## Main

A septuagenarian male was admitted to a tertiary hospital in the summer of 2020 and had tested positive for SARS-CoV-2 using reverse-transcription quantitative PCR (RT–qPCR) 35 days previously in a nasopharyngeal swab (day 1) at a local hospital (Extended Data Figs. 1, 2). His past medical history included marginal B cell lymphoma diagnosed in 2012, with previous chemotherapy including vincristine, prednisolone, cyclophosphamide and anti-CD20 B cell depletion with rituximab. It is likely that both chemotherapy and underlying lymphoma contributed to combined immunodeficiency of B and T cells (Extended Data Figs. 2, 3 and Supplementary Table 1). Computed tomography of the chest showed widespread abnormalities consistent with pneumonia associated with coronavirus disease 2019 (COVID-19) (Supplementary Fig. 1). Treatment included two 10-day courses of remdesivir with a 5-day gap in between (Extended Data Fig. 1). Two units of convalescent plasma were administered on days 63 and 65 (Extended Data Fig. 3). After clinical deterioration, remdesivir and a unit of convalescent plasma were administered on day 95, but the individual died on day 102 (Supplementary Note).

## Comparative analysis of sequential virus samples

Most samples were respiratory samples from the nose and throat or endotracheal aspirates during the period of intubation (Supplementary Table 3). Cycle threshold (Ct) values ranged from 16 to 34 and all 23 respiratory samples were successfully sequenced by a standard single-molecule sequencing approach as per the ARTIC protocol implemented by The COVID-19 Genomics UK (COG-UK; https://www.cogconsortium.uk/) Consortium; of these samples, 20 additionally underwent short-read deep sequencing using the Illumina platform (Supplementary Table 4). There was general agreement between the two methods (Extended Data Fig. 4). However, owing to the higher reliability of Illumina for low frequency variants, this was used for formal analysis2,3. Additionally, single-genome amplification and sequencing of spike using extracted RNA from respiratory samples was used as an independent method to detect the observed mutations (Extended Data Fig. 4). Finally, we detected no evidence of recombination, based on two independent methods (data not shown).

Maximum-likelihood analysis of patient-derived whole-genome consensus sequences demonstrated clustering with other local sequences from the same region (Fig. 1). The infecting strain was assigned to lineage 20B bearing the spike(D614G) variant. Environmental sampling showed evidence of virus on surfaces such as the telephone and call bell. Sequencing of these surface viruses showed clustering with those derived from the respiratory tract (Extended Data Fig. 2). All samples were consistent with having arisen from a single underlying viral population. In our phylogenetic analysis, we included sequential sequences from three other local patients identified with persistent viral RNA shedding over a period of 4 weeks or more as well as two recently reported long-term immunosuppressed individuals with SARS-CoV-24,5 (Extended Data Fig. 2 and Supplementary Table 2). Whereas the sequences from the three local patients as well as one of the previous studies5 showed little divergence with no amino acid changes in the spike protein over time, the case patient (hereafter, patient X1) showed considerable diversification. The other, previously reported patient4 showed a similar degree of diversification as patient X1. Further investigation of the sequence data suggested the existence of an underlying structure to the viral population in patient X1, with samples collected at days 93 and 95 being rooted within, but divergent from, the original population (Extended Data Figs. 5, 6). The relationship of the divergent samples to those at earlier time points argues against superinfection.

## SARS-CoV-2 viral diversity

All samples tested positive for SARS-CoV-2 by RT–qPCR and there was no sustained change in Ct values throughout the 101 days after the first two courses of remdesivir (days 41 and 54) or the first two units of convalescent plasma with polyclonal antibodies (days 63 and 65) (Extended Data Fig. 3). Notably, we were not able to culture virus from stored swab samples. Consensus sequences from short-read deep-sequencing Illumina data revealed dynamic population changes after day 65, as shown in a highlighter plot (Extended Data Fig. 6). In addition, we were also able to follow the dynamics of virus populations down to low frequencies during the entire period (Fig. 2 and Supplementary Table 4). After treatment with remdesivir on day 41, the low frequency variant analysis showed transient amino acid changes in populations at below 50% abundance in open reading frame (ORF) 1b, 3a and spike, with a C27509T (causing a T39I substitution) mutation in orf7a reaching 79% on day 45 (Fig. 2, and Supplementary Information). We found that an I513T substitution (encoded by T2343C) in NSP2 and a V157L substitution (encoded by G13936T) in RdRp had emerged from undetectable at day 54 to almost 100% frequency on day 66 (Fig. 2), with the mutation in the polymerase being the more plausible candidate for driving this sweep. Notably, spike(N501Y)—which can increase the affinity for the ACE2 receptor6 and which is present in the UK B1.1.7 lineage7—was observed on day 55 at a frequency of 33%, but was eliminated by the sweep of the NSP2 and RdRp variant.

In contrast to the early period of infection, between days 66 and 82, after the first two administrations of convalescent sera, a shift in the virus population was observed, with a variant bearing a D796H substitution in S2 and a ΔH69/ΔV70 deletion in the S1 N-terminal domain (NTD) of spike becoming the dominant population at day 82. This was identified in a nose and throat swab sample with high viral load as indicated by a Ct value of 23 (Fig. 3a). The deletion was detected transiently at baseline on the basis of to short-read deep sequencing. The ΔH69/ΔV70 deletion was due to an out-of-frame six-nucleotide deletion, which results in the sequence of codon 68 changing from ATA to ATC.

On days 86 and 89, viruses obtained from upper respiratory tract samples were characterized by the spike(Y200H, T240I) double mutant, with the mutation–deletion pair (spike(D796H, ΔH69/ΔV70)) observed on day 82 having decreased to frequencies of 10% or less (Figs. 2, 3). Spike(Y200H, T240I) were accompanied at high frequency by three other nonsynonymous variants with similar allele frequencies, encoding I513T in NSP2, V157L in RdRp and N177S in NSP15 (Fig. 2a). Both of these were also previously observed at a frequency of more than 98% in the sample on day 66 (Fig. 2a), suggesting that this new lineage emerged out of a population that existed previously.

Sequencing of a nose and throat swab sample on day 93 identified viruses characterized by spike(P330S) at the edge of the receptor-binding domain (RBD) and spike(W64G) in the S1 NTD at close to 100% abundance, whereas spike(D796H, ΔH69/ΔV70) showed an abundance of less than 1% and the variants spike(Y200H) and spike(T240I) had frequencies of less than 2%. Viruses with the spike(P330S) variant were detected in two independent samples from different sampling sites, arguing against the possibility of contamination. The divergence of these samples from the remainder of the population (Figs. 2, 3b and Extended Data Figs. 5, 6) suggests the possibility that they represent a compartmentalized subpopulation.

Patterns in the variant frequencies suggest competition between virus populations carrying different mutations; viruses with the mutation–deletion pair spike(D796H, ΔH69/ΔV70) rose to high frequency during convalescent plasma therapy, but were then outcompeted by another population in the absence of therapy. Specifically, these data are consistent with a lineage of viruses with the NSP2(I513T) and RdRp(V157L) variant, which was dominant on day 66 but was outcompeted during therapy by the mutation–deletion variant. With the lapse in therapy, the original strain—which had acquired NSP15(N1773S) and the spike(Y200H, T240I)—regained dominance, followed by the emergence of a separate population with the spike(W64G, P330S) variant.

In a final attempt to reduce the viral load, a third course of remdesivir (day 93) and a third dose of convalescent plasma (day 95) were administered. We observed the re-emergence of the spike(D796H, ΔH69/ΔV70) viral population (Figs. 2, 3). The inferred linkage of spike(D796H) and spike(ΔH69/ΔV70) was maintained as evidenced by the highly similar frequencies of the two variants, suggesting that the third unit of convalescent plasma led to the re-emergence of this population under renewed positive selection. In further support of our proposed idea of competition, frequencies of these two variants appeared to mirror changes in the NSP2(I513T) variant (Fig. 2), suggesting that these variants are markers of opposing clades in the viral population. Ct values remained low throughout this period with hyperinflammation, which eventually led to multi-organ failure and death on day 102. The repeated increase in the frequency of the viral population with convalescent plasma therapy strongly supports the hypothesis that the combination of the deletion and mutation in the spike protein conferred a selective advantage.

## Spike mutations impair neutralizing antibody potency

Using lentiviral pseudotyping we generated wild-type spike, spike(D796H, ΔH69/ΔV70) and single mutant spike proteins in enveloped virions to measure the neutralization activity of convalescent plasma against these viruses (Fig. 4). This system has been shown to give generally similar results to replication-competent viruses8,9. Spike protein from each mutant was detected in pelleted virions (Fig. 4a). We also used an HIV-1(p24) antibody to monitor the levels of lentiviral particle production (Fig. 4a and Supplementary Fig. 2). We then measured the infectivity of the pseudoviruses, correcting for virus input using measurements of the reverse transcriptase activity, and found that spike(ΔH69/ΔV70) appeared to have twofold higher infectivity over a single round of infection compared to wild-type spike (Fig. 4b and Extended Data Fig. 7). By contrast, the spike(D796H) single mutant had significantly lower infectivity compared to the wild-type spike protein and the double mutant (spike(D796H, ΔH69/ΔV70)) had a similar infectivity to wild-type spike (Fig. 4b and Extended Data Fig. 7).

We found that spike(D796H) alone and the spike(D796H, ΔH69/ΔV70) double mutant were less sensitive to neutralization by convalescent plasma samples (Fig. 4c–e and Extended Data Fig. 7). By contrast, the spike(ΔH69/ΔV70) single mutant did not reduce neutralization sensitivity. In addition, patient-derived serum from days 64 and 66 (one day before and after the infusion with the second convalescent plasma therapy) similarly showed lower potency against the spike(D796H, ΔH69/ΔV70) double mutant (Fig. 4f, g).

A panel of nineteen monoclonal antibodies isolated from three donors was previously identified to neutralize SARS-CoV-2. To establish whether the mutations that occurred in vivo (spike(D796H) and spike(ΔH69/ΔV70)) resulted in a global change in neutralization sensitivity, we tested neutralizing monoclonal antibodies targeting the seven major epitope clusters that have previously been described (excluding the non-neutralizing clusters II and V and the small (n ≤2) neutralizing clusters IV and X). The eight RBD-specific monoclonal antibodies (Extended Data Fig. 8) exhibited no major change in neutralization potency and the non-RBD-specific antibody COVA1-21 showed a 3–5-fold reduction in potency against spike(D796H, ΔH69/ΔV70) and spike(ΔH69/ΔV70), but not against spike(D796H) alone9 (Extended Data Fig. 8). We observed no differences in neutralization between single or double mutants and the wild-type spike protein, suggesting that the mechanism of escape was probably outside of these epitopes in the RBD. These data confirm the specificity of the findings from convalescent plasma and suggest that the observed mutations are related to antibodies that target regions outside of the RBD. Notably, spike(ΔH69/ΔV70)-containing viruses showed reduced neutralization sensitivity to the monoclonal antibody COVA1-21, targeting an as-yet undefined epitope outside of the RBD10.

To understand how spike(ΔH69/ΔV70) and spike(D796H) might confer antibody resistance, we assessed how these mutations might affect the spike structure (Extended Data Fig. 9). We based this analysis primarily on a structure that lacked stabilizing modifications (PDB 6XR8)11, but also referred to stabilized structures determined at different pH values12. ΔH69/ΔV70 is located in a disordered, glycosylated loop at the distal surface of the NTD, near the binding site of polyclonal antibodies derived from COV57 plasma13,14 (Extended Data Fig. 9). As this loop is flexible and highly accessible, ΔH69/ΔV70 could in principle affect antibody binding in this region. D796 is located near the base of the spike protein, in a surface loop that is structurally disordered in the prefusion conformation and becomes part of a large disordered region in the post-fusion S2 trimer11 (Extended Data Fig. 9). The loop containing residue 796 is proposed to be targeted by antibodies15, despite mutations at position 796 being relatively uncommon (Extended Data Fig. 9). In the RBD-down spike structures11,12, D796 forms contacts with residues in the neighbouring protomer, including the glycosylated residue N709 (Extended Data Fig. 9).

## Discussion

Here we documented a repeated evolutionary response by SARS-CoV-2 in the presence of antibody therapy during the course of a persistent infection in an immunocompromised host. The observation of potential selection for specific variants coinciding with the presence of antibodies from convalescent plasma is supported by the experimental finding of the twofold reduced susceptibility of these viruses to convalescent plasma containing polyclonal antibodies. In this case, the emergence of the variant was not the primary reason for treatment failure.

We noted in our analysis signs of compartmentalized viral replication based on the sequences recovered in upper respiratory tract samples. Both population-genetic and small-animal studies have shown a lack of reassortment between influenza viruses within a single host during an infection, suggesting that acute respiratory viral infection may be characterized by spatially distinct viral populations16,17. In the analysis of data, it is important to distinguish genetic changes that occur in the primary viral population from apparent changes that arise from the stochastic observation of spatially distinct subpopulations in the host. Although the samples that we obtained on days 93 and 95 of infection are genetically distinct from the other samples, the remaining samples are consistent with arising from a consistent viral population. We note that detection of viral RNA in post-mortem tissue was observed previously not only in lung tissue, but also in the spleen, liver, and heart4. Mixing of virus from different compartments—for example, via blood or the movement of secretions from lower to upper respiratory tract—could lead to fluctuations in viral populations at particular sampling sites.

This is a report of a single case and therefore limited conclusions can be drawn about generalizability.

An important limitation is that the data were derived from sampling from the upper respiratory tract and not the lower tract, thus limiting the inferences that can be drawn regarding viral populations in this single patient.

In addition to documenting the emergence of SARS-CoV-2 spike(ΔH69/ΔV70) in vivo, we show that this variant modestly increases infectivity of the spike protein in a pseudotyping assay. The deletion was observed contemporaneously with the rare S2 variant spike(D796H) after two separate courses of convalescent plasma, with other viral populations emerging. D796H, but not ΔH69/ΔV70, conferred a reduction in susceptibility to polyclonal antibodies in the units of convalescent plasma administered, although we cannot speculate as to their individual effect on sera from other individuals. It is notable that the spike(D796H, ΔH69/ΔV70) double mutant decreased in prevalence in between convalescent plasma courses, suggesting that there were other selective forces at play in the intervening period, which are possibly driven by the inflammation observed in the individual. This includes the possibility that the haplotype comprising spike (D796H, ΔH69/ΔV70) may have carried mutations in other regions that were deleterious during that intervening period. Although ΔH69/ΔV70 is expanding at a high rate18, D796 mutations are also increasing. D796H has been documented in 0.02% of global sequences and D796Y appears in 0.05% of global sequences (Extended Data Fig. 9).

The effects of convalescent plasma on virus evolution found here are unlikely to apply in immunocompetent hosts in whom viral diversity is likely to be lower owing to better immune control. Our data highlight that infection control measures may need to be tailored to the needs of immunocompromised patients and also caution in the interpretation of guidelines from the US Centres of Disease Control that recommend 20 days as the upper limit of infection-prevention precautions in immunocompromised patients who are afebrile19. Owing to the difficulty of culturing clinical isolates, use of surrogate experiments are warranted20. However, in cases in which the detection of ongoing viral evolution is possible, this serves as a clear proxy for the existence of infectious virus. In our case, we detected environmental contamination while in a single-occupancy room and the patient was moved to a negative-pressure high air-change infectious-disease isolation room.

The clinical efficacy of convalescent plasma in patients with severe COVID-19 has not been demonstrated21, and its use in different stages of infection and disease remains experimental; as such, we suggest that it should be reserved for use within clinical trials, and with rigorous monitoring of clinical and virological parameters. The data from this single patient indicate caution should be used for convalescent plasma therapy in patients with immunosuppression of both T cell and B cell arms; in these patients, the administered antibodies have little support from cytotoxic T cells, thereby reducing the chances of clearance and theoretically raising the potential for the evolution of escape mutations in SARS-CoV-2. Although we await further data, in cases in which clinical trial enrolment is not possible, convalescent plasma administered for clinical need in immunosuppressed individuals should ideally only be considered as part of observational studies, undertaken preferably in single-occupancy rooms with enhanced infection-control precautions, including environmental sampling of SARS-CoV-2 and real-time sequencing. Understanding the viral dynamics and characterizing the viral evolution in response to different selection pressures in an immunocompromised individual are necessary not only for improved patient management but also for public health benefit.

## Methods

### Clinical-sample collection and next-generation sequencing

Serial samples were collected from the patient periodically from the lower respiratory tract (sputum or endotracheal aspirate), upper respiratory tract (throat and nasal swab) and from stool. Nucleic acid extraction was done from 500 μl of sample with a dilution of MS2 bacteriophage to act as an internal control, using the easyMAG platform (Biomerieux) according to the manufacturers’ instructions. All samples were tested for presence of SARS-CoV-2 with a validated one-step RT–qPCR assay developed in conjunction with the Public Health England Clinical Microbiology22. Amplification reactions were all performed on a Rotorgene PCR instrument. Samples with a Ct value of ≤36 were considered to be positive.

Sera from recovered patients in the COVIDx study23 were used for testing of neutralization activity by SARS-CoV-2 mutants.

### SARS-CoV-2 serology by multiplex particle-based flow cytometry

Recombinant SARS-CoV-2 nucleocapsid, spike and RBD proteins were covalently coupled to distinct carboxylated bead sets (Luminex) to form a triplex and analysed as previously described24. Specific binding was reported as mean fluorescence intensities.

### Whole-blood T cell and innate stimulation assay

Whole blood was diluted 1:5 in RPMI into 96-well F plates (Corning) and activated by single stimulation with phytohaemagglutinin (10 μg ml−1; Sigma-Aldrich), lipopolysaccharide (1 μg ml−1, List Biochemicals) or by costimulation with anti-CD3 (MEM57, Abcam, 200 ng ml−1, 1:1,000) and IL-2 (Immunotools, 1,430 U ml−1, 1:1,000) antibodies. Supernatants were taken after 24 h. Levels (pg ml−1) are shown for IFNγ, IL-17, IL-2, TNF, IL-6, IL-1b and IL-10. Cytokines were measured by multiplexed particle-based flow cytometry on a Luminex analyser (Bio-Plex, Bio-Rad) using an R&D Systems custom kit (R&D Systems).

For viral genomic sequencing, total RNA was extracted from samples as described previously25. Samples were sequenced using MinION flow cells v.9.4.1 (Oxford Nanopore Technologies) following the ARTICnetwork v.3 protocol26 and BAM files assembled using the ARTICnetwork assembly pipeline27. A representative set of 10 sequences was selected and also sequenced using the Illumina MiSeq platform. Amplicons were diluted to 2 ng μl−1 and 25 μl (50 ng) were used as input for each library preparation reaction. The library preparation used KAPA Hyper Prep kit (Roche) according to the manufacturer’s instructions. In brief, amplicons were end-repaired and had A-overhang added; these were then ligated with 15 mM of NEXTflex DNA Barcodes (Bio Scientific). Post-ligation products were cleaned using AMPure beads and eluted in 25 μl. Then, 20 μl was used for library amplification by 5 cycles of PCR. For the negative controls, 1 ng was used for ligation-based library preparation. All libraries were assayed using TapeStation (Agilent Technologies) to assess fragment size and quantified by qPCR. All libraries were then pooled in equimolar ratios accordingly. Libraries were loaded at 15 nM and spiked in 5% PhiX (Illumina) and sequenced on for MiSeq 500 cycle using a Miseq Nano v.2 with 2 × 250 bp paired-end sequencing. A minimum of ten reads was required for a variant call.

### Bioinformatics processes

For long-read sequencing, genomes were assembled with reference-based assembly and a curated bioinformatics pipeline with 20× minimum coverage across the whole genome27. For short-read sequencing, FASTQ files were downloaded, poor-quality reads were identified and removed, and both Illumina and PHiX adapters were removed using TrimGalore v.0.6.628. Trimmed paired-end reads were mapped to the National Center for Biotechnology Information SARS-CoV-2 reference sequence MN908947.3 using MiniMap2-2.17 with arguments -ax and sr29. BAM files were then sorted and indexed with Samtools v.1.11 and PCR optical duplicates were removed using Picard (http://broadinstitute.github.io/picard). Consensus sequences of nucleic acids with a minimum whole-genome coverage of at least 20× were generated with BCFtools using a 0% majority threshold.

### Variant calling

Variant frequencies were validated using custom code as part of the AnCovMulti package (https://github.com/PollockLaboratory/AnCovMulti). The main idea behind this validation was to identify and remove consistent potential amplification errors and mutability near the end of Illumina reads. Furthermore, stringent filtering was applied to remove biased amplification of early laboratory-induced mutations or very low copy variations.

Filtering consisted of requiring exact initiation at a primer within 2 bp of the start of a read, a minimum of 247 bp length read, fewer than four well-separated sites divergent from the reference sequence, a maximum insertion size of three nucleotides, a maximum deletion size of 11 bp, and resolution of conflicting signal from different primers.

### Single-genome amplification and sequencing

Viral RNA extracts were reverse-transcribed from each sample to sufficiently capture the diversity of the viral population without introducing resampling bias. SuperScript IV (Thermo Fisher Scientific) and gene-specific primers were used for reverse transcription. Template RNA was degraded with RNase H (Thermo Fisher Scientific). All primers used were ‘in-house’ primers designed using the multiple sequence alignment of the patient’s consensus sequences obtained by next-generation sequencing. A partial gene sequence of spike (encoding amino acids 21–800) was amplified as one continuous length of DNA (the spike gene is approximately 1.8 kb) by nested PCR. Terminally diluted cDNA was amplified by PCR using Platinum Taq DNA Polymerase High Fidelity (Invitrogen) so that 30% of reactions were positive30. By Poisson statistics, sequences were deemed ≥80% likely to be derived from HIV-1 single genomes. We obtained between 20 and 60 single genomes at each sample time point to achieve 90% confidence of detecting variants present at ≥8% of the viral population in vivo31,32. Partial spike amplicons obtained from terminal-dilution PCR amplification were Sanger-sequenced to form a contiguous sequence using another set of eight in-house primers. Sanger sequencing was provided by Genewiz and manual sequence editing was performed using DNA Dynamo software (Blue Tractor Software).

### Phylogenetic analysis

All available full-genome SARS-CoV-2 sequences were downloaded from the GISAID database (http://gisaid.org/)33 on 16 December 2020. Duplicate and low-quality sequences (>5% nucleocapsid regions) were removed, leaving a dataset of 212,297 sequences with a length of more than 29,000 bp. All sequences were sorted by name and only sequences sequenced with United Kingdom/England identifiers were retained. From this dataset, sequences were deduplicated and, in figures in which background sequences were required, randomly subsampled using seqtk (https://github.com/lh3/seqtk). All sequences were aligned to the SARS-CoV-2 reference strain MN908947.3, using MAFFT v.7.475 with automatic flavour selection34. Major SARS-CoV-2 clade memberships were assigned to all sequences using both the Nextclade server v.0.9 (https://clades.nextstrain.org/) and Phylogenetic Assignment Of Named Global Outbreak Lineages (PANGOLIN)35.

Maximum-likelihood phylogenetic trees were produced using the above curated dataset using IQ-TREE v.2.1.236. Evolutionary model selection for trees were inferred using ModelFinder37 and trees were estimated using the GTR + F + I model with 1,000 ultrafast bootstrap replicates38. All trees were visualized with Figtree v.1.4.4 (http://tree.bio.ed.ac.uk/software/figtree/), rooted on the SARS-CoV-2 reference sequence and nodes arranged in descending order. Nodes with bootstraps values of less than 50 were collapsed using an in-house script.

### In-depth allele frequency variant calling

The SAMFIRE package v.1.0639 was used to call allele frequency trajectories from BAM file data. Reads were included in this analysis if they had a median PHRED score of at least 30, trimming the ends of reads to achieve this if necessary. Nucleotides were then filtered to have a PHRED score of at least 30; reads with fewer than 30 such reads were discarded. Distances between sequences, accounting for low-frequency variant information, were also obtained using SAMFIRE. The sequence distance metric, described in a previous paper40, combines allele frequencies across the whole genome. Where L is the length of the genome, we define q(t) as a 4 × L element vector describing the frequencies of each of the nucleotides A, C, G, and T at each locus in the viral genome sampled at time t. For any given locus i in the genome, we calculate the change in allele frequencies between the times t1 and t2 via a generalization of the Hamming distance

$$d({q}_{i}({t}_{1}),{q}_{i}({t}_{2}))=\frac{1}{2}\sum _{a\in \{{\rm{A}},{\rm{C}},{\rm{G}},{\rm{T}}\}}|{q}_{i}^{a}({t}_{1})-{q}_{i}^{a}({t}_{2})|$$

where the vertical lines indicate the absolute value of the difference. These statistics were then combined across the genome to generate the pairwise sequence distance metric

$$D({\bf{q}}({t}_{1}),{\bf{q}}({t}_{2}))=\sum _{i}d({q}_{i}({t}_{1}),{q}_{i}({t}_{2}))$$

The Mathematica software package was used to conduct a regression analysis of pairwise sequence distances against time, leading to an estimate of a mean rate of within-host sequence evolution. In contrast to the phylogenetic analysis, this approach assumed that the samples collected on days 93 and 95 arose via stochastic emission from a spatially separated subpopulation within the host, leading to a lower inferred rate of viral evolution for the bulk of the viral population.

All variants were indecently validated using custom code as part of the AnCovMulti package (https://github.com/PollockLaboratory/AnCovMulti).

### Western blot analysis

Cells were transfected with the indicated plasmid preparations and 48 h after transfection the culture supernatant was collected and passed through a 0.45-μm-pore-size filter to remove cellular debris. The filtrate was centrifuged at 15,000 rpm for 120 min to pellet virions. The pelleted virions were lysed in Laemmli reducing buffer (1 M Tris-HCl (pH 6.8), SDS, 100% glycerol, β-mercaptoethanol and bromophenol blue). Pelleted virions were subjected to electrophoresis on SDS 4–12% Bis-Tris protein gels (Thermo Fisher Scientific) under reducing conditions. This was followed by electroblotting onto polyvinylidene difluoride (PVDF) membranes. The SARS-CoV-2 spike proteins were visualized using a ChemiDoc MP imaging system (Biorad) using anti-spike S2 (Invitrogen; 1:1,000 dilution) and anti-p24 Gag (NIH AIDS Reagents; 1:1,000 dilution) antibodies.

### Recombination detection

All sequences were tested for potential recombination, as this would affect evolutionary estimates. Potential recombination events were explored with nine algorithms (RDP, MaxChi, SisScan, GeneConv, Bootscan, PhylPro, Chimera, LARD and 3SEQ), implemented in RDP5 with default settings41. To corroborate any findings, ClonalFrameML v.1.1242 was also used to infer recombination breakpoints. Neither of the programs found evidence of recombination in our data.

### Structural viewing

The PyMOL Molecular Graphics System v.2.4.0 (https://github.com/schrodinger/pymol-open-source/releases) was used to map the location of the four spike mutations of interested onto a previously published SARS-CoV-2 spike structure (PDB: 6ZGE)43.

### Testing of convalescent plasma for antibody titres

The anti-SARS-CoV-2 ELISA (IgG) assay used to test convalescent plasma for antibody titres was Euroimmun Medizinische Labordiagnostika. This indirect ELISA-based assay uses a recombinant structural spike 1 (S1) protein of SARS-CoV-2 expressed in the human cell line HEK293 for the detection of SARS-CoV-2 IgG.

### Generation of spike mutants

Amino acid substitutions were introduced into the D614G pCDNA_SARS-CoV-2_Spike plasmid as previously described44 using the QuikChange Lightening Site-Directed Mutagenesis kit, following the manufacturer’s instructions (Agilent Technologies).

### Pseudotype virus preparation

Viral vectors were prepared by transfection of HEK293T cells using the Fugene HD transfection reagent (Promega). HEK293T cells were transfected with a mixture of 11 μl of Fugene HD, 1 μg of pCDNAΔ19Spike–HA, 1 μg of p8.91 HIV-1 gag–pol expression vector45,46 and 1.5 μg of pCSFLW (expressing the firefly luciferase reporter gene with the HIV-1 packaging signal). Viral supernatant was collected at 48 and 72 h after transfection, filtered through 0.45-μm filter and stored at −80 °C. The 50% tissue-culture infectious dose (TCID50) of SARS-CoV-2 pseudovirus was determined using the Steady-Glo Luciferase assay system (Promega).

### Standardization of virus input by SYBR-Green-based product-enhanced PCR assay

The reverse transcriptase activity of virus preparations was determined by qPCR using a SYBR-Green-based product-enhanced PCR assay as previously described47. In brief, tenfold dilutions of virus supernatant were lysed in a 1:1 ratio in a 2× lysis solution (made up of 40% glycerol (v/v), 0.25% Triton X-100 (v/v), 100 mM KCl, RNase inhibitor 0.8 U ml−1, Tris HCL 100 mM, buffered to pH 7.4) for 10 min at room temperature.

Then, 12 μl of each sample lysate was added to 13 μl of a SYBR Green master mix (containing 0.5 μM of MS2 RNA forward and reverse primers, 3.5 pmol ml−1 of MS2 RNA and 0.125 U μl−1 of Ribolock RNase inhibitor and cycled in a QuantStudio. Relative amounts of reverse transcriptase activity were determined as the rate of transcription of bacteriophage MS2 RNA, with absolute reverse transcriptase activity calculated by comparing the relative amounts of reverse transcription to a reverse transcriptase standard of known activity.

### Serum/plasma pseudotype neutralization assay

Spike pseudotype assays have been shown to have similar characteristics as neutralization testing using fully infectious wild-type SARS-CoV-28. Virus neutralization assays were performed on HEK293T cells transiently transfected with ACE2 and TMPRSS2 using SARS-CoV-2 spike pseudotyped virus expressing luciferase48. Pseudotyped virus was incubated with serial dilutions of heat-inactivated human serum samples or convalescent plasma in duplicate for 1 h at 37 °C. Virus- and cell-only controls were also included. Then, freshly trypsinized HEK293T ACE2- and TMPRSS2-expressing cells were added to each well. After 48 h incubation in a 5% CO2 environment at 37 °C, luminescence was measured using Steady-Glo Luciferase assay system (Promega).

### Monoclonal antibody pseudotype neutralization assay

Virus neutralization assays were performed using HeLa cells stably expressing ACE2 and using SARS-CoV-2 spike pseudotyped virus expressing luciferase as previously described49. Pseudotyped virus was incubated with serial dilutions of purified monoclonal antibodies9 in duplicate for 1 h at 37 °C. Then, freshly trypsinized HeLa ACE2-expressing cells were added to each well. After 48 h incubation in a 5% CO2 environment at 37 °C, luminescence was measured using a Bright-Glo Luciferase assay system (Promega) and neutralization was calculated relative to virus-only controls. IC50 values were calculated in GraphPad Prism.

### Ethics

The study was approved by the East of England–Cambridge Central Research Ethics Committee (17/EE/0025). Written informed consent was obtained from both the patient and his family. Additional control patients with COVID-19 were enrolled to the NIHR BioResource Centre Cambridge under ethics review board approval (17/EE/0025).

### Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this paper.