MainDespite more than a decade of multidisciplinary research, limited remains from outside Denisova Cave have been molecularly identified as Denisovan; those that have been identified include a partial mandible and a rib from Baishiya Karst Cave (Xiahe) on the Tibetan Plateau2,3, a partial mandible (Penghu) from the Taiwan Strait4 and a nearly complete cranium (Harbin) from northeastern China5,6 (Fig. 1a). These findings necessitated a re-evaluation of Middle Pleistocene hominin systematics in Asia, which led to the proposal of new species: Homo longi (comprising Harbin, Dali and Jinniushan) and Homo juluensis (comprising Xujiayao, Xuchang, Siberian Denisovans, Xiahe, Penghu and Tam Ngu Hao 2)7. Recent phylogenetic analyses based on morphological characteristics also placed Yunxian 2, Hualongdong and Maba in the H. longi clade8, implying their potential affinity with Denisovans. However, most putative Denisovan fossils have been confined to northern and southeastern China, leaving a conspicuous gap in southwestern China. Given genomic evidence of substantial Denisovan ancestry in present-day populations in Southeast Asia and Oceania9, this southwestern region warrants in-depth investigation as a key region for Denisovan dispersal.Fig. 1: Geographic location of Bianfu Cave and morphology of two hominin teeth and three newly identified Denisovan bones.a, Map showing the geographic locations of Denisovan fossils determined by molecular evidence. The colours of the map reflect a combination of elevation and environmental context. The map was projected using the Albers equal-area conic projection. b, Western profile map of Trench 2 in Bianfu Cave, showing the stratigraphic layers containing Denisovan remains (indicated by red stars). c, Different views of the lower fourth premolar from layer 7 (YHB3518; from left to right: occlusal, buccal and distal views). d, Different views of the lower second molar from layer 7 (YHB3075; from left to right: occlusal, buccal and mesial views). e, Different views of the Bianfu Cave radius from layer 7 (BFD771, upper row) and its virtual reconstruction (bottom row); from left to right: anterior, posterior, medial, lateral and proximal views. The cross-section indicates the elliptical contour of the radial neck. f, Ectocranial and endocranial surfaces of the Bianfu Cave parietal from layer 7 (BFD767) and its virtual reconstruction to show the branches of the middle meningeal vessels (blue arrows) and vault thickness. g, Ectocranial and endocranial surfaces of the Bianfu Cave parietal from layer 9 (BFD769) and its virtual reconstruction to show the ridge along the superior sagittal sinus groove (blue arrow on top), lambdoidal suture (blue arrow in the middle) and vault thickness. Scale bars, 1 m (b), 1 cm (c,d), 1 cm (e–g). The base map in a uses cross-blended hypsometric tints from Natural Earth (available at https://www.naturalearthdata.com/downloads/10m-cross-blend-hypso/10m-cross-blended-hypso-with-shaded-relief-and-water).In addition, although the nearly complete Harbin cranium has provided the most comprehensive cranial morphology for Denisovans5,6, key anatomical aspects of Denisovan biology relevant to adaptive evolution remain poorly understood, including stature, body proportions, axial posture, locomotor behaviour and activity levels. These questions require discovery of diagnostic postcranial elements10,11. Biomechanical analyses of Neanderthal and early modern human long bones have revealed intra- and interpopulation skeletal variation across temporal and regional contexts12 and have also demonstrated evolutionary trends in mobility11, manipulative behaviours13, subsistence strategies14 and sex division of labour15. However, current Denisovan postcranial fossils are restricted to the Denisova 3 phalanx and the Xiahe 2 rib, severely limiting our understanding of their postcranial morphology and associated adaptive functional variation.Although researchers have attempted to reconstruct Denisovan skeletal morphology from methylation patterns in the high-coverage Denisova 3 genome16, some predicted traits have differed from those observed in the Denisova 3 phalanx17 and the Penghu and Xiahe 1 mandibles2,4. In addition, given the divergent Denisovan ancestry in modern populations9 and the known spatiotemporal distribution and amino acid variation among Denisovan individuals2,3,4,5, a broader fossil record is needed to define Denisovan morphological phenotype. Direct analysis of skeletal remains, especially long bones, is therefore essential for robust morphological and behavioural reconstruction. Zooarchaeology by mass spectrometry (ZooMS) is a critical tool for screening of hominin fossils from numerous morphologically unidentifiable bone fragments18. Its use at Denisova Cave19,20,21 and Baishiya Karst Cave3 has nearly doubled the number of known Denisovan specimens, thereby improving our understanding of the adaptive flexibility and evolutionary history of Denisovans and their interactions with other archaic hominins. Wider use of ZooMS could lead to the discovery of further Denisovan localities, offering insight into behaviour, subsistence strategies and cultural practices. The discovery of further postcranial remains with diagnostic features could also clarify morphological variations.Yunnan, situated at the southeastern edge of the Tibetan Plateau, serves as a crucial crossroads between East Asia, South Asia and Southeast Asia, where Denisovans are thought to have interbred with modern humans9,22. At least 15 hominin localities and 19 Palaeolithic sites with associated mammal fossils have been reported in this province, including Bianfu Cave, the earliest known Palaeolithic site in western Yunnan (Fig. 1a,b). Excavations have unearthed four hominin teeth (from layer 7), 1,366 stone tools and more than 64,000 mammalian fossils within cultural deposits dated to approximately 190–70 thousand years ago (ka)1, coinciding with the known chronological range of Denisovans in East Asia2,3,9,23. To understand hominin occupation of this site, we conducted proteomic analysis on two hominin teeth from layer 7, comprising a lower fourth premolar (YHB3518) and a lower second molar (YHB3075) (Fig. 1c,d); we also performed systematic morphological and ZooMS screening of mammalian fragments to identify other hominin bones, especially postcranial remains, to further elucidate their morphological phenotype.Further hominin bone identificationTo identify potential hominin fossils, we screened more than 60,000 bone fragments from Bianfu Cave. Even with a high-throughput ZooMS approach, analysis of all fragments would have been prohibitively costly. Therefore, we added a morphological prescreening step before the wet laboratory experiments. On the basis of manual inspection of bone size, cortical thickness and surface morphology, we classified 22 fragments as possible hominin remains, with 5 from the upper layers (layers 3–6, approximately 140–63 ka) and 17 from the lower layers (layers 7–12, approximately 195–134 ka)1 (Fig. 1b). Preliminary ZooMS analysis of randomly selected bones indicated that both upper and lower layers retained detectable collagens, with better preservation in the upper layers (24 of 36, 66.7%) than in the lower layers (3 of 18, 16.7%) (Extended Data Fig. 1 and Supplementary Data 1). ZooMS analysis of all 22 potential hominin fossils identified 5 upper-layer specimens as either Bovidae or Cervidae and 8 lower-layer specimens as Bovidae, Cervidae, Hyaenidae and Hominidae, consistent with the mammalian fauna at this site1 (Extended Data Fig. 1 and Supplementary Data 1). Notably, three fragments, comprising two parietal bone fragments (BFD767 from layer 7 and BFD769 from layer 9) and a proximal radial fragment (BFD771 from layer 7) (Fig. 1e–g), exhibited Homo-related collagen markers. An attempt to recover endogenous ancient DNA from these three hominin samples failed. The parietal from layer 7 (BFD767) was then selected for further protein extraction to obtain more data. Uranium-series (U-series) dating of three hominin bones and attached carbonates provided minimum ages of around 102 kyr for 2 samples from layer 7 and around 143 kyr for the parietal from layer 9 (BFD769) (Methods, Supplementary Text 1 and Supplementary Table 1). These ages were much younger than those of the corresponding layers, as established by a Bayesian age model integrating optically stimulated luminescence dating of sediment samples and U-series dating of speleothems, namely, between 148–139 kyr and 142–134 kyr for layer 7, and between 167–161 kyr and 160–150 kyr for layer 9 (ref. 1). Given that bones are open systems for U-series dating24, their ages may represent minimum ages, whereas the stratigraphic ages are more accurate.Shotgun proteomics was conducted to confirm the endogeneity and taxonomic identification of all three hominin bones (Supplementary Data 1). Searching against a custom-built mammalian type I collagen database attributed the strongest matches to the Homo genus. Deamidation calculations of these collagen peptides revealed elevated deamidation levels in the three putative hominin specimens comparable to or exceeding those of faunal bone fragments from the same layer, indicating endogenous protein preservation (Extended Data Fig. 2). A further database search against an expanded collagen database and further taxonomic classification using the ClassiCOL pipeline25 assigned the highest taxonomic match scores to the Homo–Pan clade. Considering the past biogeographical range of Pan26 and Middle Pleistocene faunal records from this site and surrounding regions1,27, all three specimens were attributed to the Homo genus.The endogenous proteomeTo obtain comprehensive protein profiles from the five Bianfu Cave hominin fossils, we used the proteomic data from bone or dentine fractions to search against a database that incorporated the modern human proteome and translated archaic hominin bone and dentine protein sequences, whereas the proteomic data from enamel fractions were used to search the custom-built Hominidae enamel database for each tooth (Methods). This identified 52,682 peptide-spectrum matches (PSMs) and 10,242 peptides for the parietal BFD767 (four fractions combined); 2,937 PSMs and 1,548 peptides for the parietal BFD769; 4,750 PSMs and 2,582 peptides for the radius BFD771; 26,465 PSMs and 4,869 peptides for the YHB3518 dentine (three fractions combined); 20,737 PSMs and 3,454 peptides for the YHB3518 enamel; 21,468 PSMs and 4,433 peptides for the YHB3075 dentine (three fractions combined); and 25,388 PSMs and 4,430 peptides for the YHB3075 enamel. Endogenous proteins were validated on the basis of elevated deamidation rates of glutamine (Q) and asparagine (N), resulting in 6, 6, 5, 13 and 16 endogenous proteins in the parietal BFD767, parietal BFD769, radius BFD771, tooth YHB3518 and tooth YHB3075, respectively (Extended Data Fig. 3, Supplementary Text 2, Supplementary Figs. 1–4 and Supplementary Table 2). All the endogenous proteins from bone or dentine fractions were collagens, consistent with previous findings that collagens tend to resist degradation for a longer period28. Typical enamel-specific proteomes were retrieved from two enamel fractions, and an abundance of AMELY-specific peptides indicated that both teeth belonged to male individuals.The tooth YHB3518 yielded 3,856 amino acid residues across 13 endogenous proteins, among which 62.0% and 23.6% were uniquely found in dentine and enamel fractions, respectively (Extended Data Fig. 4). For the tooth YHB3075, 4,076 residues were recovered, with 63.3% unique to dentine fractions and 36.7% unique to the acid etch enamel fraction (Supplementary Fig. 5). The parietal BFD767 yielded 3,666 amino acid residues across 6 collagens, of which 44.4% were consistently recovered across four fractions and 24.6% were uniquely found in a single fraction (Supplementary Fig. 6). This suggests that multiple fractions substantially increase amino acid coverage of endogenous proteins, yielding a proteome comparable with other high-quality Pleistocene hominin proteomes, such as those from the Penghu mandible (4,241 residues)4 and Xiahe 2 rib (4,597 residues)3. The parietal BFD769 and radius BFD771 yielded 1,972 and 2,422 residues, respectively, comparable with the Xiahe 1 proteome with 2,333 residues2, enabling population assignment and phylogenetic analysis.Species and population assignmentsTo determine the population assignments of the five Bianfu Cave specimens, we screened for amino acid substitutions within the recovered endogenous proteome (with a threshold of at least two peptides per position) unique to modern humans, Neanderthals, Denisovans, and the Pan, Gorilla and Pongo genera (Supplementary Data 2 and 3). For all recovered positions derived in non-Homo genera, the Bianfu Cave specimens retained the ancestral alleles with a 100% match rate: 20, 24, 9, 7 and 9 ancestral alleles for the tooth YHB3518, tooth YHB3075, parietal BFD767, parietal BFD769 and radius BFD771, respectively. In addition, the parietal BFD769 had one derived amino acid variant assigned to the Homo genus with a 100% match rate. These single-amino polymorphisms (SAPs) enabled us to confidently assign the five hominin specimens to the Homo genus, confirming the morphological identification of the two teeth and the preliminary species identification based on ZooMS spectra and collagen database searches for the three bone specimens (Supplementary Data 1).To investigate the relationship of the samples among hominin populations, we further evaluated all recovered SAPs within the Homo genus, retaining only those with elevated deamidation values and sufficiently reliable peptide identification across three software programs (Methods, Supplementary Text 3 and Supplementary Data 4 and 5). No Neanderthal- or modern-human-derived variants were confidently identified in any of the five Bianfu Cave specimens. By contrast, all five carried a Denisovan diagnostic marker in the collagen α-2(I) chain (COL1A2 R996K, rs1792270869), with 6, 39, 30, 2 and 7 supporting peptides for the tooth YHB3518, tooth YHB3075, parietal BFD767, parietal BFD769 and radius BFD771, respectively (Fig. 2, Table 1 and Extended Data Fig. 5). This variant is consistent across all identified Denisovan individuals, including Harbin5, Penghu4, Xiahe 1 (ref. 2), Xiahe 2 (refs. 3,29), Denisova 25 (ref. 30) and Denisova 3 (ref. 31), suggesting that the Bianfu Cave specimens also belong to Denisovans.Fig. 2: Details of the Denisovan diagnostic variant (COL1A2 R996K) identified from two Bianfu Cave samples.a,b, Amino acid coverage depths of COL1A2 identified from the radius BFD771 (a) and tooth YHB3518 (b). The Denisovan diagnostic variant at position 996 is indicated by a vertical solid line. c, Sequence alignments covering phylogenetically informative variant COL1A2 996K in the Bianfu Cave samples. Variant 996K is highlighted in red. Modifications of amino acids are indicated by lowercase letters: n, deamidation; h, His-to-Asp; p, hydroxylation.Table 1 Details of SAPs identified in the Bianfu Cave samples assigned within Homo sp.Full size tableIn addition to the diagnostic Denisovan marker (COL1A2 R996K), three other informative variants were identified from the enamel of two Bianfu Cave teeth, YHB3518 and YHB3075 (Table 1, Extended Data Figs. 6 and 7 and Supplementary Figs. 7 and 8), two from ameloblastin (AMBN) and one from amelogenin Y isoform (AMELY). Among these other variants, AMBN M273V (rs564905233) was also present in East Asian Homo erectus32, Harbin32, Penghu4, Denisova 3 (ref. 31), Denisova 25 (ref. 30) and some modern humans with high Denisovan introgression33, confirming the Denisovan affinity of the specimens. Another variant (AMELY M179L) was shared by Penghu4, Harbin32, Denisova 25 (ref. 30), Neanderthals34,35,36,37,38,39, Ust’-Ishim40 and all the modern humans in the gnomAD database, in contrast to AMELY-179M, which is present in Homo antecessor41 and great apes. Thus, the identification of AMELY M179L enabled us to attribute these two teeth to the Homo sapiens, Neanderthals and Denisovans clade. The last variant, AMBN-253A, was present in all the comparative samples except East Asian H. erectus32. Therefore, on the basis of all the informative variants, we attributed the Bianfu Cave specimens to Denisovans.To further determine the taxonomic affiliation of these samples, we constructed Bayesian phylogenetic trees from the concatenated consensus sequences of the corresponding endogenous proteins of each specimen. Uncertain amino acid positions were masked; these included positions with single-peptide depth and SAPs with indistinguishable post-translational modifications (PTMs) or those considered unreliable after manual inspection (Supplementary Data 6). We obtained 3,076 residues from the tooth YHB3518, 3,219 from tooth YHB3075, 3,005 from parietal BFD767, 1,423 from parietal BFD769 and 1,917 from radius BFD771; these values constituted 22.4%, 23.5%, 36.7%, 18.6% and 25.0% of the total protein alignments, respectively. The resulting phylogenetic trees showed a topology highly congruent with findings of nuclear genomic studies31,34,39, placing modern humans as a sister group to the Neanderthal–Denisovan clade (Fig. 3, Extended Data Fig. 8 and Supplementary Figs. 9–11). Each Bianfu Cave sample formed a monophyletic group with Denisova 3, with a posterior probability of 100%, consistent with the results of the SAP assignments and supporting a more confident attribution to Denisovans.Fig. 3: Phylogenetic placement of radius BFD771.A Bayesian phylogenetic tree was shown to determine the placement of the Bianfu Cave radius from layer 7 (BFD771), using Pongo abelii as an outgroup.Morphology of hominin bone fragmentsThe newly identified Denisovan fossils from Bianfu Cave included two small pieces of parietal bones (BFD767 from layer 7 and BFD769 from layer 9) and one partial radius (BFD771 from layer 7) (Fig. 1e–g). BFD767 was from the anterior portion of the right parietal’s lateral wall. It exhibited a weak curvature, relatively smooth external surface and two vascular impressions on the endocranial surface (identified as the two obelic branches of the middle meningeal vessels). BFD769 had a moderate curvature and represented a right parietal fragment next to the lambda, preserving a small portion of the ridge along the superior sagittal sinus groove and the lambdoidal suture. The parietals BFD767 and BFD769 measured 7.6 mm and 10.0 mm in thickness at the centre, respectively.Despite the variation in vault thickness in different areas of the parietal bone, the thickness of Bianfu Cave parietal fragments (7.6–10.0 mm) was comparable with those of Zhoukoudian H. erectus (5–16 mm) and some of the East Asian late Middle Pleistocene archaic Homo from Maba (7 mm), Xuchang (7.9 mm), Dali (11.2 mm) and Xujiayao (7.0–12.6 mm); they were thicker than that of Jinniushan (6 mm), close to the upper limits of Homo heidelbergensis (6.5–11.5 mm) and fossil H. sapiens (3.5–12.0 mm), and roughly the median value of Neanderthals (5–17 mm) (Supplementary Table 3). In terms of the cranial vault thickness pattern, the parietal BFD767 resembled some specimens of H. heidelbergensis, Neanderthals and East Asian late Middle Pleistocene archaic Homo (Xujiayao PA1485 and Maba), being thicker than that of H. sapiens but thinner than that of Asian late H. erectus (Extended Data Fig. 9). The parietal BFD769 was very thick, similar to that of Asian late H. erectus, H. heidelbergensis and East Asian late Middle Pleistocene archaic Homo (Dali, Xujiayao PA1490) and thicker than that of Neanderthals and H. sapiens (Extended Data Fig. 9; see also figure S4 in ref. 42). On the basis of both parietal fragments, considered together, the cranial vault thickness pattern of Bianfu Cave hominins is most like that of H. heidelbergensis and East Asian late Middle Pleistocene archaic Homo.BFD771 is the only Denisovan radius identified so far. The robust cortical structure and fully fused proximal epiphysis, assessed through micro-computed tomography (micro-CT) analysis of the internal morphology, indicate that the individual was probably an adult or near-adult43. The specimen preserves approximately 45 mm of the proximal portion, extending from a partially intact radial head to the distal part of the markedly elevated radial tuberosity. On the basis of the marginal morphology of the radial tuberosity and the major axis orientation of the radial mid-neck cross-section, BFD771 is most likely to be a right radius with a medially oriented radial tuberosity (Extended Data Fig. 10 and Supplementary Figs. 12 and 13). A medial orientation of the radial tuberosity has been observed in 76.5% of late Neanderthals, as well as in Australopithecus afarensis (AL 288-1p) and Homo habilis (OH 62). By contrast, an anteromedially oriented tuberosity is the predominant condition in H. antecessor, early Neanderthals from Sima de los Huesos (80%) and H. sapiens (88.8%)44,45,46.The radial head dimensions were estimated using best-fit circles corresponding to the preserved head margin (23.2 mm) and fovea articular edge (17.2 mm) (Supplementary Fig. 14a). These values indicated a large radial head, with greater metric similarity to early and late Neanderthals than to H. sapiens (Supplementary Fig. 15a). The fisher neck length (22.5 mm; Supplementary Fig. 14b) fell within the range of late Neanderthals and the upper portion of H. sapiens and was comparatively shorter than those observed in early Neanderthals (Supplementary Fig. 15b). According to cross-sectional geometric (CSG) analysis of the radial mid-neck, BFD771 had a greater total area and torsional rigidity than Neanderthals, Homo. cf. erectus, A. afarensis, Australopithecus sediba and Paranthropus boisei, falling within the variation of both modern humans and extant great apes (Supplementary Fig. 16a,b). The cross-sectional shape, evaluated by the bending rigidity ratio, remained within the range of all comparative groups (Supplementary Fig. 16c). Diffeomorphic surface matching analysis, which quantitatively evaluates overall external shape independent of size and orientation, aligned BFD771 more closely with modern humans and H. cf. erectus owing to its relatively short neck in proportion to the radial head. This configuration distinguished it from the relatively elongated necks characteristic of Neanderthals, A. afarensis, A. sediba and P. boisei, as these taxa exhibit greater morphological consistency with extant great apes (Fig. 4 and Supplementary Fig. 17).Fig. 4: Diffeomorphic surface matching analysis of the Bianfu Cave radius.The proximal radius fossil from Bianfu Cave (BFD771) was processed with three restoration protocols designed to capture a wider range of morphological variation and optimize restoration reliability. These protocols were based on the reference template representing the mean shape of all comparative specimens, modern humans and Neanderthals, respectively (Supplementary Fig. 17). We performed principal component analysis on velocity vectors derived from diffeomorphic surface matching of the proximal radii. PC1 primarily captured the relative length of the radial neck in proportion to the radial head, whereas PC2 primarily captured the radial head size relative to the transverse dimensions of the proximal diaphysis. Sample sizes for comparative taxa are indicated in parentheses.DiscussionIdentification of further remains is crucial for understanding the spatiotemporal distribution, evolutionary history and morphological phenotype of Denisovans. Here we performed systematic morphological prescreening of more than 60,000 bone fragments from Bianfu Cave and identified 22 candidate hominin specimens. Subsequent ZooMS and proteomic analyses confirmed three to be Denisovan remains; these comprised two parietal bone fragments (BFD767 and BFD769) and one proximal radial fragment (BFD771). Combined with four hominin teeth, two (YHB3518 and YHB3075) of which were identified as Denisovan by proteomic analysis, these findings establish Bianfu Cave among the known Denisovan localities, second only to Denisova Cave in Siberia with respect to its abundance of remains. This site lies along an important migration corridor linking the Qinghai-Tibet Plateau with East Asia, South Asia and Southeast Asia; it thus fills a critical geographic gap in the known distribution of Denisovans and reveals their presence across a vast range from Siberia23 and northeastern China5,6 in the north, through the Qinghai-Tibet Plateau2,3 in the centre, to the Taiwan Strait4 in the southeast, and now the Yunnan-Guizhou Plateau in the southwest.In addition to the six specimens from layer 7, we identified a Denisovan parietal fragment from layer 9, a layer from which no hominin fossils were identified during excavation. These findings extend the Denisovan occupation to approximately 167–134 ka within marine isotope stage 6 (ref. 1). During this glacial period, the relatively warm climate and conifer-dominated forest or forest-steppe at Bianfu Cave and surrounding regions may have provided a stable supply of mammal resources, especially a predominance of medium- to large-sized herbivores1. This may have enabled the Yunnan-Guizhou Plateau to serve as a hunting ‘paradise’ for hominins. Previous studies have inferred introgression of multiple genomically distinct Denisovan populations into modern humans across Asia and Oceania9. However, the genomic and proteomic data available at present for Denisovans are too scarce to enable us to characterize the Bianfu Cave population and its possible lineages accurately; thus, we require further ancient DNA evidence, especially from the transition zone between East Asia and Southeast Asia (such as Yunnan in this study), to determine their contribution to modern human genomes.Our study expands the sparse fossil record of Denisovans and provides important fossils to elucidate the morphological phenotype of Denisovans. The identification of a Denisovan radial fragment is of particular importance for improving our understanding of the postcranial phenotype of this population. Although the Denisovan proximal radius shared large proximal dimensions and a probable medially oriented radial tuberosity with Neanderthals, its overall external shape and mid-neck cross-sectional geometry were aligned more closely with those of modern humans. Radial tuberosity orientation, neck length, head dimensions and neck cross-sectional geometry are critical biomechanical features related to musculus biceps brachii (M. biceps brachii) mechanical efficiency and proximal radial load-bearing capacity43,45. The mosaic pattern of these traits in the Bianfu Cave radius, relative to Neanderthals and modern humans, indicates differential biomechanical demands and behavioural adaptations in Denisovans. We hope that similar proteomic analyses will aid in the identification of further Denisovan long bones in the future, allowing incorporation of multidisciplinary data (including morphological, molecular, chronological, ecological and subsistence-strategy data) into work to further understand the adaptive behaviour of Denisovans in response to climate change or competitive pressure from contemporary hominins.MethodsArchaeological site and samplesBianfu Cave (26° 27′ 43.65′′ N, 100° 09′ 55.45′′ E) is in Heqing County, western Yunnan Province in southwestern China, approximately 2 km south of the Yinhe River, a secondary tributary of the Yangtze River (Fig. 1a). In 2019, the Yunnan Institute of Cultural Relics and Archaeology and other cooperative institutes conducted excavations on this site. The stratigraphy comprises Holocene deposit, artefact-bearing cultural deposit and underlying cave-fissure deposit. The cultural deposit is further divided into upper (layers 3–6) and lower (layers 7–12) cultural layers (Fig. 1b). A Bayesian age model combining optically stimulated luminescence and U-series dates yielded a time range spanning approximately 190 ka to 70 ka1, making it the earliest Palaeolithic site discovered in western Yunnan. All cultural layers yielded lithic artefacts, totalling 1,366 stone artefacts, including both core and flake tools1. More than 64,000 mammal fossils were recovered from the cultural layers, among which more than 60,000 were fragmentary bones. On the basis of the limited number of morphologically identifiable mammal fossils, more than 100 individuals of deer and bovids were discovered, together with small numbers of rhinoceros, hyenas and other animals1.Four hominin teeth were unearthed in the middle of layer 7 (between 148–139 ka and 142–134 ka)1, suggesting a high likelihood of finding more hominin fossils in this site. The four hominin teeth included a lower left lateral incisor (I2), a lower left fourth premolar (P4) and two lower right second molars (M2). Morphological analysis showed that the lower M2 teeth exhibited close morphological affinity with those of Xiahe and Laos Denisovans, as well as the East Asian late Middle Pleistocene archaic Homo from Hualongdong1. Two of these teeth, a lower fourth premolar (YHB3518) and a lower second molar (YHB3075) (Fig. 1c,d), were selected for proteomic analysis to further confirm their taxonomic attribution. An acid etch approach was used for analysis of the enamel, whereas dentine powder was drilled for protein extraction, resulting in one enamel fraction and three dentine fractions for liquid chromatography coupled with tandem mass spectrometry (LC–MS/MS) analysis.After manual inspection and traditional morphological measurement of more than 60,000 bone fragments on the basis of bone size, cortical thickness, articular surface morphology, muscle attachment marks on long bone surfaces, vascular impressions on cranial internal surfaces and so on, 22 fragments were considered as potential hominin remains, five from the upper layers and 17 from the lower layers. Before conducting protein extraction on these 22 remains, we randomly selected 36 bone fragments from the upper layers and 18 from the lower layers for preliminary ZooMS analysis to assess protein preservation of the site. Thus, a total of 76 bone fragments were analysed by ZooMS in 6 batches, with each batch including an extraction blank and a positive control (either modern camel bone or dentine powder) (Supplementary Data 1). After ZooMS identification, 3 hominin remains and 14 mammalian samples from different layers were submitted to LC–MS/MS analysis, including the corresponding extraction blanks. To obtain a more comprehensive proteome of the hominin remains at this site, we performed further proteomic extraction on the parietal BFD767, yielding another three fractions for LC–MS/MS analysis. We also performed U-series dating to estimate the minimum age of three hominin bone samples.U-series datingU-series dating was conducted on three hominin bone samples and their attached carbonates. For each sample, a series of 5–11 spot pairs were analysed using a laser ablation system coupled to a multicollector inductively coupled plasma mass spectrometer (LA–MC-ICPMS). The LA system (RESOlution-LR, Applied Spectra) was equipped with a Coherent COMPex Pro102 ArF Excimer laser source and an S155 large-format sample pool. The Neptune double-focusing MC-ICPMS (Thermo Fisher) had nine Faraday cups and a secondary electron multiplier, allowing simultaneous measurement of several isotopes over a relative mass range of 17%. The Neptune interface was upgraded with a Jet-sample cone, an X-skimmer cone and a high-efficiency dry pump to improve the ion transport efficiency. By tuning the analytical parameters (carrier gas flow, torch position and zoom optics) of our LA–MC-ICPMS, we could routinely obtain 238U signals of approximately 0.8–1.0 V and 232Th/238U ratios of approximately 90–95% on the international standard NIST 612, with a spot size of 173 μm, a pulse repetition of 10 Hz, an energy fluence of 5 J cm−2 and a scan speed of 3 μm s−1. We conducted a preablation step to remove impurities from the sample surface. Then, each spot was ablated for 120 s at 10 Hz pulse repetition and 5 J cm−2 fluence. With signal intensities of 238U and 230Th, the spot size ranged from 50 μm to 180 μm. These laser ablations produced a crater approximately 150 μm deep. We measured isotopes 230Th and 234U using a secondary electron multiplier and simultaneously measured 232Th, 235U and 238U in Faraday cups47,48,49,50. For each spot, around 150 cycles were performed to acquire isotopic data, with the first 40–50 cycles collected before sample ablation and used as the gas blank. An in-house fossil bone standard (RM-B1) and a carbonate standard (RM-C1) were used as the matrix-matched standards for calibration of laser-induced elemental and isotopic fractionation for the hominin bones and carbonates, respectively.ZooMS screeningAround 100–200 mg of the sample was used for protein extraction with the conventional acid-insoluble approach51. After demineralization in 0.6 M HCl, the pellet was incubated with 50 mM ammonium bicarbonate (ABC) at 65 °C for 3 h. The supernatant was then collected and digested with trypsin. The digested peptide mixture was desalted using Pierce C18 Pipette Tips (Thermo Scientific), and then subjected to matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) analysis.Palaeoproteomic extractionFor the identified hominin parietal BFD767, a second round of protein extraction was conducted to obtain more proteins. The protocol was modified from that reported in a previous publication52. Another subsample (around 300 mg) was demineralized in 0.6 M HCl, and the acid supernatant was collected and ultrafiltered through an Amicon Ultra-4 (3 kDa) centrifugal filter unit. After the filtration unit had been washed with buffer (50 mM ABC), we obtained the ‘acid-soluble fraction 3KD’ by dissolving the proteins retained on the filter in 50 mM ABC. The pellet was further incubated in 50 mM ABC at 65 °C for 3 h. The supernatant was reduced with Tris(2-carboxyethyl) phosphine HCl (final concentration 0.1 M) at 56 °C for 30 min and alkylated using iodoacetamide (final concentration 0.1 M) at room temperature in the dark for 30 min. The supernatant was divided into two aliquots, one for trypsin digestion and the other for elastase digestion. The trypsin aliquot was ultrafiltered through an Amicon Ultra-4 (3 kDa) centrifugal filter unit, washed with buffer (ABC) and dissolved in 50 mM ABC; this was referred to as the ‘acid-insoluble fraction trypsin’. The elastase fraction was ultrafiltered through an Amicon Ultra-4 (3 kDa) centrifugal filter unit, washed with buffer (Tris-HCl solution) and dissolved in 50 mM Tris-HCl; this fraction was referred to as the ‘acid-insoluble fraction elastase’.The ‘acid-soluble fraction 3KD’ and ‘acid-insoluble fraction trypsin’ were digested with porcine trypsin (Promega, 2 μl, 0.5 μg μl−1) overnight at 37 °C, after which trifluoroacetic acid (TFA) was added to the digestion (final concentration 0.1%) to stop the reaction. The ‘acid-soluble fraction elastase’ was digested with elastase (Promega, 6 μl, 0.2 μg μl−1) overnight at 37 °C; then, TFA was added to the digestion (final concentration 0.5%) to stop the reaction. After a desalting step using Pierce C18 Pipette Tips, these peptide mixtures were dried for LC–MS/MS analysis.For two hominin teeth, YHB3518 and YHB3075, enamel proteins were extracted using an acid etch method modified from that described in ref. 53. After the initial 2-min etch using 5% (v/v) HCl, the solution was discarded. Then, two rounds of 15-min etch were carried out, and the solutions were combined to form the ‘acid etch fraction’. No reduction, alkylation or digestion was performed on this fraction. For dentine protein extraction, around 30 mg of powder (35 mg for YHB3518; 32 mg for YHB3075) was demineralized in 0.6 M HCl. The supernatant was collected and ultrafiltered through an Amicon Ultra-4 (3 kDa) centrifugal filter unit. After washing, the proteins were dissolved in 50 mM ABC and formed the ‘acid-soluble fraction 3KD’. The acid-insoluble pellet was washed and incubated with 50 mM ABC at 65°C for 3 h. The supernatant was collected as the ‘acid-insoluble fraction trypsin 1’. The remaining pellet was washed and incubated again. The supernatant with pellet was denoted the ‘acid-insoluble fraction trypsin 2’. No reduction or alkylation was conducted, and all three fractions were digested by trypsin. After desalting with Pierce C18 Pipette Tips, these peptide mixtures were analysed by MALDI-TOF MS for assessment of the collagen signal. As the ‘acid-soluble fraction 3KD’ of YHB3075 did not show any obvious collagen signal, it was combined with the ‘acid-insoluble fraction trypsin 1’ for further LC–MS/MS analysis. The other fractions were dried separately for LC–MS/MS analysis.A blank was extracted along with the samples for estimation of modern contamination during the experiment. All the sample treatments were performed in the dedicated clean room for ancient protein analysis in the Molecular Paleontology Laboratory, Institute of Vertebrate Paleontology and Paleoanthropology (IVPP), Chinese Academy of Sciences (CAS).MALDI-TOF MSOne microlitre of the peptide mixture was spotted on to an MTP384 Bruker ground-steel MALDI target plate. Then, 1 µl of α-cyano-4-hydroxycinnamic acid matrix solution (1% in 50% acetonitrile/0.1% TFA (v/v/v)) was added on top and mixed with the peptide mixture. Each sample was analysed on a Bruker autoflex maX MALDI-TOF mass spectrometer in triplicate. The acquired mass ranges were slightly different across several batches and included m/z 600 to 3,500, m/z 600 to 4,000 and m/z 700 to 3,500. Text files were converted from the raw files and processed using mMass v.5.5.0 (ref. 54).Liquid chromatography coupled with tandem mass spectrometryFor protein extraction from the three identified hominin bone samples (including the initial ZooMS extraction and the further protein extraction on the parietal BFD767), we performed LC–MS/MS analysis using two instruments from Central Laboratory at Capital Medical University: an Orbitrap Fusion Lumos mass spectrometer and an Orbitrap Exploris 480 mass spectrometer. Both machines were interfaced with an EASY-nLC 1200 system (Thermo Fisher Scientific). The peptides were loaded on to a trap column (100 μm i.d. × 2 cm, C18), followed by separation on an analytical column (150 μm i.d. × 25 cm, C18). Mobile phase A was 0.1% formic acid in water, and mobile phase B consisted of 80% acetonitrile and 0.1% formic acid. For Orbitrap Fusion Lumos, the peptides were eluted at a flow rate of 500 nl min−1 using a 120-min linear gradient program: 0–8 min, 7–11% B; 8–96 min, 11–28% B; 96–108 min, 28–40% B; 108–113 min, 40–90% B; 113–120 min, 90% B. Full MS data were acquired across an m/z range of 375–1,400 with a resolution of 120,000, an AGC target of 250% and a maximum injection time of 50 ms. The MS/MS scan was conducted at a resolution of 15,000 with an AGC target of 100%, a maximum injection time of 22 ms and a normalized collision energy of 35%. For the Orbitrap Exploris 480, peptides were eluted at a flow rate of 350 nl min−1 using a 125-min linear gradient program: 0–1 min, 1–9% B; 1–101 min, 9–28% B; 101–113 min, 28–42% B; 113–117 min, 42–99% B; 117–125 min, 99% B. Full MS data were acquired across an m/z range of 350–1,500 with a resolution of 60,000, an AGC target of 300% and a maximum injection time of 50 ms. The MS/MS scan was collected at a resolution of 15,000 with an AGC target of 75%, a maximum injection time of 22 ms and a normalized collision energy of 30%. Extraction blanks from different sample batches were analysed alongside the samples. Injection blanks were included before and after each sample analysis to minimize possible carryover between runs.For the protein extraction from two hominin teeth, we performed LC–MS/MS analyses using two instruments: an Orbitrap Fusion Lumos mass spectrometer (Thermo Fisher Scientific) from Central Laboratory at Capital Medical University; and an Orbitrap Exploris 480 mass spectrometer (Thermo Fisher Scientific) from State Key Laboratory of Genetics and Development of Complex Phenotypes at Fudan University. Both devices were coupled to an EASY-nLC 1200 HPLC system (Thermo Fisher Scientific). For the Orbitrap Fusion Lumos at Capital Medical University, the gradient program and mass spectrometry parameters were the same as those used for the hominin bone sample. For the Orbitrap Exploris 480 at Fudan University, the peptides were loaded on a 75 μm i.d. × 25 cm analytical column, which was packed in-house using reversed-phase silica of 1.9 μm (Reprosil-Pur C18 AQ, Dr. Maisch GmbH). Mobile phases A and B were as described previously. The peptides were separated using an 80-min gradient: 5–8% B, 2 min, at a flow rate of 200 nl min−1; 8–44% B, 38 min, 200 nl min−1; 44–70% B, 8 min, 200 nl min−1; 70–100% B, 2 min, 200 nl min−1; 100% B, 10 min, 200 nl min−1; 100–5% B, 2 min, 200 nl min−1; 5% B, 2 min, 300 nl min−1; 5–100% B, 6 min, 300 nl min−1; 100% B, 10 min, 300 nl min−1. Full MS scans were collected during the first 65 min, after which the column was washed and re-equilibrated for 15 min without data collection. The full MS data acquisition was conducted across an m/z range of 350–1,600 with a resolution of 60,000, an AGC target of ‘Standard’ and the maximum injection time mode set to ‘Auto’. The MS/MS spectra were acquired with a resolution of 15,000, an AGC target of ‘Standard’, a maximum injection time of 30 ms and a normalized collision energy of 30%.To provide a comparative reference for the extent of fossil degradation across different layers of Bianfu Cave, we also selected 14 animal samples for further LC–MS/MS analysis. These samples were processed on an Orbitrap Fusion Lumos mass spectrometer from State Key Laboratory of Genetics and Development of Complex Phenotypes at Fudan University, coupled to an EASY-nLC 1200 HPLC system (Thermo Scientific). Mobile phases A and B were as described previously. The peptides were separated on an analytical column (75 μm i.d. × 20 cm, C18) using an 80 min gradient: 2–5% B, 3 min, at a flow rate of 200 nl min−1; 5–35% B, 40 min, 200 nl min−1; 35–44% B, 5 min, 200 nl min−1; 44–100% B, 2 min, 200 nl min−1; 100% B, 10 min, 200 nl min−1; 100–5% B, 2 min, 200 nl min−1; 5% B, 2 min, 300 nl min−1; 5–100% B, 6 min, 300 nl min−1; 100% B, 10 min, 300 nl min−1. The full MS data acquisition was conducted over the first 65 min, with no data collected over the subsequent 15 min during column wash and re-equilibration. Full scans ranging from m/z 350 to m/z 1,600 were measured, with a resolution of 60,000, an AGC target of 100% and a maximum injection time of 50 ms. The MS/MS spectra were acquired at a resolution of 15,000 with an AGC target of 100%, a maximum injection time of 30 ms and a normalized collision energy of 30%.Proteomic data analysisTo confirm the taxonomic identification of these bone fragments, we conducted two analyses on the MS/MS data file from each LC–MS/MS run (not including the second round of protein extraction for the hominin parietal BFD767). First, the raw data file was searched against a custom mammal type I collagen database through PEAKS Online v.12 or v.11.5 (ref. 55). This database comprised two predominant bone proteins (COL1A1 and COL1A2), the sequences of which were collected from GenBank, UniProt and previous publications56,57,58,59,60,61,62,63,64,65,66. The translated protein sequences from high-coverage genomes of Denisova 3 (ref. 31), two Neanderthals (Altai Neanderthal, Vindija19)34,39 and Ust’-Ishim40 were also included. PEAKS searches, including peptide de novo, PEAKS DB, PTM and SPIDER, were performed with the following parameters: parent ion mass tolerance of 10 ppm, fragment ion mass tolerance of 0.05 Da, enzyme specificity set to SemiTryptic, maximal missed cleavages of 2, maximal modification number of 6 for each peptide, and variable modifications including deamidation (NQ), oxidation (M), hydroxylation (P), pyro-Glu from E and pyro-Glu from Q. The peptide length range was set to 8–45. The false discovery rate (FDR) was set to 1% (peptide level), and protein scores were filtered with −10log10P ≥ 20 and average local confidence (%) ≥ 50 (de novo only). A common contaminant database was also included in each database search, as in a previous study5, and we analysed the corresponding extraction blanks using the same strategy as for the samples to monitor possible modern contaminants during the experiments. Peptides detected in the negative control sample or assigned to contaminant proteins were excluded. The remaining peptides from each sample were used to calculate the extent of glutamine and asparagine deamidation using a modified Python script based on peptide intensity67.To validate the taxonomic identification results, we used another set of pipelines based on Mascot search and the ClassiCOL pipeline25. Raw data files from the bone samples were searched against a published collagen database containing 45 collagen homologues from 221 mammalian species25 using Mascot v.2.6 (Matrix Science). The parameters were set as follows: parent ion mass tolerance of 10 ppm, fragment ion mass tolerance of 0.05 Da, enzyme specificity SemiTryptic, maximal missed cleavages of 2, and variable modifications including deamidation (NQ), oxidation (M) and hydroxylation (P). After being filtered with a significance threshold of 0.01, the results were submitted to the ClassiCOL pipeline with the taxonomy restricted to Mammalia (python ClassiCOL.py -d {path_to_the_script} -l {path_to_search_results} -s MASCOT -t Mammalia -c 20).Following taxonomic identification, we performed a database search against the Hominidae bone database using raw data files from all three identified hominin bone samples to retrieve their proteomes. This database comprised the entire human proteome (from UniProt, one protein sequence per gene; 20,575 entries), supplemented with common bone and dentine protein sequences translated from 4 ancient hominin individuals (Denisova 3, Altai Neanderthal, Vindija19 Neanderthal and Ust’-Ishim)31,34,39,40. Published protein sequences for several other archaic individuals2,3,4,5,41 were also included in the database. The main search was conducted in PEAKS Online55 using the same parameters as for the search against the mammal type I collagen database with the following minor modifications. For the raw data files from the second round of protein extraction for the hominin parietal BFD767, carbamidomethylation (C) was added as a variable modification in the ‘acid-insoluble fraction trypsin’ and ‘acid-insoluble fraction elastase’, and enzyme specificity was set to ‘unspecific’ for the ‘acid-insoluble fraction elastase’. The possible mutations identified in the SPIDER search were added to the original database for iterative data search and verification. Peptides matching multiple genes were considered to be non-unique and excluded from subsequent analysis. Given the high similarity between different collagen chains, the permutation and Unipept confirmation filtering approach was used to validate the identification of low-abundance collagens (the pept2prot function from Unipept was used; Supplementary Text 2 and Supplementary Fig. 18). After this filtering and validation, the extent of glutamine and asparagine deamidation for each identified protein was calculated from peptide intensities using a Python script modified from a previous study67. For a reliable calculation, each protein was required to have at least two Q/N-containing peptides with valid intensities. Endogenous proteins were confirmed on the basis of elevated deamidation values.For the raw data files from three dentine fractions of two hominin teeth, a database search against the Hominidae bone database was carried out using the same parameters as those used for the search against the mammal type I collagen database for bone samples. The data processing strategy was the same as that used for the hominin bone samples. For the raw data files from the acid etch fraction, an initial search was conducted against the entire human proteome. The parameters were set as follows: precursor ion mass tolerance of 10 ppm; fragment ion mass tolerance of 0.05 Da; enzyme specificity of ‘unspecific’; variable modifications including deamidation (NQ), oxidation (M), hydroxylation (P), phosphorylation (STY), pyro-Glu from E and pyro-Glu from Q; maximum modifications per peptide of 3; and peptide length range of 6–45. FDR was set to 1% at the spectral level, and protein scores were filtered with −10log10P ≥ 20 and average local confidence (%) ≥ 50 (de novo only). In addition to the main enamel proteins, dentine collagens were identified in YHB3518 with relatively abundant peptides and elevated deamidation rates, probably because the acid etch process accessed the dentine region of this tooth and extracted related proteins. Two serine protease inhibitors, SERPINC1 and SERPINA1, were identified in YHB3075; these proteins exhibited elevated deamidation rates. This was not unexpected, as these proteins have also been reported in archaic and modern enamel32,68,69. Therefore, for each tooth, a specific Hominidae enamel database was built for a second round of search. In addition to the 12 commonly used proteins (AHSG, ALB, AMBN, AMELX, AMELY, AMTN, COL17A1, ENAM, KLK4, MMP20, ODAM and TUFT1), three main dentine collagens (COL1A1, COL1A2 and COL2A1) were included in the database for YHB3518, whereas two serine protease inhibitors (SERPINC1 and SERPINA1) were included in the database for YHB3075. Only species belonging to Hominidae were included; these comprised archaic hominins such as Denisovans and Neanderthals. The protein sequences, including both canonical and isoform sequences, were obtained from UniProt and the translated ‘Hominid Palaeoproteomic Reference Dataset’ available at Zenodo70. In addition, previously published proteomes from extinct archaic hominins (H. antecessor and H. erectus)32,41, Paranthropus71 and Gigantopithecus72 were incorporated. The search parameters were the same as those used in the initial search. Peptides detected in the negative control sample or assigned to contaminant proteins were excluded. The remaining peptides from each acid etch fraction were then reported and used to calculate deamidation profiles and validate their endogeneity67 (Supplementary Text 2 and Supplementary Fig. 19).Phylogenetically informative SAPs validationGenus- and population-specific variants were screened using a previously established pipeline and database, excluding variants that had single-peptide support or were indistinguishable owing to PTM5. Variants within the Homo genus that met initial thresholds (PSM count greater than or equal to 2, PSM ratio greater than or equal to 10%, and intensity ratio greater than or equal to 10%) were further validated using the strict filter criteria described in previous publications to confirm their reliability4,5,71. The peptides and PSMs supporting these sites were manually checked using the following criteria.