Clinical trial design, patients and conduct

The Algonquin study is an ongoing, multicentre, open-label, single-arm, two-part study (ClinicalTrials.gov identifier: NCT03715478)) of BPd in patients with relapsed/refractory MM. Part 1 of the trial consisted of a dose-exploration phase²². Part 2 is evaluating the safety, tolerability and clinical activity of the dose and schedule identified in part 1. Patients included in this interim analysis were enrolled in cohorts 2a and 2b, with the last patient being enrolled on 14 September 2022. The clinical data cut-off date for the presented analysis was 28 October 2025.

At screening, patients with relapsed/refractory MM were eligible if they met the following criteria: aged ≥18 years; had an Eastern Cooperative Oncology Group performance status of 0–2; had undergone autologous stem cell transplantation or were considered transplant ineligible; had experienced disease progression after ≥1 previous lines of antimyeloma treatment and must have been lenalidomide refractory and proteosome inhibitor exposed (in separate regimens or in combination); and had adequate BM, renal and cardiac function. Patients with previous pomalidomide or BCMA-target therapy exposure, concurrent corneal epithelial disease (except mild punctate keratopathy), any serious and/or unstable preexisting medical condition, a psychiatric disorder or any other condition (including laboratory abnormalities) that could interfere with the patient’s safety, obtaining informed consent or compliance with the study procedures were excluded. Sex was not considered in the study design or patient selection and was determined on the basis of self-reporting.

The Algonquin study was conducted at nine Canadian sites in accordance with the Declaration of Helsinki, International Council Harmonisation Good Clinical Practices Guidelines, local regulation governing the conduct of clinical studies, and institutional guidelines. All patients provided written, informed consent. The study protocol, amendments and informed consent were approved by the University Health Network Institutional Review Board (REB numbers 16-5260, 18-5911 and 17-5357). Eligibility criteria, trial procedures and end points, and participating sites conducting this trial, are as previously described²². In summary, belamaf was administered using the following dosages: a 30-min intravenous infusion every 4 weeks at a dose of 1.92 mg kg⁻¹ (cohort 1); or every 4, 8 or 12 weeks at a dose of 2.5 mg kg⁻¹ (cohorts 1a, 3a and 3b); or at total doses of 2.5 or 3.4 mg kg⁻¹ once every 4 weeks but split evenly with 50% of the dose administered on days 1 and 8 of every cycle (cohorts 1b and 2, respectively); or as a loading dose of 2.5 mg kg⁻¹ on cycle 1 day 1, followed by a dose reduction to 1.92 mg kg⁻¹ from cycle 2 onwards once every 4 weeks (cohort 1c). Pomalidomide was administered at 4 mg on day 21 of 28 and dexamethasone 40 mg (20 mg if aged >75 years) weekly. Patients remained on treatment until disease progression, unacceptable toxicity or consent withdrawal. Ophthalmology examinations before each dose of belamaf and preservative-free lubricant eye drops were required throughout study treatment. Dose modifications were made independently for each drug according to predefined criteria based on the nature and toxicity grade of the event.

Acquisition of patient samples

For the Algonquin patient cohort, BM and peripheral blood samples were collected from enrolled patients at the time points indicated in Fig. 1a. Patient samples included in this investigation were selected and sequenced based on availability on an on-going basis. BM samples were included for patients with available matched baseline and C2D1 samples, except for patient ALQ-19-019, in whom poor viability prohibited baseline BM sequencing. No patient peripheral blood samples were excluded except in cases when poor sample quality prohibited sequencing or in specific analyses, which is indicated in figure legends, where applicable. For the distinct cohorts of patients used to train PreGame (discovery cohort and validation cohorts 1 and 2, n = 22), BM samples were obtained from patients with MM receiving standard-of-care BM aspirates at the Princess Margaret Cancer Centre or purchased from commercial sources (iSpecimen and Discovery Life Sciences). Samples were obtained from a mixture of untreated patients and patients having received at least one or more previous lines of therapy. BM mononuclear cells were isolated using Ficoll according to the manufacturer’s recommended protocol (Sigma) and cryopreserved in fetal calf serum (FCS) + 10% DMSO. Melanoma and lung cancer samples were obtained from surgical resections of patients receiving standard-of-care therapy at the Princess Margaret Cancer Centre. Tumour samples were digested with GentleMACS (Miltenyi) and single-cell suspensions were cryopreserved. All BM biopsy samples from patients were taken from the pelvic bone. See Supplementary Table 1 for additional details on patients in our sequencing cohorts, where available. All samples were collected with informed patient consent using protocols approved by the appropriate Institutional Review Boards.

Cell lines

The MM cell lines RPMI-8226 and MM1R were purchased from the American Type Culture Collection (ATCC), ALMC1 was purchased from Sigma, and INA6, JJN3, EJM and AMO1 were purchased from DSMZ. All MM cell lines were cultured in complete IMDM, which is IMDM (Gibco) supplemented with 10% FCS, 2 mM l-glutamine, 1% non-essential amino acids, 1% penicillin–streptomycin, 10 µg ml^–1 gentamicin and 55 µM β-ME (all purchased from Gibco). Human IL-6 (BioLegend) was added to the culture media for ALMC1 (1 ng ml^–1) and INA6 (10 ng ml^–1) cells. To generate CD1D-overexpressing myeloma cell lines, the human CD1D full-length construct was transduced into MM cell lines using lentivirus, and CD1D⁺ cells were purified by FACS. Jurkat-76 cells were obtained from N.H. and cultured in the same complete IMDM as the myeloma cell lines. The melanoma cell lines A-375 (ATCC), A2058 (ATCC), 624Mel, 526Mel, 888Mel (provided by S. Rosenberg) and WM3670 (Rockland), the lung cancer cell lines A549, H1568, H1573, H1437, H1792, H1944 and H2030 (all from ATCC), and the glioblastoma cell lines LN-229 (ATCC) and A-172 (ATCC) were cultured in complete DMEM (Gibco) + 1% GlutaMAX (Gibco) and supplemented in the same manner as complete IMDM. The small-cell lung carcinoma cell line H69PR (ATCC), mesothelioma cell lines H2452 (ATCC) and MSTO-211H (ATCC), breast cancer cell lines HCC1143 (ATCC) and HCC1937 (ATCC), the pancreatic cancer cell line ASPC1 (ATCC) and the colon cancer cell lines LoVo (ATCC) and Colo205 (ATCC) were cultured in complete RPMI-1640 (Gibco) supplemented in the same manner as complete IMDM. The pancreatic cancer cell line CFPAC-1 (ATCC) was cultured in complete IMDM medium. The Merkel cell carcinoma cell lines MCC14-2, MCC26, MS-1 and MCC13 were purchased from Sigma and cultured in RPMI-1640 (Gibco) supplemented with 20% FCS, 2 mM l-glutamine, 1% penicillin–streptomycin, 25 mM HEPES and 10 µg ml^–1 gentamicin. Lenti-X 293T cells were purchased from Takara and cultured in complete DMEM as described above. Healthy primary cells used for screening were obtained from healthy donors and obtained from the ATCC and Sigma and cultured according to the manufacturers’ recommended media and protocols. Tumour cell lines used were confirmed to be free of mycoplasma.

Antibodies and flow cytometry

The following antibodies were used in this study for flow cytometry as indicated: CD3 (OKT3; BioLegend, 317308 or 317322, 1:400); B2M (A17082A; BioLegend, 395716, 1:800); CD69 (FN50; BioLegend, 310910, 1:400); CD8 (RPA-T8; BioLegend, 301049, 1:400); CD4 (OKT4; BioLegend, 317420, 1:200); NGFR (ME20.4; BioLegend, 345132, 1:200); γδTCR (B1; BioLegend, 331210, 1:100); αβ TCR (IP26; BioLegend, 306720, 1:100); CD14 (61D3; ThermoFisher, 53-0149-42, 1:50); CD2 (RPA-2.10; BioLegend, 300230, 1:75); CD235 A/B (HIR2; BioLegend, 306614, 1:200); CD138 (MI15; BioLegend, 356504, 1:100); CD319 (235614; BD, 750838, 1:20); HLA-ABC (W6/32; BioLegend, 311405, 1:50); HLA-A (1082C5; BD, 568024, 1:100); HLA-C (DT-9; BD, 566372, 1:100); BTN3 (232-5; BD, 566006, 1:100); Bw4 (REA274; Miltenyi, 130-123-760, 1:100); KIR3DL1 (DX9; BioLegend, 312707, 1:100); CD94 (HP-3D9; Invitrogen, 17-5094-4210, 1:100); GPR56 (CG4; BioLegend, 358205, 1:100); 4-1BB (4B4; BioLegend, 309809, 1:200); CD1D (51.1; BioLegend, 350305, 1:100); and CD1C (L161; BioLegend, 331505, 1:100). Trustain Fc Block (BioLegend, 422302) was used to block Fc receptors, and fixable viability dye (Invitrogen, 65-0865-14), with or without Annexin V staining (BioLegend, 640920), was used to identify live and apoptotic cells. Unless used for cell sorting or annexin V staining, cells were fixed with 4% paraformaldehyde before analyses. Flow cytometry data were acquired on a BD LSR Fortessa X20 or Beckman Coulter CytoFLEX instrument and analysed using FlowJo software (v.10, BD Biosciences)

Generation of B2M-deficient NUR77 T2A eGFP knock-in Jurkat-76 cells

To generate human NUR77 T2A eGFP knock-in Jurkat cells, we modified Jurkat-76 cells, an αβ TCR-negative derivative of the CD8⁻ human T cell lymphoma Jurkat cell line⁵². A CRISPR design tool (http://crispor.tefor.net/) was used to identify a unique single guide RNA (sgRNA) with CRISPR target site 5′-TTCCCAGGCAGGGGTCAGAA-3′, which cuts near the 3′ stop codon of the NUR77 coding region in the human genome. The corresponding sgRNA was chemically synthesized by Synthego. We then designed a 1,761-bp homology-directed repair (HDR) template with a 5′ homology region containing the 3′ coding region (exon 8) of NUR77, a furin cleavage site (RRKR), a flexible linker (SGSG), a T2A (Thosea asigna virus 2A) ribosome-skipping peptide sequence, an eGFP open reading frame and a 3′ homology region containing the NUR77 3′ UTR. To prevent CRISPR targeting of the HDR template, 17 out of the 20 nucleotides of the NUR77 sgRNA target sequence were removed from the HDR sequence. The double-stranded HDR DNA fragment was then synthesized by Twist Biosciences. The corresponding sense and antisense single-stranded DNA (ssDNA) HDR templates were prepared using a Guide-it Long ssDNA production system v.2 kit (Takara Biosciences) and PrimeStar Max DNA polymerase with the following 5′ phosphorylated and unphosphorylated sense antisense primers: NUR77_U1 5′-CGTTTGTGACCCTGGGCAAGTCATTAGC-3′ and NUR77_L1 5′-CTGTCATTTGTCTTCTTCCTAGAGTGAC-3′. RNP complexes were generated by mixing sgRNA with SpCAS9 2NLS nuclease protein (both from Synthego) in electroporation buffer at room temperature for 10 min. Sense or antisense NUR77-T2A–eGFP HDR ssDNA HDR template (1.5 µg) was added to the mix, which was used to electroporate Jurkat-76 cells using an Amaxa 4D Nucleofector instrument (Lonza) with the CL-120 program. FACS sorting of eGFP⁺ cells was used to enrich for successful HDR CRISPR targeting of Jurkat-76 cells as a measure of TCR activation of NUR77 and eGFP expression after low-dose PMA–ionomycin (Invitrogen) treatment. Following establishment of the cell line, we knocked out B2M by electroporating CAS9 and B2M sgRNA (Supplementary Table 2) and confirmed the knockout using flow cytometry.

TCR constructs

To normalize TCR surface expression, all endogenous γδ TCR signal peptide sequences were mapped using the Department of Health Technology SignalP−6.0 signal peptide prediction service and replaced with the fibroin light chain (FIBL) signal peptide sequence from Dendrolimus spectabilis. RefSeq and nucleotide sequence alignments from the NCBI database were used to create consensus sequence file templates for human TCR γ2, 3, 4, 5, 8 and 9 with either constant region 1 or 2 as well as TCR δ1, 2 and 3. Sequences were codon-optimized for optimal expression in human cells using the GenSmart codon optimization tool from GenScript. To normalize expression, only amino acid differences in the paired γδ TCR framework 3 (FR3), FR4 regions plus the unique hypervariable third complementarity-determining region 3 (CDR3) sequences from the single-cell sequence data were appended into each expression template. To facilitate cloning, a silent SacII restriction endonuclease site was engineered into the region coding for amino acids Gly15 and Thr16 of the constant region of TCRδ. The approximately 1,500 bp T2A-linked γδ TCRs with FIBL signal sequences were synthesized as fragments by Twist Biosciences and PCR-amplified using Q5 high-fidelity DNA polymerase (New England Biolabs, M0491L) and the sense primer XhoI_FiBL_GD_U1: 5′-ATTAGCGCTTCTCGAGGCCACCATGATGAGACCTATTGTG-3′ plus the antisense primer KpnI_DeltaC_L1 5′-AGGCATGCCACGTTGGTACCATTTTTCATCACGAACAC-3′ (bold bases indicate engineered restriction endonuclease cut sites). PCR products were gel-purified and cloned into a pTrans-NGFR-Puro expression vector using an In-fusion HD Cloning kit (Takara Biosciences). pTrans-NGFR-Puro is a novel engineered mammalian gene expression plasmid based on the pcDNA3.1 (Invitrogen) vector backbone. It consists of the 1,172 bp EF1A promoter with an intronic enhancer driving expression of the human TCRγ chain, followed by a furin cleavage site (RRKR), a flexible linker (SGSG), a T2A self-cleaving ribosome skipping peptide sequence, a TCRδ sequence and a BGH PolyA signal. A separate expression cassette was engineered into the vector consisting of the mouse Pgk1 promoter driving expression of the extracellular and transmembrane regions of human NGFR (amino acids 1–267) followed by a furin cleavage signal, the P2A (Porcine teschovirus-1 2A) self-cleaving peptide and puromycin for antibiotic selection. Both the EF1A promoter plus TCRγδ, and PGK NGFR-P2A-Puro cassettes were then engineered to be flanked by the left and right terminal inverted repeat sequences (L-TIR and R-TIR) from the Sleeping Beauty (SB) transposon system, as previously described⁵³. All plasmids were transformed into Stellar chemically competent Escherichia coli cells (Takara Biosciences), and transfection-grade plasmids were purified using a Nucleospin Plasmid Transfection-grade kit (Macherey-Nagel, 740490). Plasmids were used to electroporate Jurkat-76 NUR77-T2A–eGFP cells along with 5′ ARCA capped SB100x Sleeping Beauty transposase mRNA. Lentiviral γδ TCR expression plasmids were engineered into a lentiviral backbone modified from pKLV2 (Addgene plasmid 67974). In brief, each human FIBL-TCRγ-Furin-Linker-T2A-FIBL-TCRδ from pTrans-NGFR-Puro was subcloned in-frame into an amino-terminal human NGFR (amino acids 1–267)-furin (RRKR)-linker (SGSG)-P2A under the control of the 232-bp core EF1A promoter. Transfection-grade lentiviral plasmid was purified as above and used to transduce primary T cells.

Sleeping Beauty SB100X ARCA-capped mRNA production

The hyperactive SB100X variant of the Sleeping Beauty transposase has approximately 100 times the transposition efficiency of the original SB10, and has previously been shown to promote efficient transposition and genetic modification of human T cells while bypassing many of the biosafety issues and time delays associated with retroviral and lentiviral approaches⁵⁴. We decided to use the SB100X mRNA as a high-throughput means of ensuring rapid and stable insertion of our γδ TCRs into NUR77-T2A–eGFP Jurkat-76 cells. These cells were engineered to be αβ TCR-negative and to express eGFP after TCR activation through the NUR77 locus. The SB100X coding sequence (Addgene plasmid 127909)⁵⁵ was codon-optimized for expression in human cells, synthesized by Twist Biosciences as a gene fragment, and subcloned into pxInVitro, a novel in vitro mRNA expression vector. The pxInVitro vector was designed to generate in vitro mRNA products using a protocol modified from previously described methods⁵⁶. The vector contains a T7 promoter, mouse α-globin 5′ UTR, codon-optimized SB100X, followed by the mouse α-globin 3′ UTR. The pxInVitro SB100X vector was linearized at the 3′ end of the α-globin 3′ UTR with XhoI restriction enzyme. The resulting linearized plasmid was electrophoresed on a 1% agarose gel, excised and gel-purified using a Nucleospin Gel and PCR clean-up kit from Macherey-Nagel (740609). The entire T7-SB100x expression cassette was PCR-amplified using 30 cycles with Q5 high-fidelity DNA polymerase (New England Biolabs, M0491L) with pIVT Fwd (27-mer) 5′-TTGGACCCTCGTACAGAAGCTAATACG-3′ and pIVT Rev T120 primers (149-mer) 5′-T₁₂₀CTTCCTACTCAGGCTTTATTCAAAGACCA-3′. The Q5-amplified dsDNA tailed Poly(T) T7-SB100x cassette was then electrophoresed on a 1% agarose gel, gel-purified using a NucleoSpin Gel and PCR clean-up kit from Macherey-Nagel, and eluted in RNase-free NE buffer at 70 °C. A MEGAscript T7 In Vitro RNA production kit (Ambion, Life Technologies, AM1334) was used to generate capped polyA⁺ transposase mRNA from the dsDNA tailed T7-SB100X transposase cassette template according to the manufacturer’s instructions with the following modifications. To avoid mRNA degradation through cellular immunity pathways, we used pseudouridin-5′-trisphosphate (pseudo-UTP; TriLink Bio Technologies, N-1019-10) and 5′-methylcytidine-5′-triphosphate (5Me-CTP; TriLink N-1014-10)⁵⁷. We also used 5′ Anti-Reverse Cap Analog (ARCA 3′O-Me-⁷ G(5′)ppp(5′)G; TriLink, N-7003-10) for improved mRNA stability and enhanced translation⁵⁸. In brief, the T7 MEGAscript reaction was performed for 4.5 h at 37 °C, dephosphorylated with Antarctic phosphatase (New England Biolabs, M0289L) for 45 min at 37 °C, and purified using a MEGAclear kit (Ambion, AM1908). 5′-Capped transposase mRNA was quantified using a NanoDrop ND-1000 spectrophotometer (Thermo Fisher Scientific) and assessed for integrity and quality using native non-denaturing 1.2% agarose TAE electrophoresis with a single-stranded RNA ladder (NEB, N0362S). Aliquots of 400 ng ml^–1 5′ ARCA-capped transposase mRNA were frozen at −80 °C for future use.

Jurkat cell-based γδ TCR screening

γδ TCR plasmids (1 µg each) plus 1 µl of 400 ng µl^–1 transposase mRNA were combined with 1 × 10⁶ Jurkat-76 NUR77-T2A–eGFP cells in pre-warmed Lonza electroporation SE buffer (Lonza) and electroporated using an Amaxa/Lonza 4D Nucleofector System with the CK116 program according to the manufacturer’s recommended instructions. After 48 h, puromycin (Invivogen) was added to the culture to enrich for TCR-expressing cells. To screen for tumour reactivity, TCR-expressing Jurkat cells were mixed at a 1:1 ratio with MM cell lines and incubated for 16–18 h at 37 °C followed by flow cytometric analysis. The CD1C-restricted JR.2 TCR⁵⁹ served as a negative control because the MM cell lines do not express CD1C (Extended Data Fig. 9l). We defined a TR γδ TCR as causing a ≥15% increase in the absolute number of GFP⁺ Jurkat cells compared with Jurkat cells cultured without MM cells (GFP⁺ cells (%) in the CD3⁺ fraction in Jurkat cells cultured with MM cell line – GFP⁺ cells (%) in the CD3⁺ fraction in unstimulated Jurkat cells), referred to as ‘activation strength’. When culturing Jurkat cells with healthy cells, the GFP⁺CD69⁺ double-positive fraction was used to calculate the activation strength. For some experiments with cell lines or healthy cells that did not express CD1D, the CD1D-restricted TCR 9C2 was used as a negative control⁶⁰. Cells demonstrating a 13–15% increase in activation strength were considered borderline and excluded from bioinformatic analyses involving TR T cell populations to reduce potential false positives or false negatives. A 0–13% increase in activation strength was considered background and these TCRs were classified as non-reactive. The 0–13% threshold for defining non-reactive cells was established on the basis of the activation strength distribution of the negative-control TCR in the initial discovery cohort. Specifically, the 13% cut-off corresponds to approximately three standard deviations above the mean activation strength and exceeds the maximum observed value for the negative-control TCR across 98 measurements. When testing γδ TCR reactivity to patient-matched MM cells, BM mononuclear cells or CD138⁺ magnetically sorted tumour cells (Miltenyi) were mixed at a 1:1 ratio with Jurkat NUR77–GFP cells expressing the indicated γδ TCR, and GFP expression was assessed in the Jurkat cells after overnight co-culture. In some instances, Jurkat NUR77–GFP cells engineered to express KIR3DL1 and/or CD8a were used (as indicated in the figure legends), and treated with the KIR3DL1 antibody (clone DX9 at 10 μg ml^–1; R&D Systems). To stimulate the Vδ2Vγ9 TR TCR, the known Vδ2Vγ9 ligand HMBPP (Sigma) was added to the cell culture at 1 μM. A complete list of all validated MM-reactive γδ TCR CDR3 sequences can be found in Supplementary Table 3.

Lentivirus production

For lentivirus generation, Lenti-X 293T cells (Takara Biosciences) were cultured in DMEM supplemented with 10% FCS, 2 mM l-glutamine, 1% penicillin–streptomycin and 1% GlutaMAX (all from Gibco). Cells were seeded to achieve 80% confluence after an overnight incubation. The culture medium was replaced with viral production medium, which was the same DMEM formulation as above but contained only 0.1% penicillin–streptomycin. A transfection mixture was prepared in Opti-MEM (Gibco) containing pre-warmed TransIT-lenti transfection reagent (Mirus), 2 µg µl^–1 psPAX2, 2 µg µl^–1 VSVG (both from Addgene) and 2 µg µl^–1 of the lentiviral backbone plasmid containing the gene of interest. The culture supernatant was collected at 48 h after transfection, filtered through a 0.45 µm PVDF syringe filter (Millipore Sigma) and concentrated with Lenti-X concentrator (Takara Biosciences) according to manufacturer’s protocol. The pellet was resuspended in viral production medium and stored at −80 °C in single-use aliquots.

TR γδ T cell expansion from patient BM

Cryopreserved BM samples from patients with MM were thawed and flow-sorted to stain viable γδ T cells double-positive or double-negative for CD94 and GPR56. Cells were sorted into IMDM (10% human serum with l-glutamine, 1% penicillin–streptomycin, 0.1% gentamycin and fungizone) containing irradiated K562 feeder cells expressing membrane-bound IL-21, CD48 and 4-1BB ligand (provided by N.H.), with the addition of IL-2 (1,000 IU ml^–1), IL-15 (20 ng ml^–1) and PHA-L (2.5 µg ml^–1, eBioscience). Cells were cultured for 7–10 days to allow expansion and subsequently used for co-culture with MM cell line targets (either RPMI-8226 alone or a mixture of RPMI-8226, AMO1 and MM1R cells), with or without the addition of 10 µg ml^–1 of either anti-γδ-TCR blocking antibody (clone B1; BioLegend) or isotype control antibody (clone MOPC-21; BioLegend) as indicated in the figure legends. Tumour cell apoptosis was assessed at the end of co-culture by gating on CD138⁺ tumour cells, which were considered apoptotic if positive for Annexin V or the viability dye as determined by flow cytometry.

T cell cytotoxicity assay

Human PBMCs of healthy donors were purchased from StemCell Technologies. Untouched primary pan-T cells were isolated from PBMCs using a negative-selection magnetic separation kit (Miltenyi Biotec) according to the manufacturer’s instructions and cultured in T cell medium (RPMI-1640 containing 10% human serum AB (GeminiBio), 1% penicillin–streptomycin and 0.1% gentamicin). The gene encoding αβ TCR was knocked out using a P3 Primary Cell 4D Nucleofector kit (Lonza) with the EH-115 program using Cas9 and a sgRNA (sequence is provided in Supplementary Table 2) (Synthego) according to the manufacturer’s recommended protocol. After electroporation, cells were resuspended in T cell medium and co-cultured with irradiated mOKT3 cells⁶¹ at a 10:1 ratio. At 24 h after activation, 10 ng ml^–1 IL-15 (BioLegend) and 20 ng ml^–1 IL-2 (BioLegend) were added to the culture. At 48 and 72 h after initial activation, γδ TCR lentiviruses were added to the culture and spinfected for 2 h at 1,000g at 32 °C. NGFR-expressing T cells were isolated 7 days after lentiviral transduction using CD271 magnetic beads (Miltenyi Biotec) according to the manufacturer’s instructions. To assess cytotoxicity, transduced primary T cells were co-cultured with RPMI-8226 cells expressing luciferase at a 2:1 or 0.5:1 effector:target ratio for 18–20 h at 37 °C, followed by luciferase quantification using a Steady-Glo Luciferase assay (Promega) according to the manufacturer’s recommended protocol. Cytotoxicity was calculated by measuring the difference in luminescence between RPMI-8226 luciferase cells cultured with or without γδ TCR-transduced T cells. To account for varying background activity when pooling results from multiple donors, the highest effector:target ratio for each donor that demonstrated <35% background killing in the CD1C negative-control TCR was used for each donor.

γδ TCR ligand identification

γδ TCR-expressing NUR77–GFP-expressing Jurkat cells were cultured for 16–18 h with MM cell lines in which B2M, BTN3A, HLA-A, HLA-C or total HLA-ABC were knocked out using CRISPR (sgRNAs are listed in Supplementary Table 2). Cell lines were flow-sorted to achieve high purity, and the efficiency of disruptions was confirmed by flow cytometry (Extended Data Fig. 9m). To assess the reactivity of γδ TCR65 to different HLA-C alleles, artificial antigen-presenting cells expressing single HLA alleles were generated as previously described⁴³ and cultured with γδ TCR65 Jurkat cells as described above. HLA-C*03:04 mutants were synthesized as gene fragments (Twist Bioscience) and subcloned into pLenti plasmids to generate lentiviruses. These lentiviruses were used to stably express the mutants in RPMI-8226 cells that had had their endogenous HLA-C knocked out using CRISPR.

Preparation of cells for single-cell RNA CITE, TCR and BCR sequencing

For the Algonquin clinical trial patient cohort, cryopreserved patient BM samples were thawed and viable CD2⁺, CD3⁺, CD14⁺ and CD34⁺ cells were isolated by flow cytometry. Before flow-sorting, cells were labelled with TotalSeq-C Human antibody cocktail 1.0 (BioLegend), and additional TotalSeq-C antibodies recognizing BCMA (C0056), CD138 (C0831) and FcRL5 (C0829) according to the manufacturer’s recommended protocol. For the distinct MM patient cohorts used to train PreGame (discovery and validation cohorts 1 and 2, n = 22), cryopreserved patient BM samples were thawed and viable CD3⁺ T cells and CD138⁺ and CD319⁺ MM cells were isolated by flow cytometry as indicated in Extended Data Fig. 4a. Before flow-sorting, cells were labelled with TotalSeq-C antibody cocktail 1.0 (BioLegend) and TotalSeq-C antibodies recognizing CD138, BCMA and FcRL5, as well as TotalSeq-C Hashtag antibodies (BioLegend; to allow for multiplexing of patient samples) according to the recommended protocol. Melanoma and lung cells were processed as described above in that they received TotalSeq-C Human antibody cocktail 1.0, but only the CD3⁺ or CD45⁺ fraction, respectively, was enriched by flow-sorting. Sorted cells were immediately used for single-cell library generation using a 10x Genomics Chromium 5′ kit V2 by the Princess Margaret Genomics Centre. γδ TCR libraries were prepared using two previously described γδ TCR enrichment primer sets, which generated two distinct libraries for sequencing^62,63. All resulting libraries were sequenced on an Illumina NovaSeq 6000 sequencer at the Princess Margaret Genomics Centre.

Alignment of raw sequencing data

Single-cell gene expression (GEX), surface protein counts (CITE), BCR sequences and TCR sequences were obtained using the 10x CellRanger pipeline (v.7.0.1) with the GRCh38-2020-A, TotalSeq-C Human Universal Cocktail v.1 399905 (plus C0831, C0056, C0829 barcodes) and vdj-GRCh38-alts-ensembl-7.0.0 references, respectively. For multiplexed samples, raw GEX and CITE fastq files were first demultiplexed using cellranger multi and setting feature_types = Multiplexing Capture for CITE fastqs. Demultiplexed BAM files were then converted to fastq files using cellranger bamtofastq with –reads-per-fastq set to the total number of reads obtained from the previous demultiplexing alignment. Demultiplexed GEX fastq files were then aligned in tandem with CITE (specifying feature_types = Antibody Capture), BCR and TCR fastq files using cellranger multi, with force-cells set to the number of cells determined from the first demultiplexing alignment. Alignments of multiplexed melanoma samples were performed as described above but using CellRanger (v.8.0.1). For samples that were not multiplexed (Algonquin trial samples), raw GEX, CITE (specifying feature_types = Antibody Capture), TCR and BCR fastq files were aligned in tandem using cellranger multi with default parameters. NSCLC (lung) samples were aligned with CellRanger (v.6.1.2) using the cellranger count function.

Alignment of raw γδ TCR sequencing data

γδ TCR alignment is not officially supported by 10x CellRanger, and our use of two primer sets necessitated a custom alignment pipeline that was based on previously described alignment solutions^1,62,63 (Extended Data Fig. 4b). In brief, a custom VDJ reference was generated based on the vdj-GRCh38-alts-ensembl-7.0.0 reference used for αβ TCR alignment. The reference was duplicated and the regions.fa file was modified to convert the names of the TRG and TRD gene names into TRA and TRB gene names, respectively, with the following commands: sed -i ‘s/TRG/TRA/g’ regions.fa; sed -i ‘s/TRD/TRB/g’ regions.fa. Next, the raw fastq files generated from each of the two primer sets used in library preparation (henceforth, Mimitou (M) and Gherardin (G)) were individually aligned against the custom reference using cellranger multi with feature_types = VDJ-T. The raw fastq files were then aligned against the custom reference in tandem again using cellranger multi with feature_types = VDJ-T. This protocol resulted in three sets of γδ TCR sequences per barcode outputs: M only, G only and GM combined, which remained multiplexed when sample multiplexing was used. From each of the three outputs, the TRA and TRB gene names were converted back to TRG and TRD gene names, respectively, and barcoded TRG and TRD sequences were separately extracted from the filtered_contig_annotations.csv file. To retain the full sequence information (for downstream cloning and expression in modified Jurkat-76 cells) but also to retain per barcode clonotype assignments from cellranger (as for αβ TCR alignments), a ‘clonotype key’ and ‘full sequence key’ were created for each sequence. The clonotype key comprised the V, D and J genes and the nucleotide sequence of the CDR3 region, whereas the full sequence key comprised the full TCR nucleotide sequence (fwr1, cdr1, fwr2, cdr2, fwr3, cdr3 and fwr4) and gene usage. For cases when unique clonotypes differed in full sequence, full sequences were concatenated to derive lists of unique clonotypes for each chain. TRG and TRD clonotype lists were then joined to the clonotypes.csv alignment output to recapitulate γ and δ TCR pairing, and barcodes were reassigned with cellranger–determined clonotypes but now with full sequence information. These per barcode clonotype lists derived from each of the three alignments were then merged. The majority of barcodes and γδ TCR sequences were common between the three alignments, which was expected owing to the close similarities of the G and M primer sequences. However, this method improved the total number of barcodes with γδ TCR sequences recovered, and often resulted in additional γ and δ pairs that would have been seen as unpaired (that is, orphan) clonotypes if only one primer set had been used.

For compatibility with Seurat downstream, the merged list was collapsed into unique barcodes, assigning each barcode to only one clonotype (that is, γδ TCR sequence) using the following decision tree in cases when the alignments disagreed: (1) paired clonotypes were always selected over orphan clonotypes; (2) if both clonotypes were paired, but one clonotype had additional γ or δ sequences (that is, dual TCR), the dual TCR was favoured; (3) if both clonotypes were paired and not dual TCRs, the source alignment was considered in the following order: found in all (G, M, GM) alignments > G and M alignments but not GM > G or M alignments alone > found in the combined GM alignment only (when we expected alignment artefacts to be most prevalent if indeed our pipeline did introduce any artefact TCRs); and (4) the clonotype with the highest frequency was favoured. This custom pipeline enabled us to maximize the number of γδ TCR sequences obtained from true γδ T cells while theoretically limiting the number of artefactual γ and δ TCR pairs introduced. We viewed these aspects as fundamentally crucial to both accurate cell-type annotation (described below) and faithful expression of bona fide patient γδ TCRs in Jurkat-76 cells for tumour-reactivity screening (described above). To assess the performance of the custom pipeline, we also aligned raw γδ TCR fastqs from each primer set using the standard 10x CellRanger method (which is implemented but not officially supported). This method uses cellranger multi with feature_types = VDJ-T-GD (instead of VDJ-T as above) and specifies the inner-enrichment-primers argument with the sequences CCACAATCTTCTTGGATGATCTGAGACT, GTCCCAGTCTTATGGAGATTTGTTTCAGC. The total number of unique paired γδ TCR sequences obtained from each patient with MM that would be available for cloning and in vitro screening resulting from this pipeline is reported in Supplementary Table 4.

Quality control, normalization and single-cell clustering

Quality control was first performed at an individual sample level before sample integration and re-normalization. Single-cell analyses were performed in the R (v.4.2.1) statistical environment, unless otherwise indicated. For multiplexed samples, the assignment confidence table (generated by cellranger multi as described above) was used to filter out blank or unassigned droplets, multiplets and cells assigned to other multiplexed samples, thereby keeping only cells confidently assigned to the expected hashtagging antibody. To retain cells with high-quality GEX and CITE libraries only, cells were further filtered out if they were outliers in terms of low surface protein library size, expressed fewer than 200 features or showed mitochondrial gene expression of greater than 10%. Genes expressed in fewer than three cells were excluded. GEX libraries from high-quality cells were then converted into Seurat objects (v.4.3.0)⁶⁴ and normalized using SCTransform (v.0.3.5)⁶⁵ with the parameters method = “glmGamPoi”, vst.flavor = “v2”, variable.features.n = 3000, and vars.to.regress = “percent.mt” to regress unwanted variance associated with mitochondrial content. Cells were then clustered using Seurat’s shared-nearest neighbour algorithm following Louvain modularity optimization. Dimensionality reduction was performed using principal component analysis and UMAP. CITE data were normalized using the CLR method across cells as implemented in Seurat, and clustered as described above. Finally, to jointly cluster cells on the basis of GEX and CITE, a weighted nearest neighbour graph was constructed and used as the basis for the final UMAP embedding (wnnUMAP). BCR, αβ TCR and γδ TCR sequences obtained from VDJ alignment were then added to the Seurat object to retain sequences from high-quality (and demultiplexed, where applicable) cells only. Using a combination of what immune receptors were sequenced along with gene and protein expression of canonical markers of lymphoid, myeloid and malignant cells, cursory cell-type annotations were assigned, and any remaining clusters of low-quality cells were identified and excluded. Following this sample-level quality control, all high-quality cells from samples of the given sequencing cohort were merged and re-normalized using the same methods described above. To account for batch effects typical of cohorts sequenced over long periods of time, we used harmony (v.0.1.1)⁶⁶, grouping by either batch or sample and a modest theta value of 0. The harmony and CITE embeddings were then used as the basis for constructing a weighted nearest neighbour graph and clustering cells as described above. Cursory cell-type annotation assigned at the sample-level (as described above) was used to avoid overcorrection by harmony, thereby limiting the loss of relevant biological variation.

Identification of cell types

Cluster-specific differentially expressed genes and proteins were determined using Seurat’s FindAllMarkers function, and cell-type annotation was refined compared to sample-level cell annotations (Extended Data Fig. 4e). In brief, malignant MM cells were identified by clonal expression of the dominant BCR sequence in a given sample and expression of SDC1 (which encodes CD138) and TNFRSF17 (which encodes BCMA). Clusters of cells with polyclonal BCRs and expression of CD79A and CD19 were annotated as normal B cells. Expression of an αβ TCR and expression of canonical T cell markers were used to annotate the various subsets of T cells. To stringently annotate cells as γδ T cells, we considered expression of a γδ TCR, cluster membership, expression of TRDV genes and expression of other canonical T cell genes. As the gene encoding γTCR can be rearranged and expressed in αβ T cells, only cells with a paired γδ TCR sequence, or cells with an unpaired (γ or δ) TCR sequence but not a paired αβ TCR sequence, were annotated as true γδ T cells. Cells without a γδ TCR sequence but with TRDV gene expression and close clustering in γδ T cell clusters were also annotated as γδ T cells. This approach produced two transcriptional clusters of γδ T cells marked by their expression of GZMB and GZMK, with close clustering to their αβ T cell counterparts. Although the γδ TCR constant genes TRDC and TRGC1 were expressed by NK cells, NK cells could be differentiated from γδ T cells by their lack of TCR sequences. Finally, sparse myeloid cells were annotated by canonical markers (Extended Data Fig. 4e).

TCR repertoire diversity and similarity

Single-cell TCR repertoire diversity was assessed using richness or Shannon diversity where indicated, and repertoire similarity was assessed using the Morisita–Horn index. Richness was calculated as the total number of unique CDR3 sequences detected in a sample’s repertoire. Shannon diversity was calculated as \(S=-Check back often for more exciting news!_For more tech updates, stay tuned to our blog.^Keep following us for the latest insights.For more tech updates, stay tuned to our blog._{i}\mathrm{ln}{(p}_{i})\), where p_i is the abundance of the ith TCR sequence. The Morisita–Horn index was calculated using the vegdist function with method = “horn” implemented in the vegan (v.2.6-4) package⁶⁷.

PreGame development

To develop PreGame, we considered multimodal features derived from both gene and surface protein expression assays as well as cell metadata features. Only cells with confirmed TR and NTR γδ TCRs were considered. Cells expressing γδ TCRs with a borderline increase in activation strength (13–15%; defined as above) were disregarded to limit the risk of false positives. After initial iterations, we settled on extraction of five categories of features: (1) expression values of a 19-gene signature; (2) surface protein expression of a 5-protein signature; (3) S phase score; (4) G2/M phase score; and (5) cell cycle stage (Extended Data Fig. 6a,b). The gene and protein signatures were identified as differentially expressed between TR and NTR γδ T cells. Cell-cycle-related signatures were computed using CellCycleScoring implemented in Seurat (v.4.3.0), and cell cycle phase was label-encoded to numerical features. The model was initially trained using the discovery cohort, which consisted of a 27-dimensional feature matrix containing 88 TR T cells (positive class) and 361 NRT T cells (negative class). To address this class imbalance, we applied the synthetic minority over-sampling technique (SMOTE)⁶⁸, which generated synthetic positive samples to balance the training split while avoiding data leakage. Given the presence of both continuous and categorical features in our dataset, we used the SMOTENC function from the imbalanced-learn package (v.0.11.0), which is specifically designed to handle mixed data types. After oversampling, the numbers of positive and negative cells were equal. We used tenfold cross-validation, wherein the data were split into ten subsets, with nine used for training and one for testing in each iteration. Preprocessing, normalization and feature selection were performed before training splits, whereas oversampling was applied only to training splits. This process was repeated ten times so that each cell was used at least once for testing. Predicted probabilities for each cell being TR were recorded and used to compute performance metrics. To further reduce potential sampling bias, the entire training and evaluation procedure was repeated 100 times. Model performance was assessed using the AUC, which guided the selection of optimal hyperparameters. Among several ML models evaluated, random forest models achieved the best performance and were selected for downstream analysis.

Differential expression analysis of TR γδ T cells

After in vitro screening of each γδ TCR from the discovery cohort was complete (Fig. 2b), TR and NTR labels were mapped back to the single-cell data. Cells expressing a γδ TCR with a borderline increase in absolute activation strength (13–15%, defined as above) were excluded to limit the risk of false positives. TR (n = 88) and NTR (n = 361) T cells were compared using FindMarkers implemented in Seurat (v.4.3.0) to identify significantly differentially expressed genes and surface proteins, which were used as features by PreGame (as described above). The application of PreGame to the validation 1 and 2 MM cohorts, and subsequent reactivity screening, substantially increased the number of TR and NTR γδ T cells that could be leveraged for differential expression analyses. The accuracy of PreGame resulted in a disproportionate addition of TR (n = 1,616) compared with NTR (n = 374) cells (Fig. 4a). However, this imbalance was driven by several highly expanded clonotypes (Fig. 4c), with the number of unique TR (n = 48) and NTR (n = 73) clonotypes being more comparable. To address this imbalance and to prevent highly expanded clonotypes from dominating the statistical power, we opted to apply an iterative downsampling approach. In brief, γδ TCR clonotypes were randomly downsampled to 30 cells (twice the median number of cells per screened clonotype), where applicable, and FindMarkers was again used to identify differentially expressed genes and surface proteins. Before downsampling, cells with a sequenced BCR (in addition to a γδ TCR), which probably represented rare doublets, were excluded. After 1,000 iterations, log₂ fold changes, P values, FDR-adjusted P values and expressed values were averaged, and results with an average FDR-adjusted P value of less than 0.05 were deemed significant (as shown in Fig. 4e). This downsampling method reduced the bias of highly expanded clonotypes, thereby giving a more faithful representation of TR γδ T cells. However, this method also prompted our interest in differences between expanded and low-frequency γδ T cell clonotypes. We used a threshold of 2 cells (median frequency of TR clonotypes) to define a clonotype as expanded (> 2 cells) or low frequency (1 or 2 cells). We then used FindMarkers to compare expanded and low-frequency cells separately for TR or non-reactive clonotypes (cell-level comparison). Genes or surface proteins were considered significantly differentially expressed if the FDR-adjusted P value was less than 0.05 from the cell-level comparison.

TCR target enrichment and sequencing for hybrid-capture TCR sequencing (CapTCR-seq)

gDNA (76–500 ng) from peripheral blood buffy coat, fractionated PBMCs or fractionated plasma was input for library construction and subjected to mechanical shearing using a Covaris LE220 to achieve an average fragment size of approximately 250 bp. Fragmented DNA was ligated to duplex unique molecular identifier (dUMI) adapters designed with dual-indexed 12-nucleotide barcodes and T-tailed overhangs (Integrated DNA Technologies). Libraries were PCR-amplified with KAPA HiFi master mix (4–-6 cycles) and enriched for TCR loci (TRA, TRB, TRG and TRD) using a custom xGen Lockdown probe set (IDT). Final pools were sequenced on an Illumina NovaSeq 6000 platform using 2 × 150 bp paired-end chemistry. Demultiplexing was performed using bcl2fastq (v.2.20), which resulted in paired FASTQ files for each sample.

Alignment of raw CapTCR-seq data

Raw CapTCR-seq data were aligned as previously described²⁷. In brief, demultiplexing was performed using bcl2fastq (v.2.20), which resulted in paired FASTQ files for each sample. Before alignment, we performed barcode extraction and header tagging to isolate and validate the duplex barcodes embedded in the adapter sequences. These steps removed the 12-nucleotide barcodes along with flanking spacer sequences from the 5′ end of both reads and inserted them into the FASTQ header in the format @|./1 or /2. Barcodes were validated against a whitelist of known barcode sequences to ensure compatibility with downstream error correction. Reads with mismatched spacers or invalid barcode sequences were excluded. TCR sequence alignment and analysis was performed using MiXCR (v.4.6.0)^69,70. The align step was run with the rna-seq preset and parameters optimized for partial alignments as follows: -OallowPartialAlignments = true and -OvParameters.geneFeatureToAlign = VGeneWithP. This step was followed by two rounds of partial assembly using assemblePartial, extension of assembled reads using extend and full clonotype assembly using assemble. Final alignments were subjected to quality control and excluded in instances when outlier high or low total sequencing reads or total aligned reads led to an alignment rate <50%. This quality-control step did not result in exclusion of any samples from Algonquin clinical trial patients.

Extraction of TCR clonotypes and repertoire diversity calculation

Final clonotypes for each TCR locus (TRA, TRB, TRG and TRD) were exported together using exportClones with the following filters to exclude non-productive clonotypes: –filter-out-of-frames, –filter-stops, –dont-split-files, –export-productive-clones-only. Total read fraction, or abundance, were calculated as the sum of read fractions of each unique CDR3 sequence from a given grouping of TCRs. TCR diversity metrics were calculated for each TCR locus and sample source (that is, plasma or PBMC) separately using postanalysis individual with the following parameters: -f, –default-downsampling count-read-auto, –default-weight-function read, –only-productive.

MM cell line bulk RNA sequencing acquisition and transcriptional clustering

Bulk RNA sequencing data from 66 MM cell lines were previously generated by and obtained from the Keats Laboratory (https://www.keatslab.org/data-repository). FPKM counts were obtained and genes with fewer than ten counts across all lines were filtered out. The remaining genes were used to construct a UMAP embedding using the R package umap (v.0.2.10.0) with n_neighbors = 6, min_dist = 0.5, and otherwise default parameters.

Patient survival analysis

PFS in the Algonquin clinical trial cohort was analysed using the Kaplan–Meier method. Numbers of patients at risk at specific months are displayed below the survival curves. For PFS comparisons between two groups, patients in the high and low groups were stratified by the median value throughout, unless otherwise indicated. Significance between groups was assessed using log-rank tests for comparisons involving the subset of patients with BM sequencing samples or using Gehan–Breslow–Wilcoxon tests for comparisons involving the entire cohort when many patients were still on-trial. Survival analysis was conducted in R (v.4.2.1) using the survival package (v.3.4-0), PHInfiniteEstimates (v.2.9.5) and survminer (v.0.5.0). Kaplan–Meier curves were drawn using the ggsurvplot function implemented in survminer.

Statistical analyses and visualizations

Statistical tests were calculated using R (v.4.2.1) or GraphPad Prism (v.10.6.1) unless otherwise indicated. Statistical tests were performed as indicated in the figure legends and results were considered significant if P < 0.05. Packages used for preparing data visualizations included tidyverse (v.2.0.0), pheatmap (v.1.0.12), ggplot2 (v.3.4.2), ggalluvial (v.0.12.5) and/or packcircles (v.0.3.7), or as otherwise indicated.

Reporting summary

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