Jared Edward Reser, Ph.D.
October 21, 2026
Gigantopithecus blacki presents a demanding case for computational reconstruction because its documented anatomy is concentrated in jaws and teeth. I propose using this evidence to investigate two linked problems: the distribution of skeletal forms compatible with the fossils, and the genomic features that might be inferred from morphology together with comparative and molecular evidence. Fossil2Skeleton would learn relationships between preserved fragments and complete skeletal anatomy from associated specimens, then generate alternative reconstructions with measurable uncertainty. Skeleton2DNA would investigate the inverse relationship between anatomy and genomic variation, incorporating ancient enamel proteins as an independent source of molecular constraints. Two further proposals, Endocast2Brain and Skull2Face, clarify the evidential requirements of the broader program: neither has the necessary preserved cranial input for direct application to Gigantopithecus. I describe numerical representations, staged training, uncertainty propagation, and tests on withheld living species. The central empirical question is whether the surviving morphology improves predictions beyond those supplied by body size, evolutionary relationships, and comparative averages. This article presents a research proposal, with Gigantopithecus as a concrete test case for learning how much biological information can be recovered from fragmentary remains.
Years ago, I bought my graduate advisor an osteological reconstruction of Gigantopithecus. An object of that kind makes an extinct animal easier to imagine, while raising a scientific question about the relationship between the surviving specimen and the anatomy supplied by its reconstructors. Which features follow closely from the fossil evidence? Which depend on comparisons with living animals? Which remain largely open?
Those questions motivate the present proposal. A fossil jaw preserves its own structure, but it may also carry statistical information about other parts of the animal. If particular combinations of jaw dimensions and dental features repeatedly accompany particular body proportions in comparative specimens, a model can learn those relationships. The important experiment is to measure how well they predict anatomy that the model has never seen, and then determine how far the relationships can reasonably be extended to an extinct species.
I call this proposed system Fossil2Skeleton. Its purpose is to estimate missing skeletal anatomy from whatever fragments survive, while retaining alternative solutions and their uncertainties (Reser, 2026). Gigantopithecus supplies a particularly clear starting problem: jaw and tooth evidence is available, whereas most of the body is missing. A second proposal, Skeleton2DNA, asks whether morphological evidence can constrain genomic hypotheses when combined with living genomes and evolutionary information (Reser, 2025). The two systems address distinct inferential steps and require separate validation.
What the fossils and proteins establish
The major 2024 investigation of Gigantopithecus described a fossil record consisting of four mandibles and almost 2,000 isolated teeth, without postcranial remains. That study examined 22 caves in southern China and placed the extinction window at approximately 295,000 to 215,000 years ago. Its environmental and dental analyses associated the disappearance with increasing seasonality, changing vegetation, and difficulty adapting to altered food availability (Zhang et al., 2024). These findings establish a substantial research record despite the restricted anatomical sample.
Molecular evidence has also survived. Welker et al. (2019) recovered enamel protein sequences from a molar approximately 1.9 million years old. Their analysis placed Gigantopithecus as a sister lineage to orangutans, with an estimated common ancestor about 10 to 12 million years ago. This supplies information about evolutionary relationships that is independent of simply judging the appearance of the jaw. It does not provide an intact genome or establish that the animal’s body proportions matched those of a living orangutan.
Dental geometry already supports quantitative functional analysis. Yi et al. (2023), for example, used virtual tooth sections and measurements of enamel and dentine geometry to estimate crown strength and bite force. Such work illustrates how a preserved structure can constrain a biological property through an explicit model. Extending inference from the teeth to distant skeletal regions demands additional evidence that the relevant relationships remain predictive.
The reconstruction program should therefore begin with an inventory of what has actually been measured. Large teeth are observations. A proposed femur length is a prediction. Estimates of total body mass require further assumptions about the relationship between bones and soft tissues. These distinctions should remain attached to the data throughout analysis and visualization.
Individual identity is equally important. The known jaws and isolated teeth do not constitute one associated skeleton. Specimens from different individuals, sites, or periods should retain their separate records. A species-level model can learn from their shared features and variation, but combining them must not silently produce a fictional individual whose anatomy was supposedly observed in full.
Four proposed mappings and their available inputs
The broader reconstruction program contains four mappings. Each requires a particular kind of evidence, and successful prediction in one domain does not establish accuracy in the others.
Proposed system
Input and intended target
Application to Gigantopithecus
Fossil2Skeleton
Preserved skeletal fragments to distributions of missing skeletal anatomy
Jaw and tooth geometry provide a starting input for a testable reconstruction program.
Skeleton2DNA
Skeletal morphology and comparative evidence to genomic hypotheses
A further research step, combining morphology with evolutionary context and surviving proteins.
Endocast2Brain
Endocranial geometry to probabilistic brain anatomy
No direct application with the known material because a braincase or endocast is unavailable.
Skull2Face
Cranial geometry to probabilistic facial soft tissues
No direct application with the known material because the necessary skull is unavailable.
The missing cranial evidence is fundamental. The documented absence of cranial remains prevents an anatomical reconstruction of the face or endocranial cavity from beginning with the relevant fossil geometry (Welker et al., 2019). Fossil2Skeleton could generate candidate skulls as part of whole-skeleton hypotheses, but those candidates would remain predictions. Feeding them into another system would extend a chain of assumptions and would need to preserve the uncertainty inherited from every preceding step.
Learning the skeleton from the preserved jaw
Fossil2Skeleton would first require a comparative collection in which each input jaw can be associated with the remaining bones of the same individual. Three-dimensional scans could preserve outer surfaces, dental arrangement, and internal structures where imaging permits. Measurements should include absolute dimensions as well as shape, because rescaling every specimen to the same size would remove information needed to investigate body size.
The fragment and the target skeleton would each be represented numerically, using collections of values rather than a single scalar. A jaw could be encoded as surface points, mesh patches, anatomical landmarks, and local measurements. The skeleton could be represented as separate bone embeddings connected by a graph of anatomical relationships. Masks would identify which regions are observed, damaged, absent, or merely unrecorded. This prevents an unscanned bone from being interpreted as a biological absence.
Let denote the observed fossil geometry, its observation mask, a complete skeletal configuration, and comparative context. The proposed target is:
This conditional distribution expresses which skeletal configurations the trained model considers compatible with the available information. Sampling from it would produce alternative skeletons. Its usefulness would depend on the quality of the training data, the accuracy of the model, and whether its uncertainty estimates remain credible when applied beyond the training sample.
Training could proceed in three stages. The first would learn representations of bone geometry from a broad collection of scans. The second would repeatedly conceal measured bones or regions and train the model to predict them from the retained anatomy. The third would evaluate transfer to unfamiliar species and to inputs with fossil-like fragmentation or deformation. Incomplete comparative specimens could contribute to representation learning and to prediction tasks whose targets are actually documented; missing target bones should not be supplied as invented ground truth.
A transformer provides one possible implementation. Representations of candidate bones would attend to the available fragment tokens, while interactions among the candidate bones would coordinate their relative dimensions and joint geometry. A conditional diffusion model could generate alternatives by learning to remove noise from skeletal representations while consulting the observed fragments. During training, prediction errors would change the model’s parameters. During reconstruction, repeated denoising would refine a candidate under the learned model. This iterative process does not require reinforcement learning, although additional evaluative objectives could be investigated.
Relevant components have precedents. PoinTr applies transformers to completion of partial point clouds (Yu et al., 2021), and statistical shape modelling has been used to reconstruct missing portions of fossil pelves (Torres-Tamayo et al., 2025). The proposed Gigantopithecus application requires the further step of learning relationships across separate skeletal regions and assessing their transfer to a poorly known extinct ape.
What a useful reconstruction would answer
The first outputs should include measurements that can be evaluated before investing in highly detailed surfaces. Candidate targets include long-bone lengths, relative limb proportions, pelvic dimensions, vertebral proportions, and joint sizes. Some may prove predictable from jaw anatomy; others may remain weakly constrained. Each result should report a distribution and its performance on relevant comparative tests.
A central question concerns the relationship between dental specialization and total body size. A large jaw can reflect several interacting influences, including overall growth and the demands placed on the feeding apparatus. The model should be tested for whether it can distinguish these influences in known animals. Otherwise, it could assign excessive body size to an animal whose teeth are unusually large relative to its skeleton. The training design should include comparisons among species with differing dental proportions, as well as variation among individuals within species.
Evolutionary context should enter as information whose effects can be measured. An orangutan relationship may alter the prior distribution over possible anatomy, but the model should also be run with the jaw geometry alone and with alternative comparative samples. If adding a taxonomic label changes the predicted limbs more than adding detailed jaw geometry does, that dependence should be reported. It would reveal which source of information is driving the reconstruction.
Broad training may help the system learn general relationships among skeletal structures. Its advantage should be tested against more restricted primate training and against broad pretraining followed by primate-specific adaptation. Existing pelvic reconstruction work found that the best reference grouping differed among specimens (Torres-Tamayo et al., 2025). The relevant criterion is predictive performance on unfamiliar anatomy, rather than the number of species in the training set alone.
Interpretation should rely on controlled changes to inputs. Researchers could withhold particular tooth measurements, mask a mandibular region, or compare predictions before and after removing absolute scale. If estimated limb dimensions change, the experiment identifies a model dependence that can be examined further. Attention maps may suggest where to investigate, but they do not by themselves demonstrate a biological explanation. A useful explanation must connect an input feature to a reproducible change in prediction and to a relationship that can be checked in comparative specimens.
From skeletal morphology to genomic hypotheses
Skeleton2DNA extends the investigation to a more difficult inverse problem. Genomes influence development through interacting molecular processes, while nutrition, age, hormonal state, mechanical loading, and other conditions also affect the resulting skeleton. Consequently, anatomical similarity can coexist with considerable genomic difference. A method that predicts plausible genomic features from morphology must represent this multiplicity.
My earlier proposal describes a staged approach in which models first learn genomic and anatomical representations, then learn relationships between the two, including forward prediction from genomic information to morphology, and finally investigate inverse prediction from morphology to genomic possibilities (Reser, 2025). For Gigantopithecus, living primates would provide the most directly relevant comparative examples, with broader animal data potentially contributing to representation learning. DNA foundation models such as Nucleotide Transformer demonstrate that large-scale sequence pretraining can support genomic prediction tasks (Dalla-Torre et al., 2025). Establishing an accurate genome-to-skeleton relationship would remain a separate empirical challenge.
The decisive data requirement is paired information. Species reference genomes paired with species-average skeletons could support comparisons among lineages, but they would not establish relationships between individual genetic differences and individual anatomy. The latter task requires genomes and skeletal measurements from the same individuals, together with relevant developmental metadata. A successful design must distinguish these levels of analysis instead of treating species identity as evidence that specific variants caused a particular morphology.
Forward modelling would help evaluate candidate genomic explanations by asking what anatomy each candidate predicts. It would not make the inverse mapping unique, and agreement with the same model that generated a candidate would not independently validate it. The initial targets should therefore be bounded genomic features that can be tested in living species: selected sequence regions, protein-coding differences, or specified regulatory features for which a predictive relationship with morphology has actually been demonstrated. Whole-genome recovery should not be assumed from success on such limited targets.
Surviving enamel proteins offer a particularly useful constraint. Candidate coding sequences can be checked for whether their translations agree with confidently identified ancient amino acids. This would exclude some candidates while leaving others compatible. Because multiple codons can encode the same amino acid, the protein sequence cannot uniquely specify its original DNA sequence. Nor does a small set of proteins supply the unobserved regulatory and noncoding genome. The molecular evidence nevertheless anchors part of the inference to material recovered from the animal itself.
Combining the two systems without manufacturing evidence
Fossil2Skeleton could provide anatomically coordinated alternatives for Skeleton2DNA to consider. The combination must preserve their common dependence on the observed fossils. Generating a thousand skeletons from one jaw does not create a thousand additional biological observations, and repeatedly passing a preferred reconstruction through the models must not artificially increase confidence.
A schematic joint model makes the dependencies explicit. For an associated specimen, let denote genomic variables, its skeletal anatomy,the preserved morphology, measured protein evidence, and external comparative context. One possible factorization is:
Here, the first term describes how the surviving morphology relates to a skeleton, including preservation and measurement error. The second tests compatibility with observed proteins. The third describes the proposed forward relationship between genomic variables and anatomy, with unobserved developmental and environmental influences integrated out. The final term represents comparative expectations about the genomic variables. These are modelling assumptions to investigate, not a claim that all of the required distributions are presently available.
Genomic inference would then average over uncertain skeletons:
This operation prevents one arbitrarily selected skeleton from becoming a fixed anatomical fact. An implementation could approximate the integral by considering appropriately weighted skeletal candidates, or use a joint model that represents missing anatomy internally. A direct model conditioned on the observed jaw should also be tested. Generating a complete skeleton first is useful only if it improves inference, interpretability, or computational efficiency under evaluation.
The equations describe associated evidence. When jaws and protein samples come from different individuals, the model needs separate individual variables linked through a population or species-level model. Shared ancestry can relate their evidence without pretending that one individual’s proteins were extracted from another individual’s jaw. Likewise, a phylogenetic conclusion derived from the same protein data should not be added again as if it were a second independent molecular observation.
Why the brain and face require different evidence
Endocast2Brain would learn relationships between the interior geometry of a braincase and brain anatomy. Paired cranial imaging and brain imaging in living species could supply training examples. Quantitative comparison in humans has found substantial correspondence between endocast and brain morphology, with regional differences in correspondence (Dumoncel et al., 2021). Such evidence motivates a predictive mapping while leaving specific targets, such as regional dimensions or external shape, to be separately evaluated.
An endocast does not directly record neuronal connectivity or the contents of an animal’s experience. The proposed system should state exactly which aspects of anatomy its input can predict in validation. For Gigantopithecus, the immediate obstacle arises earlier: the known remains do not supply the endocranial geometry. Endocast2Brain consequently has no direct fossil input for this application.
Skull2Face would use paired observations of cranial anatomy and facial soft tissues to learn a conditional reconstruction. Training would need to represent differences associated with species, age, sex, and body condition, while separating those variables from the skeletal constraints. The intended result would be a distribution of plausible facial anatomy with explicit uncertainty. A preserved lower jaw could inform a limited portion of that problem, but it does not provide the geometry of the orbits, nasal region, or upper face needed for a full application.
These boundaries explain why the present article concentrates on Fossil2Skeleton and Skeleton2DNA. A synthetic skull or endocast could support an exploratory visualization of a particular hypothesis, but it would not turn the resulting face or brain into an independently supported reconstruction. A newly discovered cranium would change the available evidence and could make both additional mappings relevant to Gigantopithecus.
An experimental program with falsifiable predictions
The first implementation should use a manageable set of landmarks and dimensions from associated primate skeletons. Holding back individuals would test prediction within familiar species. Holding back entire species, and eventually larger evolutionary groups, would test whether the learned relationships generalize beyond close training examples. Repeated scans of the same specimen must remain in the same partition to prevent apparent success from reflecting data leakage.
The model should be compared with simple alternatives: scaling a reference skeleton, predicting from a few jaw measurements, and using phylogenetic averages. The important result is the additional predictive information contributed by detailed morphology. For the genomic branch, analogous comparisons should evaluate evolutionary information alone, evolutionary information plus protein evidence, and the further addition of skeletal morphology. These comparisons make it possible to identify which component improves performance.
Evaluation must measure uncertainty as well as error. Across a suitable held-out benchmark, approximately 90 percent of measured values should fall within intervals labelled as 90 percent predictive intervals. Interval widths also matter, since an interval broad enough to include nearly every possible outcome would achieve coverage without supplying much information. Calibration should be assessed by anatomical region and taxonomic distance, with explicit recognition that performance on living species cannot guarantee reliability for Gigantopithecus.
The unusual scale and dental morphology of Gigantopithecus make extrapolation a central issue. Testing on the largest or most anatomically unusual available comparative specimens would help expose failure modes. Disagreement among independently trained models and sensitivity to reference selection should be reported alongside the candidate reconstructions. Neither agreement nor visual realism is sufficient evidence of accuracy.
For genomic prediction, held-out sequence targets must also be protected from pretraining leakage if the experiment claims recovery of previously unseen information. A model that has already encountered a species’ genome may retrieve its sequence from a taxonomic or morphological cue. That would not establish the ability to infer an unknown extinct genome. Evaluation should specify which species, sequences, and paired observations were excluded at every stage.
What additional discoveries would change
The computational program should sharpen questions for fieldwork and collections research. Candidate reconstructions could be compared to determine which missing bone would most clearly distinguish them. If competing skeletal models agree on mandibular dimensions but differ strongly in predicted femoral proportions, an attributable femur could substantially reduce uncertainty. The same reasoning could prioritize additional scans of existing material when internal structures distinguish otherwise similar hypotheses.
This is a proposed use of expected information gain: estimate how much a possible observation would reduce disagreement among the current alternatives. The calculation would help identify valuable evidence, although it could not determine where that evidence is preserved or guarantee that it will be found. Site selection would still require geological, stratigraphic, and taphonomic expertise.
The cave record needs the same distinction between occurrence and interpretation. Fossils recovered from a cave establish a context of preservation; the further claim that the animal habitually occupied caves requires independent support. Thorough investigation should document how deposits formed and how remains accumulated. Expanding the search to suitable deposits with different preservation histories could be as important as investigating known caves more intensively.
Future discoveries would also provide prospective tests. Publishing quantitative predictions before additional bones are found would allow later specimens to challenge the model. Revised models could then incorporate the new observations while retaining the earlier versions and predictions. The result would be an accumulating record of what the method inferred successfully, where it failed, and which new evidence changed the answer.
A measurable reconstruction program
Gigantopithecus makes the purpose of probabilistic reconstruction concrete. The surviving anatomy offers a starting point from which to investigate the rest of the skeleton, and the molecular evidence provides a separate constraint on a limited part of the animal’s biology. Fossil2Skeleton would test anatomical relationships across bones. Skeleton2DNA would investigate whether morphology adds useful information to genomic inference. Endocast2Brain and Skull2Face specify further questions whose direct application awaits the appropriate cranial evidence.
The scientific contribution would be measured by predictions that outperform relevant baselines, uncertainties that match observed errors, and explanations that survive controlled tests. Even a result showing that particular body proportions remain poorly constrained would improve the reconstruction by identifying where existing confidence is unsupported. A successful program would make Gigantopithecus more precisely understood while leaving a clear record of what future fossils could still overturn.
Dalla-Torre, H., et al. (2025). Nucleotide Transformer: building and evaluating robust foundation models for human genomics. Nature Methods, 22, 287–297. https://doi.org/10.1038/s41592-024-02523-z
Dumoncel, J., Subsol, G., Durrleman, S., Bertrand, A., de Jager, E., Oettlé, A. C., Lockhat, Z., Suleman, F. E., & Beaudet, A. (2021). Are endocasts reliable proxies for brains? A 3D quantitative comparison of the extant human brain and endocast. Journal of Anatomy, 238(2), 480–488. https://doi.org/10.1111/joa.13318
Reser, J. E. (2025). AI-Mediated Reconstruction of Dinosaur DNA from Fossil Morphology and Extant Genomes (Version 1.0). AI Thought.https://doi.org/10.5281/zenodo.17604519
Reser, J. E. (2026). Fossil2Skeleton: Probabilistic Reconstruction of Complete Skeletal Anatomy from Fossil Fragments. Conceptual manuscript, October 21, 2026.
Torres-Tamayo, N., VanSickle, C., Woodert, M. A., Haeusler, M., & Schlager, S. (2025). Statistical shape modelling as a novel reconstruction tool in palaeoanthropology: A case study on fossil pelves. Methods in Ecology and Evolution, 16, 2569–2584.https://doi.org/10.1111/2041-210X.70154
Welker, F., et al. (2019). Enamel proteome shows that Gigantopithecus was an early diverging pongine. Nature, 576, 262–265.https://doi.org/10.1038/s41586-019-1728-8
Yi, Z., Zanolli, C., Liao, W., & Wang, W. (2023). Estimates of absolute crown strength and bite force in the lower postcanine dentition of Gigantopithecus blacki. Journal of Human Evolution.https://doi.org/10.1016/j.jhevol.2022.103313
Yu, X., Rao, Y., Wang, Z., Liu, Z., Lu, J., & Zhou, J. (2021). PoinTr: Diverse Point Cloud Completion with Geometry-Aware Transformers. Proceedings of the IEEE/CVF International Conference on Computer Vision. https://arxiv.org/abs/2108.08839
Zhang, Y., et al. (2024). The demise of the giant ape Gigantopithecus blacki. Nature, 625, 535–539. https://doi.org/10.1038/s41586-023-06900-0
Leave a Reply