Origin Neural brings together an extensive molecular prediction catalogue, an exceptionally compact computational model, and blockchain-anchored research provenance. Explore the discovery workflow through an interactive illustrative simulation, then inspect the recorded work and public anchors.
Interactive investor experience. Screening scenarios use synthetic, browser-generated candidates to illustrate the workflow. The catalogue and engine figures are published by Origin Neural; scientific performance can be evaluated through a partner-led benchmark. The demonstrations do not run the real Origin Neural network, and no experimental hit-rate improvement is shown here; real use needs experimental validation. The catalogue size and the anchor count are separate company-reported figures; this page does not state how many catalogue records each anchor covers.
A small model (reported: 73,352 parameters) ranks candidates by modelled binding fit, on commodity hardware.
Absorption, distribution, metabolism, excretion and toxicity proxies can deprioritise likely dead ends before they cost a bench slot.
Scaffold analysis lets a shortlist span chemical families rather than cluster in one, so a single failure mode is less likely to sink a series.
Batch hashes are timestamped on a public ledger. Origin Neural's public site publishes the anchors and a verification script.
Explore a virtual library of 20,000 compounds. Adjust target complexity, model signal and ADMET assumptions to see how ranking and filtering can focus a research team’s assay budget. All scenario outputs are illustrative.
Every simulation includes a baseline calibration control. Results reflect your selected assumptions and demonstrate the mechanics of prioritization rather than measured Origin Neural performance.
The same 20,000 virtual compounds, the same assay budget. Left: testing in random order. Right: testing in the order produced by the simulated model and ADMET filter. Each square is a compound; a bright square is a useful hit.
The left lane shows one random ordering, so its count can differ from the average over 200 random orderings in the table below. Squares are not positioned by chemistry; the layout is only a grid.
A useful hit is truly active and free of a toxicity liability.
A team that has already measured a compound on a few kinases can use those readings, together with its structure, to predict the rest of the panel. These are recorded retrospective results on public data, not a live run. Research prototype. Retrospective evidence on public kinase panels only. Not ready for pharma decisions.
Relative reduction in prediction error against structure alone, after revealing 3 measured kinases per held-out compound. Bars show 95% intervals where the study reports one.
What this shows: existing measurements make predictions for new compounds noticeably better on every public panel tested, pooled over the exact (uncensored) measurements that were not revealed. The interactive view below shows that on DAVIS the error fell for 34 of 68 compounds, half; in an exploratory split chosen after seeing the results, the gain was concentrated in the 30 compounds with at least 100 exact measurements (24 of 30 improved) and absent in the other 38 (10 of 38 improved; error rose from 0.985 to 1.034). What it does not show: kinases only; no wet-lab or prospective test; the revealed and predicted values come from the same screen, so whether the gain holds for measurements from a different assay is not tested; under this scoring the gain is smaller or absent for compounds whose revealed readings are not exact values; it needs measurements that already exist, and the cost of buying new assays is unknown. On PKIS2 the gain fell short of its pre-registered 8% target, and picking hits within a single compound did not meet its target. The compared structure-only models see no measurements of the new compound. Full method and limits are in the card under Research milestones.
Each compound below was held out, so the predictor never saw any of its values. Choose a compound and reveal up to five kinase measurements, in the order the method picks them (an ordering specific to this data set: the best three-kinase panel was not stable across data sets, so no panel is recommended), then watch the predictions for the remaining kinases move toward the measured values. These are recorded outputs of the same predictor behind the 1.159 to 0.949 result, not computed in your browser.
The simulator above is the mechanism. These are the five ways a partner can put it to work, and what each is designed to do for a partner. Scope, targets and terms are agreed per partner.
33.3M predictions already generated and sealed, scoped to a target family, disease area or scaffold class. A partner can start from a ranked shortlist instead of a blank library.
A 73,352-parameter model is designed to run on commodity hardware the partner may already own, against proprietary targets. Designed so their data stays on their own infrastructure, with no supercomputer to buy.
Scoped with the partner: a target is provided, and Origin Neural describes restarting its pipeline to deliver a new anchored prediction set to specification.
Anchoring a hash of a partner's own results gives a dated, tamper-evident integrity record for partner workflows, subject to key custody and independent legal and regulatory assessment. It evidences that a hash existed by a date, not who invented what and not patent ownership.
Joint work on a disease area or compound family, with terms built around shared outcomes rather than a flat data purchase.
Subject to target suitability and agreed inputs, a sample set against a partner-selected target, so their team can judge quality directly before any commercial terms.
Published catalogue composition, organized to help research partners identify the most relevant starting sets. PUBLISHED COMPANY METRICS Figures are reported by Origin Neural and available for partner diligence.
The catalogue offers multiple discovery entry points: more than 21 million peptide-like predictions, 12.1 million drug-like entries, and a 295,756-entry VAULT-tier shortlist. Partners can choose a focused evaluation set and independently benchmark it for their therapeutic program. About two thirds of the catalogue is peptide-like; a programme limited to conventional oral small molecules would start from the 12.1 million drug-like subset, which is a starting filter, not a guarantee of oral suitability. The reported model metrics describe model behaviour, not biological validity.
Explore complementary SmartLedger research prototypes in pathway analysis, separately checkable computational certificates and evidence-aware records. These are distinct research demonstrations, shown to illustrate the breadth of the laboratory’s approach.
Many drug and metabolic-engineering questions come down to one: which small set of reactions, if switched off, stops a network from making something? Click a reaction below to knock it out and watch the product go dark, or pick one of the minimal cut sets the page found by trying every combination.
Toy network. Twelve invented reactions, a deliberately simplified model (a reaction runs if all its inputs can be made; no stoichiometry or flux amounts). It illustrates the idea of a cut set. It is not Origin Neural's pathway engine and not a real organism.
A cut set is "minimal" when no smaller part of it already blocks the product. The page finds them by exhaustive search over every combination of the twelve reactions and re-tests each one with the same reachability rule you see in the picture. Real networks have hundreds to thousands of reactions, which is why efficient search and checkable answers matter.
The toy network has twelve reactions. Real published models have hundreds to thousands. These are recorded results from SmartLedger's pathway-analysis research prototype on two public models, each answer paired with a certificate that a separate checker, written by the same team, verifies. RECORDED · INTERNAL · NOT INDEPENDENTLY REVIEWED
Model of 469 reactions. The proof was accepted by three separately implemented checkers, all internal and AI-written: two by our own team, one by another project's AI session.
The certificate was verified with standard-library code and no solver. The same certificate with one branch deleted is rejected.
Among the model's non-essential genes, with an LB-like medium proxy we designed (exploratory). 162 lethal pairs, each certified; the cobra package agrees on all 162 lethal calls (other calls were not cross-checked that way). No experimental double-knockout data was compared.
Terminal output recorded on a laptop on 12 November 2026 (3 workers; the demo finished in about 22 seconds; file paths removed). Step 3 deletes one branch of the proof and the checker rejects it. Recorded run, public red blood cell model; cone semantics; no biological claim. This is a recording: nothing here runs in your browser.
model sha256: 83a69bb88844c5490ee690c77394646372cc8f74b3c4abb9315fe6f81e22cca0 (we measured 83a69bb8 on 2026-11-09; a different value means BiGG changed the file: numbers may differ)
== 1. all minimal cut sets of size <= 2 for NaKt (sodium-potassium pump), no mode enumeration (about 50 s) ==
COMPLETE up to size 2: 52 minimal cut sets {'1': 6, '2': 46}; certificate with 781 members verified exactly in 6 s
ok: 52 minimal cut sets (6 single reactions + 46 pairs)
== 2. verify the certificate with the exact checker (standard library only, no solver) ==
VERIFIED: 781 members; 52 minimal cut sets up to the certificate bound
== 3. tamper: delete one branch of the proof, then verify again (must be REJECTED) ==
REJECTED: closure fails: K=[] has witness support containing 15 but [15] is not a member and has no blocking proper subset among the members
ok: the tampered certificate was rejected
== 4. a REVERSIBLE target: GLCt1 (glucose transporter), size <= 2 (about 50 s) ==
COMPLETE up to size 2: 31 minimal cut sets {'1': 2, '2': 29}; certificate with 870 members verified exactly in 9 s
ok: 31 minimal cut sets (2 single + 29 pairs)
DEMO COMPLETE. Scope: cone semantics (bounds decide direction only), one published model, sizes up to the stated bound, no biological claim.
Our elementary-mode counts match efmtool on the E. coli core model (100,274 modes), and our cut-set lists match StrainDesign at five test points. SmartLedger’s distinctive research direction is a separately checkable certificate accompanying the computed answer.
Limits. A mathematical model of a published network, not a biological or drug-target result. No partner has used it as of 12 November 2026. The checkers were written by AI sessions from one written specification and have had no human review. StrainDesign, an established tool, found the same lists; what we add is a certificate that a separate checker can verify. The targets, medium and settings behind each count are not shown here, and the counts depend on them. These are recorded results from internal runs: they are not re-run in your browser, and the certificates and checkers are not published here.
Public models from BiGG Models (King et al., Nucleic Acids Research 44:D515, 2016). Red blood cell model iAB_RBC_283 (Bordbar et al., BMC Systems Biology 2011); E. coli model iML1515 (Monk et al., Nature Biotechnology 2017).
Two live demonstrations of how results stay checkable.
Origin Neural's public site describes sealing each batch of predictions into a Merkle tree and anchoring the root on a public chain. This panel is a local demonstration of the underlying idea, with eight synthetic records: edit any value and watch the hash check fail.
| Record (synthetic) | Modelled score (editable) | Leaf hash | Leaf |
|---|
What this shows: changing any single record changes its leaf hash and therefore the root, so it no longer matches the baseline, and the page marks the record that differs. What it does not show: real tamper evidence, which needs a root that was fixed and kept independently beforehand (for example anchored on a public chain). Here the page sets its own baseline on load, so it cannot detect changes made before then, or changes to both records and baseline. It also does not show the real catalogue's hashes, signatures or blockchain anchors; here A real anchor evidences that a hash existed by a block time, not that a prediction is correct.
Every attribute on a research record carries the kind of evidence behind it: a model prediction, an instrument measurement, or a human read. A gate like this, placed between signing and publishing, can refuse any record that claims more than its declared evidence supports. This is an illustrative design, not a description of an existing Origin Neural control. Change the evidence kinds and the claim, and watch the gate decide. All data is synthetic.
| Attribute (synthetic) | Value | Evidence kind |
|---|
Rule used here: a "declared evidence complete" claim needs every attribute declared as an instrument measurement or a human read; any declared model prediction caps the record at "Model-prioritised". The gate checks declared labels only; it cannot verify that a label is true, and a measurement or human read still needs its own acceptance criteria and review. The rule is illustrative. This is a data-integrity mechanism, not regulated validation: no claim is made about 21 CFR Part 11, Annex 11 or any qualification standard.
A partner names one target. We prepare a sample prediction set against it, so their team can judge quality directly.
Compare the sample with your own reference sets. The company itself recommends independent benchmarking before any commitment.
Catalogue access, engine licensing, a commissioned campaign, the verification layer, or joint research.
A research prototype of certificate-checked computation. An engine returns an answer together with a certificate; a small separate checker, written by the same team, verifies the certificate. This panel replays a recorded run of that check on two real ligand pairs from a public benchmark, then shows what happens when the certificate is damaged.
Recorded, not live. The verdicts below were recorded from real runs of the checker on 12 November 2026. Your browser is replaying stored results; it is not running the checker, and the certificates and checker are not published here. Treat this as a demonstration of behaviour, not independent evidence.
What was verified in the recording: an alignment between two real ligands is valid, and no alignment with more matched atoms exists, under the engine's objective (the number of matched atoms only). A damaged copy of the certificate is rejected, with the reason the checker names. Limits: graph objective only (the shared piece may be disconnected; no aromaticity, stereochemistry or tautomers are modelled); hydrogens suppressed; atom and bond labels are elements and SMILES bond orders; the checker has been written and tested by its authors and has not been reviewed by an outside person; development runs, not a benchmark. The graphs are taken as given; the check does not confirm that they are the molecules. This is about exact combinatorial answers. It says nothing about binding, activity, safety or the quality of any prediction, and Origin Neural's predictions are not certified by it.
Ligand pairs: openforcefield/protein-ligand-benchmark, commit fd88824f9114244f95a14b485e6d6c96c1de716d, LOMAP perturbation network edges, CC-BY-4.0 (https://github.com/openforcefield/protein-ligand-benchmark). Heavy-atom graphs derived by SmartLedger from the ligand SMILES.
Everything above is a demonstration. These are real: hash-only records and certificates anchored on public blockchains. Each link opens a public lookup run by a service that is not this page. Origin Neural's own batch anchors and verification script are on its verification page. The lookups below are SmartLedger records. A lookup alone cannot show what was hashed, who authored it, or who controls the signing key. PUBLICLY TIMESTAMPED RESEARCH RECORDS
f71ff284f2df3861d209d7d3…
e31f6004e1c8ebf31426ffa8…
40bbc255a61e1e6c5858b9a3…
2e9e5c436eb0c5ac333e546e…
d74de492659c7c214ddd0564…
Five hash-only records of SmartLedger work (software releases and research projects, not all of it research), each claimed by its authors, signed with their key and anchored on both BSV and Bitcoin. The contents and the blinding values never leave the authors' machines; the lookup shows only whether and where the hash is anchored. The block numbers above were read on 12 November 2026. If the holder supplies the matching files and evidence package and the commitment verifies, a record supports that those bytes existed no later than about the block's time and that the signer made a signed statement. A lookup alone does not show that, and a record does not show the science is correct, it is not a registered right, and it makes no regulatory-compliance claim. Fully verifying a record needs its evidence package, which the authors hold.
NotaryHash certificates for four public research records. Each page hashes a file you choose locally (it is never uploaded) and compares it with the certificate anchored in a BSV block. The file hash is computed locally in your browser. The check against the blockchain relies on header sources and may be inconclusive in some browsers; a matching hash alone is not a full check of the block. If everything matches, it supports that those exact bytes existed no later than about that block's time. It does not show a paper is correct. The page runs on a different site from this one.
Certificate 1 → Certificate 2 → Certificate 3 → Certificate 4 →
Certificates can be signed with the post-quantum algorithms ML-DSA (FIPS 204) or SLH-DSA (FIPS 205), among 16 algorithms. The software uses those algorithms; it is not a validated cryptographic module and makes no FIPS-compliance claim. Many receipts can be batched under one transaction, each with its own inclusion proof; that design is tested but has not been load-tested at scale.
Selected recorded computational research milestones, illustrating reproducibility, cross-implementation comparison and integrity testing beyond the molecular prediction catalogue.
The cross-validation loop behind the kinase study became about 11x faster per fold (closed-form leave-one-out and a Schur-complement greedy search) with results identical to 4e-16 against the original; rejected ideas are documented. Limits: one evaluation pipeline, one machine.
All 529 ligand pairs of the public protein-ligand-benchmark (CC-BY-4.0) now fall inside the size limits of our certified maximum-common-substructure engine. A 40-pair sample, and 296 pairs from nine targets, were solved with certificates accepted by its checker. Comparing the atom mappings distributed with the benchmark against those maxima: 93.9% (278 of 296) are valid common substructures under our bond labelling, and 63.6% (175 of 275 valid, proved cases) equal the maximum; in the rest a larger common substructure exists. This is a statement about a graph objective, not about any planner. Limits: heavy atoms only; atom count only (the shared piece may be disconnected); depends on the bond-labelling convention; development runs on a shared laptop; no outside person has read the checker.
Replacing a rational Gauss-Jordan routine with Dixon lifting (a published p-adic method) made that step 5.97x faster in total over 23 public netlib problems (median per problem 4.3x, range 1.5x to 23x), with all 47 solutions identical; every answer is verified exactly before use. Limits: one internal step compared with our own earlier routine; the whole solve is projected to speed up about 1.5x (not measured); these 23 problems only; shared laptop; not part of any deployed build.
In pre-registered tests on small networks trained on modular arithmetic, matching the code to the task’s symmetry cut the bits for the token tensors to about 0.6x of the best conventional codec (40 fresh models, within 1% of the original validation accuracy), while the same code on the wrong axis was 1.9x worse. A follow-up combining it with a standard low-rank fold shrank the whole model to about 0.6x of the conventional baseline (80 fresh models); most of that came from the fold, and the symmetry-specific part added a further 12 to 22%. Limits: one small architecture, synthetic only, no link to biology; the group axis was read from the network’s own outputs and only works when the network implements a clean group operation; the design was reviewed before the runs by a second AI model, not by human peers.
A 2-bit packed ternary matrix-vector product (2 bits per weight, AVX2) was about 1.75 to 2.1x faster than int8 on DRAM-sized matrices (320 and 640 MiB), on two AMD CPUs and also on an Intel laptop CPU (one run per machine). The gain is specific to DRAM-sized problems, where memory traffic dominates; for matrices that fit in cache, decode cost dominates instead. Limits: single thread, synthetic random ternary weights, so no accuracy claim, no energy measurement, and the speedup is untested for multi-thread or larger-batch matrix products. Not a general result.
On the public DAVIS kinase-inhibitor set (68 compounds, leave-one-compound-out), retrospectively revealing three existing kinase measurements per held-out compound and combining them with a structure-based predictor cut the prediction error (RMSE, on the exact (uncensored) values that were not revealed) from 1.159 to 0.949; a target-mean baseline scored 1.311. The gain appears on five public kinase panels from several labs and platforms, with partly overlapping chemistry (DAVIS, Metz, PKIS, Anastassiadis and, newest, PKIS2, a panel not examined before): the relative reduction in prediction error against structure alone is roughly 10 to 18% on the first four panels and 7.1% (interval 5.3 to 9.2%) on PKIS2, which fell short of the 8% pre-registered target; the size of the structure’s contribution varies by panel. A fresh agent that re-implemented the predictor from the written method alone, without seeing the code or results, reproduced the Metz gain (10.5%, interval 7.4 to 13.6%, against 11.9%) with the same leak check and wrong-reveal control; that is one panel, the same data loader and the same team’s protocol description. On the pre-registered PKIS screen (345 compounds by 214 kinases, largely different chemistry, held-out compounds from different chemical clusters), structure plus 3 measured kinases cut the error by 13.5% (95% interval 10.0 to 17.3%) against the structure-only baseline. On the fourth panel (Anastassiadis, 160 commercial inhibitors by 246 kinases, pre-registered), structure plus 3 measured kinases cut the error by 17% (interval 9 to 25%) against structure alone. Retrospectively, after measuring three selected kinases per compound in the same PKIS screen, the combined model ranks more cells with at least 50% inhibition near the top than structure alone (average precision 0.55 against 0.37; enrichment in the top 5% of cells 10x against 7x). The combined predictor beats fingerprint, random-forest, neural-network and small pretrained-embedding (ChemBERTa-77M) structure-only baselines, the best chosen on the test folds, by about 12% on PKIS and 11% on Metz; the public pretrained model weights were statically audited and run on a separate machine with internet access blocked. This compares a predictor that sees 3 measured responses of each new compound with models that see none; larger, graph-based or fine-tuned pretrained models and protein-aware models were not tested. Checks included a leak test, and revealing the wrong compound’s values removed the benefit. On PKIS2, pooled ranking of hit cells improved (average precision 0.458 against 0.326), but within each compound the improvement in the top five picks was only 0.011, below the 0.05 pre-registered target, so selecting hits within a single compound is not shown, and most of the error gain is a compound-level effect. Limits: kinases only; in an exploratory split chosen after seeing the results, the DAVIS gain was concentrated in the 30 compounds with 100 or more exact measurements (better for 24 of those 30) and absent in the other 38 (better for 10; error 0.985 to 1.034), and the error fell for 34 of 68 compounds overall, as the interactive view above shows; one panel (Metz) is also the main source of a second analysis, and the chemistry on the fourth panel overlaps heavily with DAVIS and Metz, so these are new labs and platforms, not independent new chemistry; fingerprint-based baselines only; same-screen measurement effects not ruled out; a smaller gain at the lowest PKIS concentration; the benefit is access to existing measurements, not shown for newly bought assays, and assay cost is unknown; no wet-lab confirmation and not tested prospectively; the three kinases are chosen per dataset, so no panel is recommended. An exploratory research result, not a decision tool.
Origin Neural’s docking-scoring module (BioPrime v4, as described by the company) was re-run by our own team, from the company’s code and its ten bundled public crystal structures of drug targets (chosen by the code’s authors), on a laptop with no network access. The test takes the crystal ligand pose and perturbs copies of it: five rotations (15 to 90 degrees) for an anchor-fidelity score, and five shifts (0.5 to 5 Å) for a score named “induced fit” in the module. It asks whether the anchor-fidelity score collapses for the rotated copies and whether the crystal pose outscores the shifted copies. Rotated copies fell below the 0.1 anchor-fidelity cut-off in at least 3 of 5 cases on 8 of 10 structures, and the crystal pose outscored all five shifted copies on 8 of 10; both held on 6. Rotated copies rejected, out of 5 (structure, protein): 1HVR HIV-1 protease 4, 1T3R HIV retropepsin 4, 6LU7 3C-like proteinase 3, 1AQ1 CDK2 4, 1M17 EGFR 2, 3POZ EGFR 5, 5L2I CDK6 3, 3ERT estrogen receptor 4, 1S19 vitamin D receptor 4, 3RFM adenosine A2A receptor 0. The two structures that missed the anchor criterion were 1M17 and 3RFM; the crystal pose did not outscore every shifted copy on 6LU7 and 5L2I. Limits: this is a self-consistency check, not docking accuracy or affinity prediction; it does not search for poses, rank compounds, or compare against other docking programs; the 0.1 and 3-of-5 cut-offs were set by the authors and not pre-registered; one run, and no wet-lab data. An internal result, not a decision tool.
A closed-form count of extreme rays was fitted on smaller cases and recorded before the largest run. It held exactly: 3,145,817 rays, and a second session's run gave the same count. Limit: a count only; the formula is fitted, not proved; nothing for odd cases.
On a public red-blood-cell network model, a recorded run identified 52 minimal cut sets up to size two and verified its 781-member certificate using a separate standard-library checker.
A safety check that stops a transfer from sending a token to fees agreed with a separately written implementation on 200,000 generated cases. Agreement on generated cases can still miss shared errors. Limit: covers the transaction shapes the software builds; one known gap is open.
A conformance method agreed 435 of 435 test rows across two live engines, and every group carries a control that fails on purpose if its data is altered, so a pass cannot be empty. Limit: covers one protocol's rules, nothing biological.