COVID-19 is detected using reverse transcription polymerase chain reaction (RT-PCR) of nasal swabs. A very sensitive and rapid detection technique using easily-collected fluids like saliva must be developed for safe and precise mass testing. Here, we introduce a metasurface platform for direct sensing of COVID-19 from unprocessed saliva. We computationally screen gold metasurfaces out of a pattern space of 2100 combinations for strongly-enhanced light-virus interaction with machine learning and use it to investigate the presence and concentration of the SARS-CoV-2. We use machine learning to identify the virus from Raman spectra with 95.2% sensitivity and specificity on 36 PCR positive and 33 negative clinical samples and to distinguish wild-type, alpha, and beta variants. Our results could pave the way for effective, safe and quantitative preventive screening and identification of variants.