HyInv Technical Report (Hyra Team, 2026-10-14)
Dated 2026-10-14; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (21 pages) · Project source
What this source contributed
Records in this release that cite this source. A record can cite several sources.
Results
Values as the source printed them. Compare results only under compatible protocols, datasets and metrics.
| Model or configuration | Metric | Value | Dataset | Evaluation |
|---|---|---|---|---|
| HyInv BF16 path (update 288, equal-diversity read-out) (HyInv report) | Sctm at fixed diversity difference (BF16 path minus the FP32 reference sampler at development diversity D = 0.30, four-point quadratic fit, paired by chain, with a 95% interval; ESMFold2, 2 designs per chain, draft temperature 0.6; checkpoint at update 288; the reference system is the FP32 reference sampler) | −0.0008 [−0.0021, +0.0022] unitless | CATH4.3 validation subset 1 (512 chains) | BF16 path minus the reference sampler on Subset 1 at fixed diversity (HyInv report) |
| HyInv BF16 path (update 288, equal-diversity read-out) (HyInv report) | Sctm at fixed diversity difference (BF16 path minus the FP32 reference sampler at development diversity D = 0.30, four-point quadratic fit, paired by chain, with a 95% interval; ESMFold2, 2 designs per chain, draft temperature 0.6; checkpoint at update 288; the reference system is the FP32 reference sampler) | +0.0005 [−0.0013, +0.0032] unitless | CATH4.3 validation subset 2 (512 chains) | BF16 path minus the reference sampler on Subset 2 at fixed diversity (HyInv report) |
| HyInv BF16 path (update 288, equal-diversity read-out) (HyInv report) | Sctm at fixed diversity difference (BF16 path minus the FP32 reference sampler at development diversity D = 0.30, four-point quadratic fit, paired by chain, with a 95% interval; ESMFold2, 2 designs per chain, draft temperature 0.6; checkpoint at update 288; the reference system is the FP32 reference sampler) | +0.0001 [−0.0032, +0.0018] unitless | CATH4.3 validation subset 3 (505 chains) | BF16 path minus the reference sampler on Subset 3 at fixed diversity (HyInv report) |
| HyInv BF16 path (update 288, equal-diversity read-out) (HyInv report) | Sctm at fixed diversity difference (BF16 path minus the FP32 reference sampler at development diversity D = 0.30, four-point quadratic fit, paired by chain, with a 95% interval; ESMFold2, 2 designs per chain, draft temperature 0.6; checkpoint at update 288; the reference system is the FP32 reference sampler) | −0.0001 [−0.0009, +0.0016] unitless | CATH4.3 validation (1,529 chains, development panel) | BF16 path minus the reference sampler on Whole development panel at fixed diversity (HyInv report) |
| CMLM (HyInv report) | lDDT (CA-only lDDT-style score without superposition, over all CA pairs below 15 Angstrom in the native structure, as the fraction whose distance error is below 0.5, 1, 2 and 4 Angstrom averaged over the four thresholds with the pairs pooled over the whole design; between 0 and 1) | 0.822 unitless | CATH4.2 test (1,120 chains) | CMLM on CATH4.2 test (CA-lDDT of designed sequences under ESMFold2, HyInv report) |
| CMLM (HyInv report) | Sequence recovery (Fraction of valid sites at which the design carries the native residue, averaged over the designs of a target and then over targets; descriptive in this report, with weight zero in HyInv's reward and no role in checkpoint selection) | 0.454 fraction | CATH4.2 test (1,120 chains) | CMLM on CATH4.2 test (Native sequence recovery of designed sequences, HyInv report) |
| CMLM (HyInv report) | Plddt (ESMFold2 mean per-residue confidence over valid sites on a scale from 0 to 100, averaged over the designs of a target and then over targets; from a second fold of the same sequences with the confidence head on) | 77.9 unitless | CATH4.2 test (1,120 chains) | CMLM on CATH4.2 test (pLDDT of ESMFold2 predictions of designed sequences, HyInv report) |
| CMLM (HyInv report) | Sctm (Mean TM-score of the predicted structure to the native backbone, averaged over the designs of a target and then over targets (macro-average); computed on the CA trace at valid sites, normalised by the native length and clipped at one; ESMFold2 as the structure predictor, which is also the reinforcement-learning reward of HyInv) | 0.8063 unitless | CATH4.2 test (1,120 chains) | CMLM on CATH4.2 test (scTM of designed sequences under ESMFold2, HyInv report) |
| CMLM (HyInv report) | Scrmsd (CA RMSD between the ESMFold2 prediction of a design and the native backbone at valid sites after optimal Kabsch superposition with the residue correspondence fixed by position, averaged over the designs of a target and then over targets) | 4.95 angstrom | CATH4.2 test (1,120 chains) | CMLM on CATH4.2 test (Self-consistency RMSD of designed sequences under ESMFold2, HyInv report) |
| CMLM (HyInv report) | Sequence diversity (Mean pairwise normalised Hamming distance between the K designs of a chain over valid sites, then a macro-average over targets; counted at valid sites for every method, because baseline sequences carry X at invalid sites and counting those as identical would lower their diversity) | 0.105 fraction | CATH4.2 test (1,120 chains) | CMLM on CATH4.2 test (Sequence diversity of a design set, HyInv report) |
| CMLM (HyInv report) | lDDT (CA-only lDDT-style score without superposition, over all CA pairs below 15 Angstrom in the native structure, as the fraction whose distance error is below 0.5, 1, 2 and 4 Angstrom averaged over the four thresholds with the pairs pooled over the whole design; between 0 and 1) | 0.843 unitless | CATH4.3 test (1,882 chains) | CMLM on CATH4.3 test (CA-lDDT of designed sequences under ESMFold2, HyInv report) |
| CMLM (HyInv report) | Sequence recovery (Fraction of valid sites at which the design carries the native residue, averaged over the designs of a target and then over targets; descriptive in this report, with weight zero in HyInv's reward and no role in checkpoint selection) | 0.507 fraction | CATH4.3 test (1,882 chains) | CMLM on CATH4.3 test (Native sequence recovery of designed sequences, HyInv report) |
| CMLM (HyInv report) | Plddt (ESMFold2 mean per-residue confidence over valid sites on a scale from 0 to 100, averaged over the designs of a target and then over targets; from a second fold of the same sequences with the confidence head on) | 79.8 unitless | CATH4.3 test (1,882 chains) | CMLM on CATH4.3 test (pLDDT of ESMFold2 predictions of designed sequences, HyInv report) |
| CMLM (HyInv report) | Sctm (Mean TM-score of the predicted structure to the native backbone, averaged over the designs of a target and then over targets (macro-average); computed on the CA trace at valid sites, normalised by the native length and clipped at one; ESMFold2 as the structure predictor, which is also the reinforcement-learning reward of HyInv) | 0.8093 unitless | CATH4.3 test (1,882 chains) | CMLM on CATH4.3 test (scTM of designed sequences under ESMFold2, HyInv report) |
| CMLM (HyInv report) | Scrmsd (CA RMSD between the ESMFold2 prediction of a design and the native backbone at valid sites after optimal Kabsch superposition with the residue correspondence fixed by position, averaged over the designs of a target and then over targets) | 4.61 angstrom | CATH4.3 test (1,882 chains) | CMLM on CATH4.3 test (Self-consistency RMSD of designed sequences under ESMFold2, HyInv report) |
| CMLM (HyInv report) | Sequence diversity (Mean pairwise normalised Hamming distance between the K designs of a chain over valid sites, then a macro-average over targets; counted at valid sites for every method, because baseline sequences carry X at invalid sites and counting those as identical would lower their diversity) | 0.104 fraction | CATH4.3 test (1,882 chains) | CMLM on CATH4.3 test (Sequence diversity of a design set, HyInv report) |
| CMLM (HyInv report) | lDDT (CA-only lDDT-style score without superposition, over all CA pairs below 15 Angstrom in the native structure, as the fraction whose distance error is below 0.5, 1, 2 and 4 Angstrom averaged over the four thresholds with the pairs pooled over the whole design; between 0 and 1) | 0.793 unitless | RecentPDB (121 chains from 111 recent PDB entries) | CMLM on RecentPDB (CA-lDDT of designed sequences under ESMFold2, HyInv report) |
| CMLM (HyInv report) | Sequence recovery (Fraction of valid sites at which the design carries the native residue, averaged over the designs of a target and then over targets; descriptive in this report, with weight zero in HyInv's reward and no role in checkpoint selection) | 0.477 fraction | RecentPDB (121 chains from 111 recent PDB entries) | CMLM on RecentPDB (Native sequence recovery of designed sequences, HyInv report) |
| CMLM (HyInv report) | Plddt (ESMFold2 mean per-residue confidence over valid sites on a scale from 0 to 100, averaged over the designs of a target and then over targets; from a second fold of the same sequences with the confidence head on) | 77.8 unitless | RecentPDB (121 chains from 111 recent PDB entries) | CMLM on RecentPDB (pLDDT of ESMFold2 predictions of designed sequences, HyInv report) |
| CMLM (HyInv report) | Sctm (Mean TM-score of the predicted structure to the native backbone, averaged over the designs of a target and then over targets (macro-average); computed on the CA trace at valid sites, normalised by the native length and clipped at one; ESMFold2 as the structure predictor, which is also the reinforcement-learning reward of HyInv) | 0.7741 unitless | RecentPDB (121 chains from 111 recent PDB entries) | CMLM on RecentPDB (scTM of designed sequences under ESMFold2, HyInv report) |
| CMLM (HyInv report) | Scrmsd (CA RMSD between the ESMFold2 prediction of a design and the native backbone at valid sites after optimal Kabsch superposition with the residue correspondence fixed by position, averaged over the designs of a target and then over targets) | 6.67 angstrom | RecentPDB (121 chains from 111 recent PDB entries) | CMLM on RecentPDB (Self-consistency RMSD of designed sequences under ESMFold2, HyInv report) |
| CMLM (HyInv report) | Sequence diversity (Mean pairwise normalised Hamming distance between the K designs of a chain over valid sites, then a macro-average over targets; counted at valid sites for every method, because baseline sequences carry X at invalid sites and counting those as identical would lower their diversity) | 0.095 fraction | RecentPDB (121 chains from 111 recent PDB entries) | CMLM on RecentPDB (Sequence diversity of a design set, HyInv report) |
| CMLM (HyInv report) | Sctm (Mean TM-score of the predicted structure to the native backbone, averaged over the designs of a target and then over targets; Boltz2 as the structure predictor, with the alignment built from the designed sequence by MMseqs2 against UniRef50; RecentPDB only) | 0.8217 unitless | RecentPDB (121 chains from 111 recent PDB entries) | CMLM on RecentPDB Boltz2 (scTM of designed sequences under Boltz2, HyInv report) |
| CMLM (HyInv report) | Sctm (Mean TM-score of the predicted structure to the native backbone, averaged over the designs of a target and then over targets; Protenix2 as the structure predictor, with the alignment built from the designed sequence by MMseqs2 against UniRef50; RecentPDB only) | 0.8303 unitless | RecentPDB (121 chains from 111 recent PDB entries) | CMLM on RecentPDB Protenix2 (scTM of designed sequences under Protenix2, HyInv report) |
| CMLM (HyInv report) | lDDT (CA-only lDDT-style score without superposition, over all CA pairs below 15 Angstrom in the native structure, as the fraction whose distance error is below 0.5, 1, 2 and 4 Angstrom averaged over the four thresholds with the pairs pooled over the whole design; between 0 and 1) | 0.854 unitless | TS50 (50 chains) | CMLM on TS50 (CA-lDDT of designed sequences under ESMFold2, HyInv report) |
| CMLM (HyInv report) | Sequence recovery (Fraction of valid sites at which the design carries the native residue, averaged over the designs of a target and then over targets; descriptive in this report, with weight zero in HyInv's reward and no role in checkpoint selection) | 0.521 fraction | TS50 (50 chains) | CMLM on TS50 (Native sequence recovery of designed sequences, HyInv report) |
| CMLM (HyInv report) | Plddt (ESMFold2 mean per-residue confidence over valid sites on a scale from 0 to 100, averaged over the designs of a target and then over targets; from a second fold of the same sequences with the confidence head on) | 81.2 unitless | TS50 (50 chains) | CMLM on TS50 (pLDDT of ESMFold2 predictions of designed sequences, HyInv report) |
| CMLM (HyInv report) | Sctm (Mean TM-score of the predicted structure to the native backbone, averaged over the designs of a target and then over targets (macro-average); computed on the CA trace at valid sites, normalised by the native length and clipped at one; ESMFold2 as the structure predictor, which is also the reinforcement-learning reward of HyInv) | 0.8709 unitless | TS50 (50 chains) | CMLM on TS50 (scTM of designed sequences under ESMFold2, HyInv report) |
| CMLM (HyInv report) | Scrmsd (CA RMSD between the ESMFold2 prediction of a design and the native backbone at valid sites after optimal Kabsch superposition with the residue correspondence fixed by position, averaged over the designs of a target and then over targets) | 3.52 angstrom | TS50 (50 chains) | CMLM on TS50 (Self-consistency RMSD of designed sequences under ESMFold2, HyInv report) |
| CMLM (HyInv report) | Sequence diversity (Mean pairwise normalised Hamming distance between the K designs of a chain over valid sites, then a macro-average over targets; counted at valid sites for every method, because baseline sequences carry X at invalid sites and counting those as identical would lower their diversity) | 0.092 fraction | TS50 (50 chains) | CMLM on TS50 (Sequence diversity of a design set, HyInv report) |
| CMLM (HyInv report) | lDDT (CA-only lDDT-style score without superposition, over all CA pairs below 15 Angstrom in the native structure, as the fraction whose distance error is below 0.5, 1, 2 and 4 Angstrom averaged over the four thresholds with the pairs pooled over the whole design; between 0 and 1) | 0.879 unitless | TS500 (500 chains) | CMLM on TS500 (CA-lDDT of designed sequences under ESMFold2, HyInv report) |
| CMLM (HyInv report) | Sequence recovery (Fraction of valid sites at which the design carries the native residue, averaged over the designs of a target and then over targets; descriptive in this report, with weight zero in HyInv's reward and no role in checkpoint selection) | 0.558 fraction | TS500 (500 chains) | CMLM on TS500 (Native sequence recovery of designed sequences, HyInv report) |
| CMLM (HyInv report) | Plddt (ESMFold2 mean per-residue confidence over valid sites on a scale from 0 to 100, averaged over the designs of a target and then over targets; from a second fold of the same sequences with the confidence head on) | 88.4 unitless | TS500 (500 chains) | CMLM on TS500 (pLDDT of ESMFold2 predictions of designed sequences, HyInv report) |
| CMLM (HyInv report) | Sctm (Mean TM-score of the predicted structure to the native backbone, averaged over the designs of a target and then over targets (macro-average); computed on the CA trace at valid sites, normalised by the native length and clipped at one; ESMFold2 as the structure predictor, which is also the reinforcement-learning reward of HyInv) | 0.8794 unitless | TS500 (500 chains) | CMLM on TS500 (scTM of designed sequences under ESMFold2, HyInv report) |
| CMLM (HyInv report) | Scrmsd (CA RMSD between the ESMFold2 prediction of a design and the native backbone at valid sites after optimal Kabsch superposition with the residue correspondence fixed by position, averaged over the designs of a target and then over targets) | 4.37 angstrom | TS500 (500 chains) | CMLM on TS500 (Self-consistency RMSD of designed sequences under ESMFold2, HyInv report) |
| CMLM (HyInv report) | Sequence diversity (Mean pairwise normalised Hamming distance between the K designs of a chain over valid sites, then a macro-average over targets; counted at valid sites for every method, because baseline sequences carry X at invalid sites and counting those as identical would lower their diversity) | 0.075 fraction | TS500 (500 chains) | CMLM on TS500 (Sequence diversity of a design set, HyInv report) |
| HyInv decoupled draft Td = 0.3 (update 288) (HyInv report) | Sctm at fixed diversity difference (Setting minus the unmodified single-temperature sampler of the same checkpoint on the same panel, both read at fixed diversity (D = 0.30 on the development panel, 0.32 on CATH4.3 test, 0.30 on RecentPDB), paired, with a 95% interval; ESMFold2; checkpoint at update 288; the reference system is the unmodified sampler; read at diversity D = 0.32 Both values are read from least-squares quadratic fits of scTM against design-set diversity over a rewrite-temperature grid, so the level is interpolated from a fit and is not a measurement at a matched sampling setting.) | +0.0074 [+0.0060, +0.0096] unitless | CATH4.3 test (1,882 chains) | Decoupled draft, Td = 0.3 over the unmodified sampler on CATH4.3 test at equal diversity (HyInv report) |
| HyInv decoupled draft Td = 0.3 (update 288) (HyInv report) | Sctm at fixed diversity difference (Setting minus the unmodified single-temperature sampler of the same checkpoint on the same panel, both read at fixed diversity (D = 0.30 on the development panel, 0.32 on CATH4.3 test, 0.30 on RecentPDB), paired, with a 95% interval; ESMFold2; checkpoint at update 288; the reference system is the unmodified sampler; read at development diversity D = 0.30 Both values are read from least-squares quadratic fits of scTM against design-set diversity over a rewrite-temperature grid, so the level is interpolated from a fit and is not a measurement at a matched sampling setting.) | +0.0054 [+0.0022, +0.0067] unitless | CATH4.3 validation (1,529 chains, development panel) | Decoupled draft, Td = 0.3 over the unmodified sampler on Development at equal diversity (HyInv report) |
| HyInv decoupled draft Td = 0.3 (update 288) (HyInv report) | Sctm at fixed diversity difference (Setting minus the unmodified single-temperature sampler of the same checkpoint on the same panel, both read at fixed diversity (D = 0.30 on the development panel, 0.32 on CATH4.3 test, 0.30 on RecentPDB), paired, with a 95% interval; ESMFold2; checkpoint at update 288; the reference system is the unmodified sampler; read at diversity D = 0.30 Both values are read from least-squares quadratic fits of scTM against design-set diversity over a rewrite-temperature grid, so the level is interpolated from a fit and is not a measurement at a matched sampling setting.) | +0.0089 [+0.0036, +0.0145] unitless | RecentPDB (121 chains from 111 recent PDB entries) | Decoupled draft, Td = 0.3 over the unmodified sampler on RecentPDB at equal diversity (HyInv report) |
| HyInv decoupled draft Td = 0.6 (update 288) (HyInv report) | Sctm at fixed diversity difference (Setting minus the unmodified single-temperature sampler of the same checkpoint on the same panel, both read at fixed diversity (D = 0.30 on the development panel, 0.32 on CATH4.3 test, 0.30 on RecentPDB), paired, with a 95% interval; ESMFold2; checkpoint at update 288; the reference system is the unmodified sampler; read at diversity D = 0.32 Both values are read from least-squares quadratic fits of scTM against design-set diversity over a rewrite-temperature grid, so the level is interpolated from a fit and is not a measurement at a matched sampling setting.) | +0.0078 [+0.0064, +0.0107] unitless | CATH4.3 test (1,882 chains) | Decoupled draft, Td = 0.6 (HyInv default) over the unmodified sampler on CATH4.3 test at equal diversity (HyInv report) |
| HyInv decoupled draft Td = 0.6 (update 288) (HyInv report) | Sctm at fixed diversity difference (Setting minus the unmodified single-temperature sampler of the same checkpoint on the same panel, both read at fixed diversity (D = 0.30 on the development panel, 0.32 on CATH4.3 test, 0.30 on RecentPDB), paired, with a 95% interval; ESMFold2; checkpoint at update 288; the reference system is the unmodified sampler; read at development diversity D = 0.30 Both values are read from least-squares quadratic fits of scTM against design-set diversity over a rewrite-temperature grid, so the level is interpolated from a fit and is not a measurement at a matched sampling setting.) | +0.0072 [+0.0062, +0.0118] unitless | CATH4.3 validation (1,529 chains, development panel) | Decoupled draft, Td = 0.6 (HyInv default) over the unmodified sampler on Development at equal diversity (HyInv report) |
| HyInv decoupled draft Td = 0.6 (update 288) (HyInv report) | Sctm at fixed diversity difference (Setting minus the unmodified single-temperature sampler of the same checkpoint on the same panel, both read at fixed diversity (D = 0.30 on the development panel, 0.32 on CATH4.3 test, 0.30 on RecentPDB), paired, with a 95% interval; ESMFold2; checkpoint at update 288; the reference system is the unmodified sampler; read at diversity D = 0.30 Both values are read from least-squares quadratic fits of scTM against design-set diversity over a rewrite-temperature grid, so the level is interpolated from a fit and is not a measurement at a matched sampling setting.) | +0.0131 [+0.0065, +0.0196] unitless | RecentPDB (121 chains from 111 recent PDB entries) | Decoupled draft, Td = 0.6 (HyInv default) over the unmodified sampler on RecentPDB at equal diversity (HyInv report) |
| DPLM-2 3B (HyInv report) | lDDT (CA-only lDDT-style score without superposition, over all CA pairs below 15 Angstrom in the native structure, as the fraction whose distance error is below 0.5, 1, 2 and 4 Angstrom averaged over the four thresholds with the pairs pooled over the whole design; between 0 and 1) | 0.820 unitless | CATH4.2 test (1,120 chains) | DPLM-2 3B on CATH4.2 test (CA-lDDT of designed sequences under ESMFold2, HyInv report) |
| DPLM-2 3B (HyInv report) | Sequence recovery (Fraction of valid sites at which the design carries the native residue, averaged over the designs of a target and then over targets; descriptive in this report, with weight zero in HyInv's reward and no role in checkpoint selection) | 0.479 fraction | CATH4.2 test (1,120 chains) | DPLM-2 3B on CATH4.2 test (Native sequence recovery of designed sequences, HyInv report) |
| DPLM-2 3B (HyInv report) | Plddt (ESMFold2 mean per-residue confidence over valid sites on a scale from 0 to 100, averaged over the designs of a target and then over targets; from a second fold of the same sequences with the confidence head on) | 80.7 unitless | CATH4.2 test (1,120 chains) | DPLM-2 3B on CATH4.2 test (pLDDT of ESMFold2 predictions of designed sequences, HyInv report) |
| DPLM-2 3B (HyInv report) | Sctm (Mean TM-score of the predicted structure to the native backbone, averaged over the designs of a target and then over targets (macro-average); computed on the CA trace at valid sites, normalised by the native length and clipped at one; ESMFold2 as the structure predictor, which is also the reinforcement-learning reward of HyInv) | 0.8093 unitless | CATH4.2 test (1,120 chains) | DPLM-2 3B on CATH4.2 test (scTM of designed sequences under ESMFold2, HyInv report) |
| DPLM-2 3B (HyInv report) | Scrmsd (CA RMSD between the ESMFold2 prediction of a design and the native backbone at valid sites after optimal Kabsch superposition with the residue correspondence fixed by position, averaged over the designs of a target and then over targets) | 4.74 angstrom | CATH4.2 test (1,120 chains) | DPLM-2 3B on CATH4.2 test (Self-consistency RMSD of designed sequences under ESMFold2, HyInv report) |
| DPLM-2 3B (HyInv report) | Sequence diversity (Mean pairwise normalised Hamming distance between the K designs of a chain over valid sites, then a macro-average over targets; counted at valid sites for every method, because baseline sequences carry X at invalid sites and counting those as identical would lower their diversity) | 0.000 fraction | CATH4.2 test (1,120 chains) | DPLM-2 3B on CATH4.2 test (Sequence diversity of a design set, HyInv report) |
| DPLM-2 3B (HyInv report) | lDDT (CA-only lDDT-style score without superposition, over all CA pairs below 15 Angstrom in the native structure, as the fraction whose distance error is below 0.5, 1, 2 and 4 Angstrom averaged over the four thresholds with the pairs pooled over the whole design; between 0 and 1) | 0.823 unitless | CATH4.3 test (1,882 chains) | DPLM-2 3B on CATH4.3 test (CA-lDDT of designed sequences under ESMFold2, HyInv report) |
| DPLM-2 3B (HyInv report) | Sequence recovery (Fraction of valid sites at which the design carries the native residue, averaged over the designs of a target and then over targets; descriptive in this report, with weight zero in HyInv's reward and no role in checkpoint selection) | 0.473 fraction | CATH4.3 test (1,882 chains) | DPLM-2 3B on CATH4.3 test (Native sequence recovery of designed sequences, HyInv report) |
Claims
Statements this source supports, each with where in the source it was checked.
- HyInv report: application enzymes: Application enzymes Enzymes with the catalytic site kept. With the catalytic residues fixed, HyInv redesigns TEM-1 beta-lactamase (1BTL, 263 aa), subtilisin BPN' (2ST1, 275 aa) and IsPETase (5XJH, 300 aa) and keeps the predicted catalytic geometry at the level of the natural sequence; the displayed designs have catalytic heavy-atom RMSD 0.67, 0.71 and 0.23 Angstrom. The report states that the named enzymes are homologous to the training data and were chosen for display, and that the figure illustrates the preservation of a fixed geometry, not the design of new enzymes. Figure 11c prints median catalytic-site RMSD bars for HyInv, ProteinMPNN and the native sequence over 128 designs per method, but the bar order cannot be recovered from the text layer, so those nine values are recorded as a gap rather than read by eye. (Section 6 'Enzymes with the catalytic site kept'; Figure 11 panels a and c; Appendix G 'Loop and window redesign and named enzymes'; Table 10 row 'Named enzymes')
- HyInv report: application homology screen: Application homology screen Homology screen of the application panels (Table 10). De novo backbones: 74 backbones, 0 with a passing hit, 28 with partial hits only. Pocket library: 12, 0, 7. Loop and window, strict: 2, 0, 2. Loop and window, relaxed: 6, 0, 4. Named enzymes: 4 backbones, all 4 with a passing hit and 0 partial. A hit passes if its sequence identity to the pretraining library or to the benchmark reference set is at least 0.3 and it covers at least 0.8 of the query or of the target; partial hits stay below the coverage limit. Strict backbones have no passing hit, relaxed backbones relax one criterion of the strict rule, fixed before any design, and the named enzymes are homologous to training data by design. The report states the table documents the panel rule only. (Table 10 'Homology screen of the application panels' and its footnote; Appendix G 'Homology of the panels')
- HyInv report: application homology screen: Application homology screen rule A hit passes the rule if its sequence identity to the pretraining library or to the benchmark reference set is at least 0.3 and it covers at least 0.8 of the query or of the target; partial hits stay below the coverage limit. Strict backbones have no passing hit; relaxed backbones relax one criterion of the strict rule, fixed before any design was made. The report states the table documents the panel rule only. The per-panel counts are in the claims for each row of Table 10. (Table 10 'Homology screen of the application panels', its footnote; Appendix G 'Homology of the panels')
- HyInv report: application loops: Application loops Loop redesign. HyInv redesigns loop windows of 16 to 25 residues in a fixed context, with the window within 0.5 to 1.8 Angstrom of the experimental structure; the three displayed cases are 9T2I (178 aa, window of 25, scTM 0.97, window RMSD 1.3 to 1.8 Angstrom), 10UT (156 aa, window of 16, scTM 0.99, 0.5 to 0.7 Angstrom) and 9SNE (172 aa, window of 20, scTM 0.96, 0.6 to 0.7 Angstrom), each the four highest-scTM designs of 128. The panel is 8 backbones chosen by rule plus 4 named enzymes, 128 designs per backbone and region, 24,672 designs in all; a design succeeds if its whole-chain scTM is at least 0.9 and the region CA RMSD after fitting the unchanged context is at most 1.5 Angstrom, and the success rate is read only where the native refold meets the criterion on a majority of backbones. No success rate is printed as a cell. (Section 6 'Loop redesign'; Figure 12; Appendix G 'Loop and window redesign and named enzymes')
- HyInv report: application pocket library: Application pocket library Pocket-constrained libraries. With the 15 ligand-contacting residues of a 148-residue binder fixed (backbone 9NZE, ligand A1B7L), 373 of 384 designs pass on that one backbone, which the report says is shown for display; the mean pass rate over twelve pocket backbones is 43%. A design passes if its scTM is at least 0.90 and its site CA RMSD at most 1.0 Angstrom. The five displayed designs have scTM 0.98 to 0.99, pocket RMSD 0.37 to 0.73 Angstrom and 41 to 50% pairwise identity. The panel is 12 backbones, ten chosen by rule and two single cases chosen earlier, with 384 designs per backbone and cell; the endpoint is the number of distinct sequences after clustering at 50% identity among the passing designs of a plate, against ProteinMPNN at the temperature whose diversity at the free positions is closest to HyInv's. Retention of the pinned residues is a check of the code and not a result. (Section 6 'Pocket-constrained libraries'; Figure 10 panels a and c; Appendix G 'Pocket-constrained libraries')
- HyInv report: application statistics: Application statistics For the application tasks, monomers are folded with ESMFold2 with one fold seed shared by all generators for the same backbone and chunk, so generators are paired, and a subset of cells is also folded with Boltz2 and Protenix2. Comparators are ProteinMPNN and ProteinZero V2 only. The diversity-matched baseline temperature is chosen from sequences alone, before any fold is read, as the grid value whose mean pairwise Hamming distance over the designed positions is closest to HyInv's; where HyInv has lower diversity than the lowest baseline temperature the match is as close as the grid allows. Differences are paired by backbone, averaged over designs first, with 95% percentile intervals from 10,000 resamples, and the report notes that with four to twelve backbones the intervals are wide. Intervals are not adjusted for multiple comparisons. (Appendix G 'Predictors, comparators and statistics')
- HyInv report: architecture: Architecture A message-passing graph network with 1,957,573 parameters following the encoder-decoder design of ProteinMPNN: width 128, 4 encoder layers and 3 decoder layers, each residue connected to its 48 nearest neighbours with rotation- and translation-invariant distance and orientation features, random initialisation. The encoder runs once per backbone; the decoder predicts amino acid distributions at masked positions from the backbone and the observed sequence context. At inference the model needs neither a protein language model, a structure predictor nor an alignment. (Section 2.1 'Architecture'; Appendix C 'Configuration'; AI4Science/hyinv/README.md opening paragraph)
- HyInv report: baseline diversity mismatch: Baseline diversity mismatch The ten baselines are single published settings at one temperature each, not tuned to the panels, so comparisons are at comparable rather than exactly equal diversity. Of the 70 comparisons (ten baselines in seven panel-predictor columns) all 70 point estimates favour HyInv and 61 have an interval that excludes zero; the nine that include zero are ProteinMPNN (TS500, TS50), ProteinMPNN grouped (TS50), ProteinZero V2 ESM1 (CATH4.3 test, TS500, TS50), ProteinZero V2 ESM2 (CATH4.3 test, TS50) and LigandMPNN (TS50). PiFold, CMLM, LM-Design and DPLM-2 3B are single points at much lower diversity than HyInv, so their differences are not at matched diversity. ProteinMPNN is the only baseline with a temperature sweep, and it is the only baseline for which the report reads an equal-diversity difference. HyInv's diversity is above ProteinMPNN with a clear interval on CATH4.3 test and CATH4.2 test and above ProteinZero V2 (ESM2) on the same two panels, but 0.015 below ProteinZero V2 (ESM2) on RecentPDB; at the development curve's slope of 0.07 scTM per unit of diversity the report puts the largest offset (0.016) at a worth of at most 0.0011 scTM. (Appendix D.1, D.2 and D.3; Appendix A 'Scores at equal diversity'; Section 5.2)
- HyInv report: baseline reproduction check: Baseline reproduction check Some ESMFold2 cells were regenerated for the report with the same weights, temperature and sampler. A regenerated cell is used only where the same procedure reproduces the released score on RecentPDB; the printed absolute scTM differences are 0.0009 (ProteinMPNN grouped, CATH4.3 test), 0.0007 (ProteinMPNN low-noise grouped, four panels), 0.0006 (ProteinZero V2 ESM1, CATH4.3 test), 0.0004 (PiFold), 0.0022 (CMLM), 0.0005 (LM-Design 650M), 0.0001 (DPLM-2 3B) and 0.0005 (LigandMPNN), each on four panels unless stated, all within 0.003. The fraction of identical sequences is 100.0% for every baseline except DPLM-2 3B at 99.2%. ProteinMPNN and ProteinZero V2 (ESM2) were not regenerated. All RecentPDB cells under the three predictors are released values, and the ProteinMPNN temperatures other than the published one were generated for this report. (Table 5 'Baseline settings, weights and reproduction check' and the paragraph after it)
- HyInv report: baseline reproduction check: Baseline reproduction check rule Some ESMFold2 cells were regenerated for the report with the same weights, temperature and sampler. A regenerated cell is used only where the same procedure reproduces the released score on RecentPDB; the printed absolute differences run from 0.0001 to 0.0022, all within 0.003, and the identical-sequence fraction is 100.0% for every regenerated baseline except DPLM-2 3B at 99.2%. ProteinMPNN and ProteinZero V2 (ESM2) were not regenerated: the table prints a dash in both columns for each of them. All RecentPDB cells under the three predictors are released values, and the ProteinMPNN temperatures other than the published one were generated for this report. All ten baselines use the released weights. The per-baseline figures are in the claims for each row of Table 5. (Table 5 'Baseline settings, weights and reproduction check', the paragraph after the table)
- HyInv report: BF16 speed-up against the reference sampler: Reference sampler speedup bf16 Section 4 states HyInv designs 128 sequences for a 402-residue backbone in 50.6 ms on one NVIDIA H20, '52.1x faster than the reference implementation of the sampler and 57x faster than ProteinMPNN'; the abstract gives the same figure as 'about 52x faster than the reference sampler'. This is HyInv's optimised BF16 path against HyInv's own FP32 reference sampler, not against a baseline model, and it is the printed ratio of the two Table 1 times for the 402-residue request (2,633 and 50.6 ms). The 57x recorded as a result is the ratio against ProteinMPNN's released implementation. Both were timed with the supervised-stage weights at rewrite temperature 0.23, not the released checkpoint. (Section 4 'Efficient inference', first sentence; abstract)
- HyInv report: checkpoint selection: Checkpoint selection The checkpoint and the sampler setting were chosen on the development panel alone. The checkpoint is the latest reinforcement-learning snapshot whose development scTM at diversity 0.30 is within 0.002 of the best (update 640); the best-scoring snapshot differs from it by -0.0004, below the noise of one evaluation. The default sampler setting (draft 0.6, rewrite 0.34, development diversity 0.30) was read off the development curve before the test panels were generated with this checkpoint. The development panel also chose the reinforcement-learning learning rate and training configuration. (Appendix A 'Checkpoint and sampler setting')
- HyInv report: decoding rounds ablation: Decoding rounds ablation The default of 8 draft and 8 rewrite rounds was kept because shorter schedules cost more quality than they save time: the scTM at fixed diversity of six schedules with fewer rounds (6+8, 4+8, 8+6, 8+4, 6+6 and 4+4 draft and rewrite rounds) is lower than the 16-pass default by 0.0028 to 0.0074. The per-schedule values appear only in Figure 8b, which prints no cell labels, so only the range is recorded. (Appendix F 'Quality of the BF16 path', final sentences; Section 5.4 'Decoding rounds'; Figure 8b)
- HyInv report: developer: Developer Released by Tencent Hunyuan and trained by the team's Hyra 2.0 research agent. The named authors of the technical report are Haowei Lin, Cai Zhou and Xiangxin Zhou, all affiliated with Tencent Hunyuan. (Technical report 'Author Information' and Abstract; repository README.md header and 'Model releases')
- HyInv report: evaluation panel CATH4.2 test: Panels cath42 test CATH4.2 test: 1,120 chains in 109 clusters, folded by ESMFold2, used for the comparison. (Table 3 'Evaluation panels', row 'CATH4.2 test')
- HyInv report: evaluation panel CATH4.3 test: Panels cath43 test CATH4.3 test: 1,882 chains in 103 CATH topology clusters, folded by ESMFold2, used for the comparison and as a sampler check. (Table 3 'Evaluation panels', row 'CATH4.3 test')
- HyInv report: evaluation panel CATH4.3 validation: Panels cath43 validation CATH4.3 validation: 1,529 chains in 91 clusters, folded by ESMFold2, the development panel: it fixed the sampler setting and the checkpoint and monitored training, and each test panel was evaluated once. (Table 3 'Evaluation panels', row 'CATH4.3 validation')
- HyInv report: evaluation panel RecentPDB: Panels recentpdb RecentPDB: 121 chains in 121 clusters, folded by ESMFold2, Boltz2 and Protenix2, used for the comparison and as a sampler check. (Table 3 'Evaluation panels', row 'RecentPDB')
- HyInv report: evaluation panel TS50: Panels ts50 TS50: 50 chains in 50 clusters, folded by ESMFold2, used for the comparison. (Table 3 'Evaluation panels', row 'TS50')
- HyInv report: evaluation panel TS500: Panels ts500 TS500: 500 chains in 500 clusters, folded by ESMFold2, used for the comparison. (Table 3 'Evaluation panels', row 'TS500')
- HyInv report: exact FP32 speed-up against the reference sampler: Reference sampler speedup exact fp32 The abstract states 'an exact FP32 path that returns the reference sequences is 8.7x faster', against HyInv's own FP32 reference sampler. Table 1's printed times for the 402-residue request give the same ratio (2,633 and 303 ms). The 9.6x recorded as a result is the ratio against ProteinMPNN's released implementation, so the two figures describe different comparisons and neither is a speed-up of HyInv over a baseline at equal optimisation. (Abstract, final speed sentence)
- HyInv report: homology screen of De novo backbones: Application homology screen de novo backbones Homology screen of the application panels. De novo backbones: 74 backbones, 0 with a hit that passes the rule, 28 with partial hits only. (Table 10 'Homology screen of the application panels', row 'De novo backbones')
- HyInv report: homology screen of Loop and window, relaxed: Application homology screen loop window relaxed Homology screen of the application panels. Loop and window, relaxed: 6 backbones, 0 passing hits, 4 with partial hits only. (Table 10 'Homology screen of the application panels', row 'Loop and window, relaxed')
- HyInv report: homology screen of Loop and window, strict: Application homology screen loop window strict Homology screen of the application panels. Loop and window, strict: 2 backbones, 0 passing hits, 2 with partial hits only. (Table 10 'Homology screen of the application panels', row 'Loop and window, strict')
- HyInv report: homology screen of Named enzymes: Application homology screen named enzymes Homology screen of the application panels. Named enzymes: 4 backbones, all 4 with a passing hit and 0 with partial hits only; the named enzymes are homologous to training data by design. (Table 10 'Homology screen of the application panels', row 'Named enzymes')
Models, benchmarks and other records
Models (6)
Methods (1)
Configurations (23)
- CMLM (HyInv report)
- DPLM-2 3B (HyInv report)
- HyInv BF16 path (timing configuration) (HyInv report)
- HyInv BF16 path (update 288, equal-diversity read-out) (HyInv report)
- HyInv decoupled draft Td = 0.3 (update 288) (HyInv report)
- HyInv decoupled draft Td = 0.6 (update 288) (HyInv report)
- HyInv default (HyInv report)
- HyInv exact FP32 path (timing configuration) (HyInv report)
- HyInv FP32 reference sampler (timing configuration) (HyInv report)
- HyInv FP32 reference sampler (update 288, equal-diversity read-out) (HyInv report)
- HyInv min-p 0.1 truncation (update 288) (HyInv report)
- HyInv plain sampling (update 640) (HyInv report)
- HyInv supervised checkpoint, plain sampling (HyInv report)
- HyInv top-p 0.9 truncation (update 288) (HyInv report)
- HyInv unmodified sampler (update 288) (HyInv report)
- LigandMPNN (HyInv report)
- LM-Design 650M (HyInv report)
- PiFold (HyInv report)
- ProteinMPNN (HyInv report)
- ProteinMPNN (low-noise, grouped) (HyInv report)
- ProteinMPNN grouped (HyInv report)
- ProteinZero V2 (ESM1) (HyInv report)
- ProteinZero V2 (ESM2) (HyInv report)
Benchmarks (1)
Protocols (19)
- BF16 path minus reference sampler in scTM at fixed diversity (HyInv report)
- CA-lDDT of designed sequences under ESMFold2 (HyInv report)
- Gain in scTM at fixed diversity from reinforcement learning (HyInv report)
- Gain in scTM at fixed diversity over the unmodified sampler (HyInv report)
- HyInv above the ProteinMPNN temperature curve at HyInv's diversity (HyInv report)
- Latency of a design request on one NVIDIA H20 (HyInv report)
- Native sequence recovery of designed sequences (HyInv report)
- Novel and foldable designs per de novo backbone at matched diversity (HyInv report)
- Paired scTM difference from HyInv under Boltz2 (HyInv report)
- Paired scTM difference from HyInv under ESMFold2 (HyInv report)
- Paired scTM difference from HyInv under Protenix2 (HyInv report)
- pLDDT of ESMFold2 predictions of designed sequences (HyInv report)
- scTM at fixed sequence diversity on the development panel (HyInv report)
- scTM of designed sequences under Boltz2 (HyInv report)
- scTM of designed sequences under ESMFold2 (HyInv report)
- scTM of designed sequences under Protenix2 (HyInv report)
- Self-consistency RMSD of designed sequences under ESMFold2 (HyInv report)
- Sequence diversity of a design set (HyInv report)
- Speed-up of a design request over ProteinMPNN on one NVIDIA H20 (HyInv report)
Datasets (8)
Dataset subsets (3)
Evaluations (485)
- BF16 path minus the reference sampler on Subset 1 at fixed diversity (HyInv report)
- BF16 path minus the reference sampler on Subset 2 at fixed diversity (HyInv report)
- BF16 path minus the reference sampler on Subset 3 at fixed diversity (HyInv report)
- BF16 path minus the reference sampler on Whole development panel at fixed diversity (HyInv report)
- CMLM on CATH4.2 test (CA-lDDT of designed sequences under ESMFold2, HyInv report)
- CMLM on CATH4.2 test (Native sequence recovery of designed sequences, HyInv report)
- CMLM on CATH4.2 test (pLDDT of ESMFold2 predictions of designed sequences, HyInv report)
- CMLM on CATH4.2 test (scTM of designed sequences under ESMFold2, HyInv report)
- CMLM on CATH4.2 test (Self-consistency RMSD of designed sequences under ESMFold2, HyInv report)
- CMLM on CATH4.2 test (Sequence diversity of a design set, HyInv report)
- CMLM on CATH4.3 test (CA-lDDT of designed sequences under ESMFold2, HyInv report)
- CMLM on CATH4.3 test (Native sequence recovery of designed sequences, HyInv report)
- CMLM on CATH4.3 test (pLDDT of ESMFold2 predictions of designed sequences, HyInv report)
- CMLM on CATH4.3 test (scTM of designed sequences under ESMFold2, HyInv report)
- CMLM on CATH4.3 test (Self-consistency RMSD of designed sequences under ESMFold2, HyInv report)
- CMLM on CATH4.3 test (Sequence diversity of a design set, HyInv report)
- CMLM on RecentPDB (CA-lDDT of designed sequences under ESMFold2, HyInv report)
- CMLM on RecentPDB (Native sequence recovery of designed sequences, HyInv report)
- CMLM on RecentPDB (pLDDT of ESMFold2 predictions of designed sequences, HyInv report)
- CMLM on RecentPDB (scTM of designed sequences under ESMFold2, HyInv report)
- CMLM on RecentPDB (Self-consistency RMSD of designed sequences under ESMFold2, HyInv report)
- CMLM on RecentPDB (Sequence diversity of a design set, HyInv report)
- CMLM on RecentPDB Boltz2 (scTM of designed sequences under Boltz2, HyInv report)
- CMLM on RecentPDB Protenix2 (scTM of designed sequences under Protenix2, HyInv report)
- CMLM on TS50 (CA-lDDT of designed sequences under ESMFold2, HyInv report)
Sources (3)
Evidence
Source checking verifies the cited claim or transcription. It does not establish independent reproduction.
Evidence table
Inspect claims, sources and review details
Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.
One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.
24 evidence rows matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|---|---|
| Archive note The PDF itself is not committed; its gzipped text layer is in artifacts/HyInv_Technical_Report.txt.gz and the PDF can be refetched at the pinned commit and checked against artifact_sha256. Source metadata | HyInv Technical Report (Hyra Team, 2026-10-14) No field-specific location recorded Version: Dated 2026-10-14; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (21 pages) | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and column count asserted. |
| Archive sha256 49f48831fd845c24ca557d06f31aec6d0c363ba709f9a8f6aa62f06f0aa00f27 Source metadata | HyInv Technical Report (Hyra Team, 2026-10-14) No field-specific location recorded Version: Dated 2026-10-14; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (21 pages) | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and column count asserted. |
| Archive transformation receipt gzip -n9 of the pdftotext -layout output of the pinned PDF (poppler pdftotext, macOS). Source metadata | HyInv Technical Report (Hyra Team, 2026-10-14) No field-specific location recorded Version: Dated 2026-10-14; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (21 pages) | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and column count asserted. |
| Artifact SHA-256 0f10b390be399c341d1f7497c5a3ca0ac857ce78172c4732bdc86a9c66e46419 Source metadata | HyInv Technical Report (Hyra Team, 2026-10-14) No field-specific location recorded Version: Dated 2026-10-14; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (21 pages) | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and column count asserted. |
| Artifact url https://raw.githubusercontent.com/Tencent-Hunyuan/Hyra-results/0dcc533cc9f8f76c90ecbfe944df816b6672328a/AI4Science/hyinv/HyInv_Technical_Report.pdf Source metadata | HyInv Technical Report (Hyra Team, 2026-10-14) No field-specific location recorded Version: Dated 2026-10-14; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (21 pages) | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and column count asserted. |
Evidence concerns3 values
| HyInv Technical Report (Hyra Team, 2026-10-14) No field-specific location recorded Version: Dated 2026-10-14; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (21 pages) | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and column count asserted. |
| Extraction method pdftotext -layout text layer, parsed by extract/parse.py with every row label and column count asserted. Source metadata | HyInv Technical Report (Hyra Team, 2026-10-14) No field-specific location recorded Version: Dated 2026-10-14; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (21 pages) | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and column count asserted. |
| Licence Apache-2.0 (repository LICENSE) Source metadata | HyInv Technical Report (Hyra Team, 2026-10-14) No field-specific location recorded Version: Dated 2026-10-14; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (21 pages) | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and column count asserted. |
Limitations9 values
| HyInv Technical Report (Hyra Team, 2026-10-14) No field-specific location recorded Version: Dated 2026-10-14; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (21 pages) | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and column count asserted. |
| Media type application/pdf Source metadata | HyInv Technical Report (Hyra Team, 2026-10-14) No field-specific location recorded Version: Dated 2026-10-14; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (21 pages) | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label and column count asserted. |
Sources and history
Release 2026-11-21-adc15377fab2 · Record review: source checked
Technical report for HyInv, with the five-panel inverse-folding comparison against ten published baselines, the sampler and numerics ablations, the inference-time measurements and the four application protocols.
Technical metadata and extraction receipts
Stable ID: hyinv-20261106-source-report
- areas
- proteins-complexes; protein-structure
- contexts
- research
- archive note
- The PDF itself is not committed; its gzipped text layer is in artifacts/HyInv_Technical_Report.txt.gz and the PDF can be refetched at the pinned commit and checked against artifact_sha256.
- archive sha256
- 49f48831fd845c24ca557d06f31aec6d0c363ba709f9a8f6aa62f06f0aa00f27
- archive transformation receipt
- gzip -n9 of the pdftotext -layout output of the pinned PDF (poppler pdftotext, macOS).
- artifact sha256
- 0f10b390be399c341d1f7497c5a3ca0ac857ce78172c4732bdc86a9c66e46419
- artifact url
- https://raw.githubusercontent.com/Tencent-Hunyuan/Hyra-results/0dcc533cc9f8f76c90ecbfe944df816b6672328a/AI4Science/hyinv/HyInv_Technical_Report.pdf
- evidence concerns
- applies to: hyinv-20261106-data-cath42-test; hyinv-20261106-data-cath43-test; hyinv-20261106-data-cath43-validation; hyinv-20261106-data-cath43-validation-subset-1; hyinv-20261106-data-cath43-validation-subset-2; hyinv-20261106-data-cath43-validation-subset-3; hyinv-20261106-data-de-novo-backbones; hyinv-20261106-data-recentpdb; hyinv-20261106-data-ts50; hyinv-20261106-data-ts500; artifact sha256: 0f10b390be399c341d1f7497c5a3ca0ac857ce78172c4732bdc86a9c66e46419; message: Appendix C states 'We do not claim zero sequence overlap between the training data and the evaluation panels', and Appendix G's homology screen is a coverage rule under which partial hits remain (Table 10: 28 of 74 de novo backbones and 7 of 12 pocket backbones have partial hits, and all 4 named enzymes are homologous to training data by design). Results on these panels are therefore not clean held-out measurements and must not be compared with results from sources that do establish no overlap.; review method: ai-assisted-source-review; reviewed at: 2026-11-06T00:00:00Z; source id: hyinv-20261106-source-report; source locator: Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10; applies to: hyinv-20261106-protocol-sctm-boltz2; hyinv-20261106-protocol-sctm-difference-boltz2; hyinv-20261106-protocol-sctm-difference-esmfold2; hyinv-20261106-protocol-sctm-difference-protenix2; hyinv-20261106-protocol-sctm-esmfold2; hyinv-20261106-protocol-sctm-protenix2; artifact sha256: 0f10b390be399c341d1f7497c5a3ca0ac857ce78172c4732bdc86a9c66e46419; message: The ten baselines are single published settings at one temperature each, so their design-set diversity differs from HyInv's: on CATH4.3 test HyInv samples at diversity 0.331 against 0.093 for PiFold, 0.104 for CMLM, 0.144 for LM-Design 650M and 0.000 for DPLM-2 3B, which decodes by argmax. Appendix D.2 states that these four differences are not at matched diversity, and Appendix A states that even for the closest baselines the comparisons are 'close to, but not exactly at, equal diversity', the largest offset of 0.016 being worth at most 0.0011 scTM at the development slope. The scTM and scTM difference results of the protocols named here must not be read as a matched-diversity ranking. The results read at a stated diversity level (Tables 8 and 9 and Appendix D.3) are not covered by this entry, because they are the ones the report reads on a fitted curve at a fixed diversity; nor are the recovery, diversity, pLDDT, scRMSD and lDDT results, whose protocols carry the same single-setting caveat in their own limitations and whose metric qualifiers say the value is at each method's own setting.; review method: ai-assisted-source-review; reviewed at: 2026-11-06T00:00:00Z; source id: hyinv-20261106-source-report; source locator: Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'; applies to: hyinv-20261106-protocol-design-latency-h20; hyinv-20261106-protocol-design-speedup-vs-proteinmpnn-h20; hyinv-20261106-protocol-sctm-fixed-diversity-bf16-minus-reference; artifact sha256: 0f10b390be399c341d1f7497c5a3ca0ac857ce78172c4732bdc86a9c66e46419; message: The inference-time results were measured with the supervised-stage weights and a rewrite temperature of 0.23, not with the released checkpoint and its default temperatures, and the quality read-outs of the two numerical paths used the RL checkpoint at update 288, not update 640. Speed and quality in this report are therefore measured on different weights, and the 57x figure compares HyInv's optimised BF16 path with ProteinMPNN's released implementation rather than with an equally optimised baseline.; review method: ai-assisted-source-review; reviewed at: 2026-11-06T00:00:00Z; source id: hyinv-20261106-source-report; source locator: Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
- extraction method
- pdftotext -layout text layer, parsed by extract/parse.py with every row label and column count asserted.
- licence
- Apache-2.0 (repository LICENSE)
- limitations
- Author-released report, not peer reviewed and not independently refereed. All three authors are at Tencent Hunyuan, which released the model, and the baselines were run by the same team.; Every quantity is computed on predicted structures and designed sequences. Appendix B states that none is an experimental measurement, so no binding, stability or activity is measured anywhere in the report.; Appendix C states: 'We do not claim zero sequence overlap between the training data and the evaluation panels.' The RL pool was filtered against the panels with a same-source homology mask and the supervised corpus by exact-sequence removal and a homology filter, which is weaker than no overlap.; The ten baselines are each at one published single-temperature setting and were not tuned to the panels, so the scTM comparisons are at comparable rather than equal sequence diversity. Appendix D.2 states that PiFold, CMLM, LM-Design and DPLM-2 3B are single points at much lower diversity than HyInv.; ESMFold2 is both the primary scoring predictor and the reinforcement-learning reward, so the five ESMFold2 panels favour the trained model by construction; only RecentPDB is also folded by Boltz2 and Protenix2.; Some baseline ESMFold2 cells were regenerated for the report rather than taken from the released scores (Table 5), and the ProteinMPNN temperatures other than the published one were generated for the report.; Appendix A records run-to-run noise of about +/-0.0015 scTM per panel and says differences between checkpoints below about 0.003 are not resolvable; intervals are not adjusted for multiple comparisons.; Latencies were measured on one NVIDIA H20 and exclude model loading, CUDA graph capture and kernel compilation. Baselines ran as released, unbatched where the released adapter handles one structure per call, so Appendix F states the speed-ups are against those implementations and not against the best achievable speed of the baseline models.; Figures 1, 5, 6, 7, 8, 9b, 11c, 13, 16 and 17 carry results that are not printed as labelled cells, or whose bar order cannot be recovered from the text layer; they are recorded as gaps in notes.md, not read by eye.
- media type
- application/pdf
- publication status
- official_project_source
- retrieved at
- 2026-11-06T09:52:00Z
- source type
- vendor technical report
- url
- https://github.com/Tencent-Hunyuan/Hyra-results/blob/0dcc533cc9f8f76c90ecbfe944df816b6672328a/AI4Science/hyinv/HyInv_Technical_Report.pdf
- version
- Dated 2026-10-14; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (21 pages)
- version note
- Page 1 of the report prints the date 2026-10-14, which is the date issue #232 records and the one kept in version. The pinned PDF's own metadata gives CreationDate and ModDate D:20261030105208+08'00' (2026-10-30 02:52 UTC), produced by pdfTeX-1.40.25, so the bytes that carry artifact_sha256 were typeset sixteen days after the printed date. No earlier PDF of this report is published at or before the pinned commit, so the two dates cannot be compared against another version of the bytes, and nothing in the batch rests on either date beyond naming the report.
Related records
- companion of: HyInv LICENSE (MIT, commit 0dcc533c)
- companion of: HyInv package README (Tencent Hunyuan, commit 0dcc533c)
- companion of: Hyra Results repository README (Tencent Hunyuan, commit 0dcc533c)