Designing sensitive viral diagnostics with machine learning

Abstract

Design of nucleic acid-based viral diagnostics typically follows heuristic rules and, to contend with viral variation, focuses on a genome’s conserved regions. A design process could, instead, directly optimize diagnostic effectiveness using a learned model of sensitivity for targets and their variants. Toward that goal, we screen 19,209 diagnostic–target pairs, concentrated on CRISPR-based diagnostics, and train a deep neural network to accurately predict diagnostic readout. We join this model with combinatorial optimization to maximize sensitivity over the full spectrum of a virus’s genomic variation. We introduce Activity-informed Design with All-inclusive Patrolling of Targets (ADAPT), a system for automated design, and use it to design diagnostics for 1,933 vertebrate-infecting viral species within 2 hours for most species and within 24 hours for all but three. We experimentally show that ADAPT’s designs are sensitive and specific to the lineage level and permit lower limits of detection, across a virus’s variation, than the outputs of standard design techniques. Our strategy could facilitate a proactive resource of assays for detecting pathogens.

Multiplexed detection of bacterial nucleic acids using Cas13 in droplet microarrays

“…Using the resulting amplicon sequences, we then applied ADAPT ( 30 ) to design crRNA guides that optimally distinguish each species from all others. These guides were predicted not to cross-react (> 3 base pair mismatches) with amplicons corresponding to all included strains from the 51 nontarget pathogens in the panel.…”

Multiplexed CRISPR-based microfluidic platform for clinical testing of respiratory viruses and identification of SARS-CoV-2 variants

“…This included all viruses covered by BioFire RP2.1-SARS-CoV-2, four other human-associated coronaviruses and both influenza strains- as well as a few additional illness-inducing viruses 41 . To generate maximally active virus-specific crRNAs and PCR primers to detect the 21 viruses, we applied the assay design method ADAPT (Activity-informed Design with All-inclusive Patrolling of Targets; described in the Methods ) 35 . We were able to encompass the full genomic diversity of these viral families by including multiple primers, if needed.…”

“…The crRNAs for SNP discrimination were designed using a generative sequence design algorithm (Mantena, S. et al, manuscript in preparation). This approach uses ADAPT’s predictive model to predict the activity of candidate crRNA sequences against on-target and off-target sequences 35 . These predictions of candidate crRNA activity steer the generative algorithm’s optimization process, where it seeks to design crRNA probes that have maximal predicted on-target activity and minimal predicted off-target activity.…”

“…The crRNA target binding events are highly specific and altered by the presence of sequence variation. Maximally active crRNA design has been accelerated by machine learning and other computational methods 35 . Nonetheless, most CRISPR diagnostics detect one to three targets per sample 30 , 36 – 39 , which is not sufficient for differential diagnosis via comprehensive microbe or variant identification.…”

Prognostic Prediction Models for Postoperative Patients with Stage I to III Colorectal Cancer: A Retrospective Study Based on Machine Learning Methods

“…Thus, we explored the possibility of using machine learning (ML) to build a prognostic model for CRC. ML is a branch of arti cial intelligence (AI) in which a computer generates rules underlying or based on raw data (Mitchell, 2003); the use of ML in medicine has gradually become common (Grimm et al 2022;Kim et al 2022;Metsky et al 2022; Xie and Zhuang and Niu and Ai et al 2022). ML can be used to directly compare the accuracy of two or more quantitative tests for the same disease/condition (Tripepi et al 2009).…”