it is unable to uncover novel strategies that optimize guide design. By contrast, algorithms that create artificial guide sequences could achieve these goals while enabling superior diagnostic sensitivity and specificity, particularly for challenges in infectious diseases.

Designing artificial guide sequences necessitates exploring a vast high-dimensional space of RNA sequences. Each guide has a fitness according to the task at hand, such as multitarget detection (Fig. 1a) or variant identification (Fig. 2a). Highly fit artificial guide sequences should be nearby in sequence space to the complements of their targets, but this region, with ~10⁷–10⁹ sequences (Supplementary Note 1), is sufficiently large to motivate the development of algorithms that efficiently search the sequence space.

Here, we implement and evaluate a set of algorithms for designing artificial CRISPR–Cas13a guide RNAs. These algorithms combine a machine-learned model—trained to predict the enzymatic activity that results from a guide–target interaction—with search algorithms to explore a fitness landscape of guides. Our predictive model is the CNN we previously developed and implemented in ADAPT⁶. It predicts the enzymatic activity of Cas13a, a proxy of a given guide’s sensitivity for detecting a given target. We implement and evaluate two design algorithms: (1) a generative model that employs a conditional Wasserstein Generative Adversarial Network and activation maximization (WGAN-AM) to explore a continuous latent space of sequences^13,20,21 (Extended Data Fig. 1a and Supplementary Fig. 1) and (2) an evolutionary algorithm^22–24 that performs iterative rounds of random mutation, fitness evaluation and selection to explore the fitness landscape in discrete steps (Extended Data Fig. 1b and Supplementary Fig. 2). Our model-directed exploration process—including the predictive model, search algorithms and fitness functions—is conditioned on user-provided target sequences, making it broadly applicable to any input targets rather than being restricted to specific genomic sites. We built a software package to make these design approaches easily accessible for Cas13 guide RNA design. It is called Building Artificial Diagnostic Guides by Exploring Regions of Sequences (BADGERS) and is freely available at https://github.com/broadinstitute/badgers-cas13 (ref. 25).

First, we applied the algorithms in BADGERS to design guides, for use with Cas13a-based detection, that are sensitive across a pathogen’s genomic variation, that is, for multitarget detection. The fitness function that we maximize is the average predicted activity across sequenced genomes (Fig. 1a and Methods). We focus on five RNA viruses, chosen for their extensive diversity (enterovirus B, Lassa and dengue) or public health relevance (influenza A and severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2)). We employed the design algorithms implemented in BADGERS to design guides and benchmarked them against algorithms representing current, state-of-the-art design approaches (Methods). These baseline methods include a simple, commonly used algorithm that complements the consensus sequence at the target site (‘consensus’) and a more advanced algorithm that we term ‘model-based choice’ (MBC), which reflects the method implemented in ADAPT, which we previously developed⁶. MBC employs clustering to compute a candidate set of guides and selects the guide, from this predetermined list, with maximal fitness based on our model’s predictions. Although the predictive model of guide activity used by MBC and the algorithms in BADGERS is the same, MBC is conceptually and algorithmically distinct; MBC does not use the predictive model to explore a sequence space and it cannot generate artificial guide sequences with several mismatches from the targets (Methods).

We computationally evaluated the multitarget detection performance of the different design methods across these five pathogens using the same predictive model of guide activity (Methods). BADGERS-designed guides were predicted to detect more diversity than guides designed by baseline methods across nearly all genomic regions of all five pathogens. In several regions within dengue, Lassa and influenza A viruses, the BADGERS-designed guides were predicted to detect 10–20% more sequence variation (Fig. 1b and Supplementary Fig. 3). The sequence and nucleotide composition of a viral pathogen vary across different genomic regions and influence the activity of diagnostic guides; thus, BADGERS-designed guides may achieve a coverage improvement in some regions but not others. The BADGERS-designed guides also achieved a higher mean fitness genome wide (Fig. 1c and Supplementary Fig. 4a). At most sites, the guides generated by the algorithms in BADGERS are several mismatches away from any sequenced genomes, suggesting that the WGAN-AM and evolutionary methods’ ability to search fitness landscapes and generate artificial sequences contributes to their superior performance (Fig. 1d and Supplementary Fig. 5). For example, in Lassa virus segment S, the WGAN-AM and evolutionary algorithms designed guides that were distinct from any observed sequence at ~92% and ~99.9% of genomic sites, respectively.

Computational predictions are inherently limited so we set out to experimentally benchmark the multitarget detection performance of the BADGERS-designed guides. We used a random sampling strategy (Methods) to select three representative genomic sites from each of the five viruses we considered (15 sites in total) and employed both the algorithms in BADGERS and baseline methods to design guides for these sites. We tested these guides against multiple synthetic targets that represent the genomic diversity of each virus (Methods), measuring the fluorescence readout over time. Cas13a is activated through guide-mediated recognition of a target; once active, it engages in collateral cleavage of a fluorescent reporter to generate the fluorescent readout, with greater fluorescent signal indicating stronger guide–target affinity at a fixed target concentration. All guides were tested against the target concentrations 10¹⁰ copies per microlitre, 10⁹ copies per microlitre and 10⁸ copies per microlitre, which represent typical postamplification concentrations of viral pathogens^6,26,27.

For all but one of the 15 tested sites, the BADGERS-designed guides achieved higher activity across sequence diversity than the baseline guides (Fig. 1e–h, and Supplementary Figs. 6 and 7) and often enabled a lower limit of detection as well (Fig. 1i and Supplementary Fig. 6). The algorithms in BADGERS provide the greatest advantage in highly diverse viruses, for example, Lassa virus. This advantage is less pronounced at more conserved sites from less diverse viruses, for example, SARS-CoV-2 (Supplementary Figs. 3 and 8).

We next applied the algorithms in BADGERS to our second diagnostic objective: to design guides that optimally identify specific SNPs or distinguish between pathogen lineages, that is, variant identification (Fig. 2a). This objective can be conceptualized as designing a guide that achieves high activity on an on-target sequence and minimal activity on an off-target sequence; the fitness function is approximately the difference between these activities (Fig. 2a and Methods).

To computationally evaluate the performance of BADGERS’s generative design approach in designing guides that identify SNPs, we considered a set of 100 pairs of randomly generated on-target and off-target sequences, each differing by one nucleotide (Methods). We applied the algorithms in BADGERS to these target sets and also designed guides using the canonical strategy¹⁷ for this problem: this strategy places the SNP at position 26 of the guide and introduces a synthetic mismatch, at position 24 of the guide, to both the on- and off-target sequences (Methods). MBC’s design approach cannot create guides that can differentiate between lineages down to the single-nucleotide level, so it could not be applied to this task. In silico, the BADGERS-designed guides exhibited similar predicted on-target activity to those designed by this synthetic mismatch method, but achieved lower off-target activity (Fig. 2b–d).

We benchmarked the design algorithms’ variant identification performance experimentally, first by designing guides for six clinically relevant SNPs and comparing them with the baseline canonical synthetic mismatch approach. The six SNPs include four associated with antimicrobial resistance in Plasmodium falciparum (Pfcrt K76T, Pfdhps A437G, Pfk13 C580Y and Pfmdr1 Y184F), one associated with microcephaly in Zika virus (PrM S139N) and one associated with immune evasion in SARS-CoV-2’s spike protein (K417N/K417T)^28–34. Experimentally, across nearly all of these tasks, the BADGERS-designed guides exhibited lower off-target activities while maintaining similar on-target activities to baseline guides (Fig. 2e,f and Supplementary Fig. 9a–c). Several baseline guides (including those for the S139, A437G and K417T targets) had an off-target signal nearly identical to the on-target signal throughout the reaction, which complicates SNP identification. In contrast, the BADGERS-designed guides achieved an on-target signal two to three times above the off-target signal (Fig. 2e,f).

We further benchmarked the design algorithms’ variant discrimination performance experimentally for designing guides that differentiate viral lineages, focusing on dengue virus (DENV) serotypes 1–4 and seven key SARS-CoV-2 lineages (Methods). In the DENV task, one serotype is the on-target set (for example, DENV-1) and the other three serotypes comprise the off-target set (for example, DENV-2, 3 and 4). Both the BADGERS-designed guides and baseline guides differentiated the DENV subtypes with high specificity at a target concentration of 10⁸ copies per microlitre, probably because there are DENV genomic regions with sufficient dissimilarity across serotypes (Supplementary Fig. 9d). However, when tested at a lower target concentration of 10⁶ copies per microlitre, the baseline guides exhibited low on-target fluorescence for serotypes 2 and 4, while the BADGERS-designed guides achieved, as desired, a high on-target fluorescence for those serotypes (Fig. 2g). In the SARS-CoV-2 task, we designed guides that differentiate the ancestral Wuhan lineage (that is, Ref), Alpha, Beta, Delta, Gamma, Omicron/BA.1 and Omicron/BA.2 (Methods). The BADGERS-designed guides enabled robust lineage discrimination, while the baseline guides for Alpha, Gamma and Ref had off-target activities that were nearly as high as their on-target activities throughout the reaction (Supplementary Fig. 9e).

Overall, the WGAN-AM and evolutionary search algorithms performed similarly in our two tasks, despite employing distinct techniques and creating different guides. On average, the evolutionary algorithm designed guides that are more distinct from observed sequences than the WGAN-AM’s guides (Figs. 1d and 2c and Supplementary Fig. 5). Fundamentally, the WGAN-AM and evolutionary methods represent two different classes of search algorithms: the WGAN-AM samples sequences from a generative network while the evolutionary algorithm simulates natural selection. To explore the utility of alternative search algorithms, we evaluated the performance of another method in each class. Specifically, we performed in silico benchmarking of conditioning by adaptive sampling (CbAS)³⁵, which employs a generative network, and AdaLead³⁶, which employs an evolutionary-inspired approach (Methods). All four search algorithms provide a substantial improvement over the baseline approaches. Different search algorithms do exhibit subtle differences in performance depending on the specific task; for example, the WGAN-AM–designed guides achieved a lower mean fitness on the multitarget detection objective than guides designed by the other three search algorithms, although this was not the case on the variant identification objective (Supplementary Fig. 4).

Next, we sought to understand whether the algorithms in BADGERS leverage known rules governing Cas13a activity to enhance guide performance. Remarkably, we found that the generative design algorithms implicitly learn Cas13a’s protospacer and mismatch preferences without having them explicitly encoded and rationally apply this biological understanding among our two design objectives to generate optimally fit guides. For example, guide–target mismatches in close proximity to one another are known to decrease enzymatic activity^37–39. When applied to the first objective of multitarget detection, the BADGERS-designed artificial guide sequences reduced the number of adjacent guide–target mismatches at mismatch-sensitive positions relative to baseline guides, enabling superior sensitivity (Supplementary Fig. 10). In contrast, when applied to the variant identification objective, the algorithms in BADGERS introduced artificial mismatches within 4 nt of the SNP for the majority of target sets (WGAN-AM, 61% and evolutionary, 81%), reducing off-target activity (Fig. 2c). Additionally, for nearly 75% of simulated targets, the algorithms in BADGERS further optimized guide specificity by placing the SNP between positions 18 and 23, a region where mismatches are known⁶ to be highly deleterious to LwaCas13a activity (Fig. 2d).

Furthermore, we found that the algorithms in BADGERS take advantage of another biological property of Cas13a to design guides optimally suited for both objectives. Previous work has found that Cas13a guides exhibit reduced activity when targeting a genomic site with complementarity to Cas13a’s tag sequence, particularly targets with a G nucleotide at the position immediately 3′ to the protospacer (the protospacer-flanking site (PFS))^39,40. In the variant identification objective, when the targeted SNP is a mutation from a G to a non-G nucleotide (as it is in the experimentally tested S139N and C580Y targets), the algorithms in BADGERS positioned the G at the PFS to achieve minimal off-target activity (schematized in Extended Data Fig. 2 and experimental results in Fig. 2e and Supplementary Fig. 9a). However, in the multitarget detection objective, the algorithms in BADGERS exhibited an unexpected design decision. For genomic sites with a G at the PFS, the algorithms often introduced a mismatch directly adjacent to the tag sequence at position 28 in the protospacer (the same as position 1 in the spacer), which we term the tag-adjacent mismatch (TAM). Intriguingly, among the genomic sites with a G at the PFS, the TAM was present in 45.7% of guides designed by the WGAN-AM algorithm and 76.2% of guides designed by the evolutionary algorithm (Supplementary Fig. 11).

We hypothesized that the TAM could represent a different design strategy, discovered by the algorithms in BADGERS, that enables improved guide activity on targets with a G at the PFS, so we explored the TAM’s relevance further. Mechanistically, the TAM may disrupt base pairing between the G at the PFS of the target and the C in the CRISPR RNA (crRNA) tag sequence, promoting the separation of the Cas13–target complex and thereby increasing collateral cleavage activity and the resulting fluorescent signal. To test our hypothesis, we designed a library of guides against targets having unfavorable nucleotides (GN) and favorable nucleotides (AN, TN or CN) for the first two positions of the antitag region; the first nucleotide of the antitag region is the PFS (Methods and Fig. 2h). Guides with a TAM targeting a sequence with a G at the PFS exhibited substantially greater activity than guides without this TAM (Fig. 2h and Supplementary Fig. 12). Furthermore, the TAM conferred the greatest benefit when the targeted sequence was GT in the antitag region. This is mechanistically explainable: the first two nucleotides of Cas13a’s tag sequence complement GT, so the TAM’s disruption of this double base pairing may enhance activity more than for targets with a non-GT allele (Fig. 2h and Supplementary Fig. 12). As expected, introducing a TAM to guides targeting a sequence with a non-G at the PFS did not increase activity. To our knowledge, this design principle of introducing a mismatch in the spacer to rescue the activity of guides with extended tag:antitag complementarity has not been described before, suggesting that these algorithms are capable of elucidating new design principles.

In this study, we demonstrate that generative design algorithms can yield maximally fit guide sequences for pathogen detection and surveillance, and can illuminate guide design rules. These algorithms leverage a predictive model to explore a sequence landscape, only evaluating ~4,000 out of ~10¹⁶ potential guides throughout their search process, yet are capable of designing artificial guide sequences that outperform those derived directly from nature (Supplementary Note 1). Crucially, our approach is not trained on nor restricted to a specific design task, but rather can generate guides for any conditioned target set. The machine-learned models of guide activity that steer our optimization process enable our method to exploit features of Cas13a’s underlying biology (for example, antitag and mismatch preferences) to generate guides that optimize diagnostic performance. During the early 2022 surge in coronavirus disease 2019 (COVID-19) cases, we applied a prototype version of our method to design guides that distinguish SARS-CoV-2 variants, demonstrating that BADGERS-designed panels achieve high concordance with next-generation sequencing in a clinical context⁴¹. Furthermore, these methods can reveal interpretable guide design rules for CRISPR-based technologies. The TAM proposed by the generative design algorithms may enable Cas13a-based technologies to more efficiently target a broader range of sequences and improve our understanding of the enzyme’s nuclease activation mechanisms.

While false positive sequences, which score highly computationally yet are nonfunctional experimentally, are a challenge in generative sequence design, none of the BADGERS-designed guide sequences fell into this category. This may be partly due to the nature of our tasks as our focus on nucleotide base pairing entails substantially less structural complexity and a smaller sequence space relative to other tasks (for example, protein design). Nevertheless, the design algorithms that we evaluated still rely on predictive models of guide activity, which may return inaccurate predictions for guide–target pairs outside of the training domain. Certain design approaches—such as the WGAN-AM, which was trained to generate active guide–target pairs, or CbAS, which considers uncertainty in activity predictions—may be less susceptible to generating out-of-domain sequences than methods such as the evolutionary algorithm that are less constrained in their search. Additionally, these methods may not perform as well when applied to pathogens whose sequence composition is highly dissimilar from the viral sequences that comprised the predictive model’s training dataset. Also, because different orthologs of Cas13 are known to have varying sequence preferences, the design algorithms that we evaluated may not be reliable in designing guides for non-Cas13a enzymes. Furthermore, although these algorithms do enhance diagnostic sensitivity, many regions of viral genomes are hypervariable and, in some cases, even the guides designed by our methods do not detect the full extent of sequence diversity (for example, DENV and enterovirus B; Fig. 1f,g).

In the realm of pathogen diagnostics, to our knowledge, all current design methods for nucleic acid technologies derive oligonucleotide sequences directly from pathogen genomes^1–6. Our results suggest that generative sequence design algorithms offer substantial benefits relative to approaches (like the one in ADAPT) that are not geared toward designing artificial guide sequences. Namely, the algorithms in BADGERS created guides that achieved superior performance to MBC (the technique employed by ADAPT) on the multitarget detection objective and enabled the discrimination of lineages down to single-nucleotide resolution, a task ADAPT cannot perform. To achieve this, BADGERS’s methods engineered guide sequences by leveraging rational design principles (for example, G–U base pairing and nucleotide and position-specific effects of mismatches) and discovered strategies for guide design (including the terminal adjacent mismatch).

Though we focused on Cas13 guide RNAs in this study, artificial sequences stand to optimize the design of diagnostic sequences beyond CRISPR guide RNAs, including the primers and fluorescent probes employed in PCR, loop-mediated isothermal amplification and recombinase polymerase amplification. While the specific search algorithms employed in this work to generate artificial sequences achieved strong performance, the field of biological sequence design is rapidly advancing with the recent application of other classes of design methods (for example, autoregressive and diffusion-based generative methods) and future work should investigate how these approaches may further benefit diagnostic design. Beyond diagnostics, generative design methods could design rationally mismatched guide RNAs for base editors to minimize bystander editing rates¹⁶. For certain CRISPR screens, where the objective is to introduce maximal sequence diversity into protein-coding genes, these methods could design guides that maximize the number of distinct edits made at the targeted sites⁴². Furthermore, while high-fidelity variants of CRISPR enzymes have lowered off-target editing rates⁴³, these class of algorithms could optimize guide RNA sequences to minimize these off-target effects further. Recent work has also highlighted the importance of accounting for human genetic diversity when designing CRISPR therapeutics^43,44, and generative design approaches are poised to generate guide sequences that optimally account for this diversity.

Generative methods are transforming the field of biological sequence design, with recent work focusing on generating long protein and DNA sequences^10,12,45. Here, we show that algorithms that explore sequence landscapes to design artificial sequences can not only optimize shorter RNA sequences, such as crRNAs, but can simultaneously elucidate mechanistic design rules and serve needs in pathogen surveillance.