Configuration: LM-Design 650M (HyInv report) Protocol: Latency of a design request on one NVIDIA H20 (HyInv report) Dataset: Timing backbones for the HyInv inference-time measurements 15 s Runtime (Median wall-clock time of one design request on one NVIDIA H20, 128 candidates per request with a maximum candidate batch of 32 unless the request is the 10,000-design library; the timed region excludes model loading, CUDA graph capture and kernel compilation; the backbone length is named per result; the request is 128 designs for a 402-residue backbone)second · lower
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source Latency of a design request for lm-design-650m on one NVIDIA H20 (HyInv report)
hyinv-20261106-protocol-design-latency-h20
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Figure 4a, bar 'LM-Design 650M', printed label (page 5 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: CA-lDDT of designed sequences under ESMFold2 (HyInv report) Dataset: CATH4.2 test (1,120 chains) 0.835 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)unitless · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source LM-Design 650M on CATH4.2 test (CA-lDDT of designed sequences under ESMFold2, HyInv report)
hyinv-20261106-protocol-lddt-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Figure 15, 'CA-lDDT' block, row 'LM-Design 650M', column 'CATH4.2 test' (page 17 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: CA-lDDT of designed sequences under ESMFold2 (HyInv report) Dataset: CATH4.3 test (1,882 chains) 0.848 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)unitless · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source LM-Design 650M on CATH4.3 test (CA-lDDT of designed sequences under ESMFold2, HyInv report)
hyinv-20261106-protocol-lddt-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Figure 15, 'CA-lDDT' block, row 'LM-Design 650M', column 'CATH4.3 test' (page 17 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: CA-lDDT of designed sequences under ESMFold2 (HyInv report) Dataset: RecentPDB (121 chains from 111 recent PDB entries) 0.804 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)unitless · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source LM-Design 650M on RecentPDB (CA-lDDT of designed sequences under ESMFold2, HyInv report)
hyinv-20261106-protocol-lddt-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Figure 15, 'CA-lDDT' block, row 'LM-Design 650M', column 'RecentPDB' (page 17 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: CA-lDDT of designed sequences under ESMFold2 (HyInv report) Dataset: TS50 (50 chains) 0.841 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)unitless · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source LM-Design 650M on TS50 (CA-lDDT of designed sequences under ESMFold2, HyInv report)
hyinv-20261106-protocol-lddt-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Figure 15, 'CA-lDDT' block, row 'LM-Design 650M', column 'TS50' (page 17 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: CA-lDDT of designed sequences under ESMFold2 (HyInv report) Dataset: TS500 (500 chains) 0.881 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)unitless · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source LM-Design 650M on TS500 (CA-lDDT of designed sequences under ESMFold2, HyInv report)
hyinv-20261106-protocol-lddt-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Figure 15, 'CA-lDDT' block, row 'LM-Design 650M', column 'TS500' (page 17 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: pLDDT of ESMFold2 predictions of designed sequences (HyInv report) Dataset: CATH4.2 test (1,120 chains) 81.5 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)unitless · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source LM-Design 650M on CATH4.2 test (pLDDT of ESMFold2 predictions of designed sequences, HyInv report)
hyinv-20261106-protocol-plddt-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Figure 15, 'pLDDT' block, row 'LM-Design 650M', column 'CATH4.2 test' (page 17 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: pLDDT of ESMFold2 predictions of designed sequences (HyInv report) Dataset: CATH4.3 test (1,882 chains) 82.6 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)unitless · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source LM-Design 650M on CATH4.3 test (pLDDT of ESMFold2 predictions of designed sequences, HyInv report)
hyinv-20261106-protocol-plddt-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Figure 15, 'pLDDT' block, row 'LM-Design 650M', column 'CATH4.3 test' (page 17 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: pLDDT of ESMFold2 predictions of designed sequences (HyInv report) Dataset: RecentPDB (121 chains from 111 recent PDB entries) 80.6 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)unitless · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source LM-Design 650M on RecentPDB (pLDDT of ESMFold2 predictions of designed sequences, HyInv report)
hyinv-20261106-protocol-plddt-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Figure 15, 'pLDDT' block, row 'LM-Design 650M', column 'RecentPDB' (page 17 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: pLDDT of ESMFold2 predictions of designed sequences (HyInv report) Dataset: TS50 (50 chains) 83.1 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)unitless · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source LM-Design 650M on TS50 (pLDDT of ESMFold2 predictions of designed sequences, HyInv report)
hyinv-20261106-protocol-plddt-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Figure 15, 'pLDDT' block, row 'LM-Design 650M', column 'TS50' (page 17 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: pLDDT of ESMFold2 predictions of designed sequences (HyInv report) Dataset: TS500 (500 chains) 90 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)unitless · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source LM-Design 650M on TS500 (pLDDT of ESMFold2 predictions of designed sequences, HyInv report)
hyinv-20261106-protocol-plddt-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Figure 15, 'pLDDT' block, row 'LM-Design 650M', column 'TS500' (page 17 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: Self-consistency RMSD of designed sequences under ESMFold2 (HyInv report) Dataset: CATH4.2 test (1,120 chains) 4.66 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)angstrom · lower
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source LM-Design 650M on CATH4.2 test (Self-consistency RMSD of designed sequences under ESMFold2, HyInv report)
hyinv-20261106-protocol-scrmsd-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Figure 15, 'scRMSD' block, row 'LM-Design 650M', column 'CATH4.2 test' (page 17 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: Self-consistency RMSD of designed sequences under ESMFold2 (HyInv report) Dataset: CATH4.3 test (1,882 chains) 4.59 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)angstrom · lower
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source LM-Design 650M on CATH4.3 test (Self-consistency RMSD of designed sequences under ESMFold2, HyInv report)
hyinv-20261106-protocol-scrmsd-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Figure 15, 'scRMSD' block, row 'LM-Design 650M', column 'CATH4.3 test' (page 17 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: Self-consistency RMSD of designed sequences under ESMFold2 (HyInv report) Dataset: RecentPDB (121 chains from 111 recent PDB entries) 6.37 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)angstrom · lower
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source LM-Design 650M on RecentPDB (Self-consistency RMSD of designed sequences under ESMFold2, HyInv report)
hyinv-20261106-protocol-scrmsd-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Figure 15, 'scRMSD' block, row 'LM-Design 650M', column 'RecentPDB' (page 17 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: Self-consistency RMSD of designed sequences under ESMFold2 (HyInv report) Dataset: TS50 (50 chains) 3.46 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)angstrom · lower
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source LM-Design 650M on TS50 (Self-consistency RMSD of designed sequences under ESMFold2, HyInv report)
hyinv-20261106-protocol-scrmsd-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Figure 15, 'scRMSD' block, row 'LM-Design 650M', column 'TS50' (page 17 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: Self-consistency RMSD of designed sequences under ESMFold2 (HyInv report) Dataset: TS500 (500 chains) 4.35 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)angstrom · lower
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source LM-Design 650M on TS500 (Self-consistency RMSD of designed sequences under ESMFold2, HyInv report)
hyinv-20261106-protocol-scrmsd-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Figure 15, 'scRMSD' block, row 'LM-Design 650M', column 'TS500' (page 17 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: scTM of designed sequences under Boltz2 (HyInv report) Dataset: RecentPDB (121 chains from 111 recent PDB entries) 0.824 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)unitless · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source LM-Design 650M on RecentPDB Boltz2 (scTM of designed sequences under Boltz2, HyInv report)
hyinv-20261106-protocol-sctm-boltz2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 6, row 'LM-Design 650M', column 'RecentPDB, Boltz2' (page 16 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: Paired scTM difference from HyInv under Boltz2 (HyInv report) Dataset: RecentPDB (121 chains from 111 recent PDB entries) +49.6 [34.4, 65.7] Sctm difference (HyInv default minus the baseline, paired by chain on RecentPDB with shared folding randomness, printed in units of 10^-3 scTM with a 95% percentile interval; Boltz2 as the structure predictor; the reference system is the HyInv default setting; the baseline is LM-Design 650M)unitless · higher
Uncertainty: 95% CI 34.4 to 65.7. Printed in units of 10^-3 scTM, as the point estimate is. The interval excludes zero.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source HyInv minus LM-Design 650M on RecentPDB Boltz2 (paired scTM difference, HyInv report)
hyinv-20261106-protocol-sctm-difference-boltz2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 7, row 'LM-Design 650M', column 'RecentPDB, Boltz2' (page 16 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: Paired scTM difference from HyInv under ESMFold2 (HyInv report) Dataset: CATH4.2 test (1,120 chains) +36.2 [23.6, 54.9] Sctm difference (HyInv default minus the baseline, paired by chain on the same panel with shared folding randomness, printed in units of 10^-3 scTM with a 95% percentile interval; ESMFold2 as the structure predictor; the reference system is the HyInv default setting; the baseline is LM-Design 650M)unitless · higher
Uncertainty: 95% CI 23.6 to 54.9. Printed in units of 10^-3 scTM, as the point estimate is. The interval excludes zero.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source HyInv minus LM-Design 650M on CATH4.2 test (paired scTM difference, HyInv report)
hyinv-20261106-protocol-sctm-difference-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 7, row 'LM-Design 650M', column 'CATH4.2 test' (page 16 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: Paired scTM difference from HyInv under ESMFold2 (HyInv report) Dataset: CATH4.3 test (1,882 chains) +25.9 [11.1, 41.1] Sctm difference (HyInv default minus the baseline, paired by chain on the same panel with shared folding randomness, printed in units of 10^-3 scTM with a 95% percentile interval; ESMFold2 as the structure predictor; the reference system is the HyInv default setting; the baseline is LM-Design 650M)unitless · higher
Uncertainty: 95% CI 11.1 to 41.1. Printed in units of 10^-3 scTM, as the point estimate is. The interval excludes zero.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source HyInv minus LM-Design 650M on CATH4.3 test (paired scTM difference, HyInv report)
hyinv-20261106-protocol-sctm-difference-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 7, row 'LM-Design 650M', column 'CATH4.3 test' (page 16 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: Paired scTM difference from HyInv under ESMFold2 (HyInv report) Dataset: RecentPDB (121 chains from 111 recent PDB entries) +76.5 [55.3, 99.8] Sctm difference (HyInv default minus the baseline, paired by chain on the same panel with shared folding randomness, printed in units of 10^-3 scTM with a 95% percentile interval; ESMFold2 as the structure predictor; the reference system is the HyInv default setting; the baseline is LM-Design 650M)unitless · higher
Uncertainty: 95% CI 55.3 to 99.8. Printed in units of 10^-3 scTM, as the point estimate is. The interval excludes zero.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source HyInv minus LM-Design 650M on RecentPDB (paired scTM difference, HyInv report)
hyinv-20261106-protocol-sctm-difference-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 7, row 'LM-Design 650M', column 'RecentPDB, ESMFold2' (page 16 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: Paired scTM difference from HyInv under ESMFold2 (HyInv report) Dataset: TS50 (50 chains) +58.8 [31.3, 90.2] Sctm difference (HyInv default minus the baseline, paired by chain on the same panel with shared folding randomness, printed in units of 10^-3 scTM with a 95% percentile interval; ESMFold2 as the structure predictor; the reference system is the HyInv default setting; the baseline is LM-Design 650M)unitless · higher
Uncertainty: 95% CI 31.3 to 90.2. Printed in units of 10^-3 scTM, as the point estimate is. The interval excludes zero.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source HyInv minus LM-Design 650M on TS50 (paired scTM difference, HyInv report)
hyinv-20261106-protocol-sctm-difference-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 7, row 'LM-Design 650M', column 'TS50' (page 16 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: Paired scTM difference from HyInv under ESMFold2 (HyInv report) Dataset: TS500 (500 chains) +18.2 [12.0, 24.5] Sctm difference (HyInv default minus the baseline, paired by chain on the same panel with shared folding randomness, printed in units of 10^-3 scTM with a 95% percentile interval; ESMFold2 as the structure predictor; the reference system is the HyInv default setting; the baseline is LM-Design 650M)unitless · higher
Uncertainty: 95% CI 12.0 to 24.5. Printed in units of 10^-3 scTM, as the point estimate is. The interval excludes zero.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source HyInv minus LM-Design 650M on TS500 (paired scTM difference, HyInv report)
hyinv-20261106-protocol-sctm-difference-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 7, row 'LM-Design 650M', column 'TS500' (page 16 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: Paired scTM difference from HyInv under Protenix2 (HyInv report) Dataset: RecentPDB (121 chains from 111 recent PDB entries) +45.9 [30.7, 62.9] Sctm difference (HyInv default minus the baseline, paired by chain on RecentPDB with shared folding randomness, printed in units of 10^-3 scTM with a 95% percentile interval; Protenix2 as the structure predictor; the reference system is the HyInv default setting; the baseline is LM-Design 650M)unitless · higher
Uncertainty: 95% CI 30.7 to 62.9. Printed in units of 10^-3 scTM, as the point estimate is. The interval excludes zero.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source HyInv minus LM-Design 650M on RecentPDB Protenix2 (paired scTM difference, HyInv report)
hyinv-20261106-protocol-sctm-difference-protenix2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 7, row 'LM-Design 650M', column 'RecentPDB, Protenix2' (page 16 of 21) Configuration: LM-Design 650M (HyInv report) Protocol: scTM of designed sequences under ESMFold2 (HyInv report) Dataset: CATH4.2 test (1,120 chains) 0.819 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)unitless · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons 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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
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.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source LM-Design 650M on CATH4.2 test (scTM of designed sequences under ESMFold2, HyInv report)
hyinv-20261106-protocol-sctm-esmfold2
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 6, row 'LM-Design 650M', column 'CATH4.2 test' (page 16 of 21)