Organizations: Department of Computer and Information Science, University of Pennsylvania · College of Arts and Sciences, University of Pennsylvania · Department of Bioengineering, University of Pennsylvania
Molecular conjugates, including PROTACs and peptide-drug conjugates (PDCs), derive their function from the joint behavior of multiple coupled components, yet most generative approaches design these components independently and combine them only after generation. Such staged pipelines ignore cross-component dependencies and often produce conjugates that are chemically invalid, fall outside empirical conjugate distributions, or lose function upon assembly. In this work, we introduce Coupled Objective-Guided Discrete Flows for Molecular Conjugate Assembly (CupOFMoCA), a discrete generative method that enforces cross-component chemical constraints during conjugate generation. CupOFMoCA restricts guided updates to chemically feasible conjugate states and reweights local transitions using molecular design objectives. Across three PDC systems, coupled generation achieved 100% conjugation validity among valid peptides and increased valid yield over matched post-hoc filtering, with comparable predicted binding scores across the full valid sets. In three-seed evaluations of three PROTAC systems, CupOFMoCA achieved the highest assembly validity on two systems and generated more assembly-valid molecules per second than the evaluated baselines. We further evaluated the generated PROTACs with post-hoc DegradeMaster scoring and explicit-solvent molecular dynamics, supporting their predicted activity and conformational properties. Together, these results establish CupOFMoCA as an efficient generator of molecular conjugates, supporting therapeutic development via coordinated control of assembly and molecular properties.
Figures & tables
Figure 1: Overview of CupOFMoCA for molecular conjugate generation. (A) A pre-trained discrete flow model is guided from unassembled components to valid assembled conjugates, including peptide-drug conjugates (PDCs) and PROTACs, under objective guidance and hard validity constraints (attachment compatibility, linker feasibility, and representational validity). (B) Objective-guided sampling with hard validity constraints. We reweight candidate probabilities with objective scores and exclude invalid updates with a validity mask. The illustrated PDC acceptance rule retains updates that preserve or increase the score. For PROTACs, we accept valid tilted proposals and return the highest-scoring valid state encountered.
ERBB2
GNRHR
NRP1
Method
Conj. Val.
Binding
Conj. Val.
Binding
Conj. Val.
Binding
moPPIt
7.1%
7.59 (0.07)
8.1%
6.78 (0.46)
65.0%
7.94 (0.06)
BoltzGen
Exp
51.0%
7.81 (0.05)
16.0%
6.63 (1.05)
-
-
AF
21.2%
7.82 (0.11)
3.0%
7.44 (0.45)
13.7%
7.05 (1.81)
AReUReDi
Table 1: Top-20 de novo PDCs across three receptor targets generated by AA-based methods (peptide length: 10 AAs). Each conjugate comprises a peptide linked to a target-specific cytotoxic payload via the corresponding linker (see Section F ). Conjugation Validity: fraction of valid peptides yielding chemically valid conjugates. Binding affinity: predicted peptide-receptor binding score (pKd). Values are mean (std). Bold : best per column; underline : second-best.
ERBB2
GNRHR
NRP1
Method
Conj. Val.
Binding
Conj. Val.
Binding
Conj. Val.
Binding
PepTune
20.8%
6.75 (0.27)
5.6%
4.55
13.7%
7.05 (1.81)
AReUReDi
Staged
30.1%
7.73 (0.11)
26.8%
7.69 (0.75)
23.0%
8.62 (0.50)
Guided
100.0%
8.02 (0.19)
100.0%
8.56 (0.35)
100.0%
9.16 (0.36)
CupOFMoCA
Table 2: Top-20 de novo PDCs across three receptor targets generated by SMILES-based methods (peptide length: 80 SMILES tokens). Each conjugate comprises a peptide linked to a target-specific cytotoxic payload via the corresponding linker.
EGFR-VHL
BRD9-VHL
CDK6-CRBN
Method
Std. Val.
Asm. Val.
Std. Val.
Asm. Val.
Std. Val.
Asm. Val.
LinkerNet
98.0
46.0
100.0
16.0
100.0
96.0
DiffPROTACs
98.0
80.0
100.0
0.0
86.0
74.0
DiffLinker
86.0
0.0
74.0
0.0
92.0
0.0
Ours
95.3 ± 2.1
85.0 ± 3.6
94.3 ± 2.1
82.3 ± 4.0
94.7 ± 1.5
83.0 ± 4.2
DegradeMaster predicted degradation probability
Table 3: Assembly validity and predicted degradation across PROTAC systems. Standard validity is the fraction of RDKit-parseable molecules; assembly validity additionally requires canonical attachment and an analyzer pass. With the selected objective weights (Appendix ), we generated 100 candidates for each of three seeds ( 42 , 123 , 456 ), obtaining 255 , 247 , and 249 assembly-valid molecules for EGFR-VHL, BRD9-VHL, and CDK6-CRBN, respectively. Validity is mean ± std across seeds; each baseline used one batch of 100 . DegradeMaster scores are mean (std) over assembly-valid molecules, and threshold columns report the fraction scoring ≥0.5 . Bold : best per column; underline : second-best. Dashes denote no valid assemblies.
EGFR-VHL
BRD9-VHL
CDK6-CRBN
Method
Div
Sim
>0.8
Div
Sim
>0.8
Div
Sim
>0.8
LinkerNet
0.895
0.816
78.3%
0.829
0.825
75.0%
0.833
0.776
45.8%
DiffPROTACs
0.763
0.779
35.0%
-
-
-
0.735
0.694
2.7%
Ours
0.772
0.630
0.0%
0.746
0.777
26.0%
0.716
0.545
0.0%
Table 4: Diversity and training-set similarity across PROTAC systems. Div: mean pairwise Tanimoto within the generated batch (lower = more diverse). Sim: mean max-Tanimoto to PROTAC-DB (lower = less similar to training data). “Sim >0.8 ” is the fraction exceeding the similarity threshold. Our rows pool three seeds ( 255 / 247 / 249 assembly-valid molecules); baselines are single batches of N=100 . Bold : best per column; underline : second-best. Dashes indicate methods that produce 0% valid assemblies.
Appendix figures & tables34 assets
Supplementary material from the paper’s appendix.
Appendix
Length
Configuration
ERBB2
GNRHR
NRP1
80
Unconditional
22.8 / 6.12
22.8 / 4.48
6.1 / 5.20
Staged
22.7 / 7.70
18.3 / 8.09
23.7 / 8.68
Gate-only
100.0 / 7.33
100.0 / 6.59
100.0 / 7.88
Guided
100.0 / 7.69
100.0 / 7.55
100.0 / 8.58
100
Unconditional
24.5 / 6.22
24.5 / 4.93
5.0 / 5.47
Staged
36.7 / 7.92
34.7 / 8.15
28.1 / 8.65
Appendix
Table A1 : PDC validity-gate and affinity-guidance ablation. Each cell reports conjugation validity (%) / mean predicted pKd over the full valid set. The pKd predictor also supplies the affinity-guidance score and therefore evaluates the optimized objective.
Group
n
Mean NLL (SEM)
Δ NLL
95% CI
PDCdb, all
393
1.5052 (0.0090)
−0.0172
[ −0.0360 , 0.0017]
PDCdb, CPP-like
149
1.4630 (0.0138)
−0.0594
[ −0.0874 , −0.0313 ]
PDCdb, non-CPP
244
1.5349 (0.0118)
0.0124
[ −0.0115 , 0.0364]
Pretraining validation
1,414
1.5224 (0.0039)
reference
Appendix
Table A2 : Zero-shot NLL under base PepReDi. SEM is computed across sequences; confidence intervals for the NLL difference use 5,000 sequence-level bootstrap samples.
Method
Best mean affinity
Best valid yield/GPU-hour
AReUReDi-Staged
3
1
CupOFMoCA-Staged
8
3
AReUReDi-Guided
1
0
CupOFMoCA-Guided
0
8
Appendix
Table A3 : Number of wins across 12 target-representation blocks. Valid yield/GPU-hour counts chemically valid conjugates produced per unit compute.
Figure A1 : Joint validity-binding tradeoff for de novo PDC generation across three receptor targets. Each point shows one method’s conjugate validity (x-axis, fraction of valid peptides that yield a valid conjugate) against the mean predicted binding affinity of its top-20 PDCs (y-axis, pKd, error bars: ±1 std). Method family is encoded by colour (CupOFMoCA Staged: tan; CupOFMoCA Guided: dark taupe; AReUReDi Staged/Guided: light/dark green; baselines: blue/grey/lavender; pharmaceutical reference: gold). Marker shape encodes the input/output representation (square: amino-acid sequence model; diamond: unnatural-AA model; circle: SMILES staged generation, including PepTune; triangle: SMILES guided/coupled generation; star: pharma reference). Guided methods achieved high conjugation validity across all three targets, with representation-dependent predicted binding scores. SMILES-guided samples are triangles; guided AA and UAA samples are squares and diamonds, respectively. Pharma validity is set to 100% by definition (well-formed reference compounds).
Figure A2 : Decomposition of 100 generation runs by outcome class for SMILES-based methods at peptide length 80. For each method × target, the bar splits the 100 runs into three categories: invalid peptide (red, generation failed at the SMILES level), valid peptide that fails post-hoc conjugation (orange, lacks the chemical features required for linker attachment), and valid PDC (green, full success). Coupled methods (AReUReDi Guided, CupOFMoCA Guided) eliminate the orange band almost entirely on ERBB2 and GNRHR by generating conjugation-compatible peptides directly, recovering 92-94% valid PDCs vs. 17-28% for staged baselines and 3-10% for PepTune. On NRP1, coupled methods yielded 67-78% valid PDCs across all generation attempts, compared with 7-23% for staged methods.
ERBB2
GNRHR
NRP1
Method
Conj. Val.
Binding
Conj. Val.
Binding
Conj. Val.
Binding
PepTune
17.9%
6.82 (0.50)
17.2%
5.85 (1.25)
11.3%
7.25(1.11)
AReUReDi
Staged
40.9%
7.94 (0.16)
39.1%
9.15 (0.35)
18.8%
8.75 (0.64)
Guided
100.0%
7.97 (0.20)
100.0%
8.86 (0.32)
100.0%
9.23 (0.24)
CupOFMoCA
Appendix
Table A4 : Top-20 de novo PDCs across three receptor targets generated by SMILES-based methods (peptide length: 100 SMILES tokens). Each conjugate comprises a peptide linked to a target-specific cytotoxic payload via the corresponding linker. Conjugation Validity: fraction of valid peptides yielding chemically valid conjugates. Binding affinity: predicted peptide-receptor binding score (pKd). Values are mean (std). Bold : best per column; underline : second-best.
ERBB2
GNRHR
NRP1
Method
Conj. Val.
Binding
Conj. Val.
Binding
Conj. Val.
Binding
moPPIt
7.1%
7.59 (0.07)
8.1%
6.78 (0.46)
65.0%
7.73 (0.34)
BoltzGen
Exp
51.0%
7.64 (0.22)
16.0%
6.63 (1.05)
-
-
AF
21.2%
7.80 (0.13)
3.0%
7.44 (0.45)
13.7%
7.05 (1.81)
AReUReDi
Appendix
Table A5 : Full-valid-set de novo PDCs across three receptor targets generated by AA-based methods (peptide length: 10 AAs). Metrics are computed over all valid conjugates from each method (not filtered to top-20). Each conjugate comprises a peptide linked to a target-specific cytotoxic payload via the corresponding linker. Conjugation Validity: fraction of valid peptides yielding chemically valid conjugates. Binding affinity: predicted peptide-receptor binding score (pKd). Values are mean (std). Bold : best per column; underline : second-best.
ERBB2
GNRHR
NRP1
Method
Conj. Val.
Binding
Conj. Val.
Binding
Conj. Val.
Binding
PepTune
20.8%
6.75 (0.27)
5.6%
4.55
13.7%
7.05 (1.81)
AReUReDi
Staged
30.1%
7.65 (0.16)
26.8%
7.54 (0.86)
23.0%
8.62 (0.50)
Guided
100.0%
7.68 (0.25)
100.0%
7.64 (0.68)
100.0%
8.59 (0.50)
CupOFMoCA
Appendix
Table A6 : Full-valid-set de novo PDCs across three receptor targets generated by SMILES-based methods (peptide length: 80 SMILES tokens). Metrics are computed over all valid conjugates from each method (not filtered to top-20). Each conjugate comprises a peptide linked to a target-specific cytotoxic payload via the corresponding linker. Conjugation Validity: fraction of valid peptides yielding chemically valid conjugates. Binding affinity: predicted peptide-receptor binding score (pKd). Values are mean (std). Bold : best per column; underline : second-best.
ERBB2
GNRHR
NRP1
Method
Conj. Val.
Binding
Conj. Val.
Binding
Conj. Val.
Binding
PepTune
17.9%
6.82 (0.50)
17.2%
5.85 (1.25)
11.3%
7.25 (1.11)
AReUReDi
Staged
40.9%
7.79 (0.21)
39.1%
8.57 (0.93)
18.8%
8.75 (0.64)
Guided
100.0%
7.69 (0.21)
100.0%
7.95 (0.63)
100.0%
8.66 (0.66)
CupOFMoCA
Appendix
Table A7 : Full-valid-set de novo PDCs across three receptor targets generated by SMILES-based methods (peptide length: 100 SMILES tokens). Metrics are computed over all valid conjugates from each method (not filtered to top-20). Each conjugate comprises a peptide linked to a target-specific cytotoxic payload via the corresponding linker. Conjugation Validity: fraction of valid peptides yielding chemically valid conjugates. Binding affinity: predicted peptide-receptor binding score (pKd). Values are mean (std). Bold : best per column; underline : second-best.
Target
Linker-payload (reaction)
Theoretical
Pharma
Generated
ΔMW
ΔMW ( n )
ΔMW ( n )
ERBB2
SSCpt, ester/carbonate (amide)
537.10 / 553.10
538.8 (1)
538.6±0.64 (92)
GNRHR
AoaDau (amide)
582.18
583.5±0.32 (4)
583.7±0.48 (94)
NRP1
DSPASN38, pos10/pos20 (disulfide)
478.12
479.0±0.10 (4)
479.6±0.41 (78)
Appendix
Table A8 : Per-attachment mass added by each receptor-specific linker-payload is conserved between PDCdb references and generated PDCs, confirming identical conjugation chemistry. ΔMW is computed as conjugate MW − peptide MW. Theoretical values: SSCpt (amide, − H 2 O) ester/carbonate, AoaDau (amide, − H 2 O), DSPASN38 (disulfide, −2 H).
Target
Pharma reference
Generated (80 SMILES tokens)
Direction of offset
peptide MW (Da)
seq. length
peptide MW (Da)
ERBB2
1852 ( n=1 )
18
1531 (1205-1986)
generated lighter than pharma
GNRHR
1559 (1072-2300)
13.2 (8-21)
1544 (1218-1907)
matched (within ∼15 Da)
NRP1
744 (577-974)
6.8 (5-10)
1396 (1123-1671)
generated heavier than pharma
Appendix
Table A9 : Peptide-length comparison between PDCdb references and generated PDCs. With the linker-payload mass held constant (Table A8 ), the apparent offsets in Figure A3 reduce to peptide-mass differences. Generated peptides are sampled at a fixed budget of 80 SMILES tokens and frequently use non-canonical residues, so a residue count is not directly comparable to the canonical pharma peptides; we report peptide MW (Da) for both sides and seq. length only for the pharma references.
Figure A3 : Molecular property distributions of generated PDCs vs. PDCdb pharmaceutical references at peptide length 80 SMILES tokens. Box plots show the distributions of molecular weight (MW), lipophilicity (LogP), topological polar surface area (TPSA, Å 2 ), and heavy-atom count (HA) over all valid conjugates from each SMILES-based method ( n=67−94 per cell for AReUReDi/CupOFMoCA Guided; n=17−28 for staged variants), compared against PDCdb reference compounds (gold, n=1 for ERBB2; n=4 for GNRHR; n=4 for NRP1). For ERBB2, the single available reference appears as a degenerate line. All methods use the same receptor-specific linker-payload as the references (ERBB2: SSCpt; GNRHR: AoaDau; NRP1: DSPASN38), and we verified that the per-attachment ΔMW (conjugate − peptide) is conserved between pharma and generated sets within ∼1 Da, confirming identical conjugation chemistry on both sides. Differences between gold and generated boxes therefore reflect peptide-length differences rather than chemistry: the GNRHR pharma references are 8-21-residue peptides (mean peptide MW ≈1.56 kDa) and overlap the generated distribution; the single ERBB2 reference is an 18-residue peptide ( ≈1.85 kDa) and sits above the generations; the NRP1 references are short 5-9-residue C-end-rule binders (e.g. CRGDK; mean peptide MW ≈0.74 kDa) and sit below. Generated distributions across methods are highly overlapping, indicating that coupled / guided generation does not collapse the chemistry to a narrow region of property space.
BoltzGen
AReUReDi
CupOFMoCA
Property
moPPIt
Exp
AF
Staged (AA)
Guided (AA)
Staged (UAA)
Guided (UAA)
Staged (AA)
Guided (AA)
Staged (UAA)
Guided (UAA)
Conj. Validity
7.1%
51.0%
21.2%
17.0%
100.0%
3.1%
100.0%
15.0%
100.0%
5.1%
100.0%
Binding affinity
7.59 (0.07)
7.64 (0.22)
7.80 (0.13)
7.65 (0.10)
7.57 (0.16)
7.68 (0.08)
7.72 (0.08)
7.56 (0.19)
7.56 (0.14)
7.77 (0.05)
7.71 (0.09)
MW (Da)
1797.19 (94.33)
1862.70 (262.50)
1652.95 (339.07)
1958.14 (104.89)
1987.75 (110.63)
2460.77 (288.16)
2558.94 (454.66)
2062.10 (90.96)
2037.29 (98.27)
2238.06 (380.48)
2465.91 (404.95)
LogP
2.23 (1.93)
−0.84 (2.37)
−1.03 (2.89)
0.93 (1.94)
0.53 (2.54)
8.93 (2.53)
8.96 (5.73)
0.89 (2.23)
0.52 (1.99)
4.59 (3.27)
6.65 (5.25)
TPSA (Å 2 )
556.56 (30.58)
663.67 (120.94)
571.81 (171.79)
648.83 (78.14)
671.25 (83.64)
668.58 (64.45)
723.24 (97.92)
708.49 (84.67)
698.27 (78.00)
670.06 (82.08)
724.58 (111.19)
Appendix
Table A10 : Full-valid-set molecular properties for AA-based methods on target ERBB2 (peptide length: 10 AAs). All conjugates comprise a peptide linked to a target-specific cytotoxic payload via the corresponding linker). Conjugation Validity: fraction of valid peptides yielding chemically valid conjugates. Binding affinity: predicted peptide-receptor binding score (pKd). MW: molecular weight; HBD/HBA: hydrogen-bond donors/acceptors. Values are mean (std).
BoltzGen
AReUReDi
CupOFMoCA
Property
moPPIt
Exp
AF
Staged (AA)
Guided (AA)
Staged (UAA)
Guided (UAA)
Staged (AA)
Guided (AA)
Staged (UAA)
Guided (UAA)
Conj. Validity
8.1%
16.0%
3.0%
17.6%
100.0%
5.1%
100.0%
15.0%
100.0%
4.0%
100.0%
Binding affinity
6.78 (0.46)
6.63 (1.05)
7.44 (0.45)
6.58 (0.42)
6.48 (0.49)
7.68 (0.33)
7.39 (0.47)
6.30 (0.35)
6.40 (0.50)
6.98 (0.84)
7.57 (0.45)
MW (Da)
1807.91 (122.97)
1914.95 (267.16)
1629.48 (178.58)
2057.01 (112.57)
2044.31 (106.20)
2431.32 (446.81)
2506.25 (395.02)
2081.44 (99.03)
2074.30 (110.34)
2852.63 (370.48)
2650.75 (445.82)
LogP
−0.08 (3.66)
−3.07 (3.00)
0.36 (1.88)
−2.58 (1.74)
−2.51 (1.40)
2.62 (3.27)
4.77 (6.72)
−1.87 (1.42)
−2.39 (1.79)
8.73 (11.32)
7.00 (8.38)
TPSA (Å 2 )
630.78 (69.49)
712.12 (138.16)
582.29 (78.16)
794.03 (91.01)
782.39 (73.53)
835.99 (164.41)
806.31 (98.75)
787.20 (74.23)
795.54 (86.98)
836.49 (137.88)
802.97 (104.47)
Appendix
Table A11 : Full-valid-set molecular properties for AA-based methods on target GNRHR (peptide length: 10 AAs). All conjugates comprise a peptide linked to a target-specific cytotoxic payload via the corresponding linker. Conjugation Validity: fraction of valid peptides yielding chemically valid conjugates. Binding affinity: predicted peptide-receptor binding score (pKd). MW: molecular weight; HBD/HBA: hydrogen-bond donors/acceptors. Values are mean (std).
AReUReDi
CupOFMoCA
Property
moPPIt
BoltzGen AF
Staged (AA)
Guided (AA)
Staged (UAA)
Guided (UAA)
Staged (AA)
Guided (AA)
Staged (UAA)
Guided (UAA)
Conj. Validity
65.0%
13.7%
19.0%
98.0%
37.4%
100.0%
25.0%
100.0%
19.4%
100.0%
Binding affinity
7.73 (0.34)
7.05 (1.81)
8.23 (0.38)
8.24 (0.37)
7.92 (0.32)
7.94 (0.27)
8.16 (0.32)
8.28 (0.42)
8.02 (0.40)
7.95 (0.22)
MW (Da)
1754.71 (96.21)
1680.54 (141.90)
1838.20 (115.56)
1842.31 (82.55)
2431.86 (271.82)
2325.85 (252.47)
1854.46 (94.83)
1859.13 (87.62)
2269.39 (317.96)
2244.83 (296.12)
LogP
4.11 (2.09)
2.22 (5.38)
2.87 (2.07)
3.00 (2.07)
10.74 (4.53)
8.75 (3.94)
3.73 (2.07)
3.40 (2.16)
10.37 (4.20)
8.23 (4.44)
TPSA (Å 2 )
525.45 (55.18)
518.17 (130.08)
566.37 (70.85)
566.61 (55.68)
626.38 (64.34)
625.54 (78.18)
556.44 (65.69)
566.99 (72.48)
569.27 (61.28)
596.79 (61.02)
Appendix
Table A12 : Full-valid-set molecular properties for AA-based methods on target NRP1 (peptide length: 10 AAs). All conjugates comprise a peptide linked to SN-38 payload via Disulfanylpropanoic Acid linker. Conjugation Validity: fraction of valid peptides yielding chemically valid conjugates. Binding affinity: predicted peptide-receptor binding score (pKd). MW: molecular weight; HBD/HBA: hydrogen-bond donors/acceptors. Values are mean (std).
Table A19 : Molecular properties of PDCs generated with variable vs. fixed linkers (GNRHR-Dau: fixed Aoa; ERBB2-Cpt: fixed SS) for the same receptor-payload pairs. All results are from 100 generation runs at peptide length 80 SMILES tokens. Values shown as mean (std).
Table A20 : Linker type diversity for the two receptor-payload pairs. DB refers to the PDCdb database distribution of linker types. Joint stands for generation with randomly sampled linker type drawn from the implemented linker types’ frequencies in PDCdb.
Receptor
Variant
Length
Validity
Uniqueness
Diversity
SNN
ERBB2
Coupled
80
0.919
1.000
0.844
0.187
100
0.949
1.000
0.835
0.194
Var. linker
80
0.919
1.000
0.842
0.188
100
0.960
1.000
0.835
0.194
GNRHR
Coupled
80
0.939
1.000
0.843
0.180
100
0.960
1.000
0.832
0.187
Appendix
Table A21 : Generation diversity metrics across receptors, peptide lengths, and linker configurations. Validity: fraction of generated peptides forming valid conjugates. Uniqueness: fraction of distinct samples among generated. Diversity: average pairwise Tanimoto distance. SNN: similarity to nearest neighbor in training set (lower = more novel).
Metric
Peptide
Conjugate
Pharma
Ours
Pharma
Ours
Conf.
0.6653
0.6878
0.6415
0.6971
ipTM
0.7217
0.7571
0.6281
0.8043
pLDDT
0.6511
0.6705
0.6448
0.6703
ipLDDT
0.4344
0.4976
0.4021
0.5023
Appendix
Table A22 : Structure prediction quality for top 10 ERBB2-targeting designs (length 80). Comparison of peptides and PDCs binding to ERBB2 receptor, with metrics measured on the receptor-ligand complex using Boltz-2. Pharma represents reference pharmaceutical designs, while Coupled represents our generation method. conf. = confidence score, ipTM = interface predicted TM-score, pLDDT = complex pLDDT, ipLDDT = complex ipLDDT.
Metric
Peptide
Conjugate
Pharma
Ours
Pharma
Ours
Conf.
0.8037
0.7764
0.7111
0.7251
ipTM
0.9217
0.8826
0.8581
0.8695
pLDDT
0.7742
0.7498
0.6744
0.6890
ipLDDT
0.6522
0.5827
0.4639
0.4961
Appendix
Table A23 : Structure prediction quality for top 10 GNRHR-targeting designs (length 80). Comparison of peptides and PDCs binding to GNRHR receptor, with metrics measured on the receptor-ligand complex using Boltz-2.
Metric
Peptide
Conjugate
Pharma
Ours
Pharma
Ours
Conf.
0.7314
0.6545
0.6560
0.6356
ipTM
0.6673
0.3914
0.4359
0.3486
pLDDT
0.7474
0.7203
0.7111
0.7074
ipLDDT
0.6563
0.4421
0.4547
0.4036
Appendix
Table A24 : Structure prediction quality for top 10 NRP1-targeting designs (length 80). Comparison of peptides and PDCs binding to NRP1 receptor, with metrics measured on the receptor-ligand complex using Boltz-2.
Figure B1 : Top three PROTACs generated for each target. For each target system (BRD9-VHL, EGFR-VHL, CDK6-CRBN), the three PROTACs with the lowest synthetic-accessibility (SA) score among the top five by DegradeMaster (DM) probability are shown horizontally. Each panel depicts the assembled conjugate with the de novo -generated linker highlighted in blue; the warhead and E3-ligand fragments are the fixed inputs. DM denotes the predicted degradation-activity probability (higher is better) and SA the synthetic-accessibility score (lower is better, range 1-10).
Target
Property
LinkerNet
DiffPROTACs
Ours
EGFR-VHL
SA score
1.89 ± 0.46
2.95 ± 0.91
2.76 ± 0.57
QED
0.487 ± 0.072
0.424 ± 0.133
0.51 ± 0.13
logP in range
87.0%
75.0%
70.3%
MW in range
100.0%
80.0%
71.4%
RotB in range
100.0%
67.5%
84.5%
Uniqueness
65.2%
97.5%
96.2%
Appendix
Table B1 : Molecular properties of valid PROTACs across three target systems (EGFR-VHL, BRD9-VHL, CDK6-CRBN). SA ↓ , all others ↑ . SA and QED reported as mean ± std across molecules. The logP , MW, and RotB rows report the fraction of valid PROTACs falling within PROTAC-DB-derived windows ( logP∈[2.5,7.4] , MW∈[771,1108] Da, RotB∈[12,25] ); note that Table uses a wider MW window ( 700 - 1200 Da) and is not directly comparable on that row. Our column pools three seeds ( 286 / 283 / 284 RDKit-valid linkers) under the optimized objective weights; baseline columns are single batches of N=100 . Uniqueness is the per-seed mean; pooled over all three seeds it is 92.7 / 92.9 / 91.9% . Dashes (-) indicate 0% assembly validity on that system.
Linker
Role
3D PSA (Å 2 ) ↓
Penalty ↓
C1
n -propylsuccinyl
Generated
140
2.2
C2
allyl succinyl
Generated
141
1.9
C3
gem-dimethylsuccinate
Generated
178
2.7
C4
butylmalonate
Generated
184
3.3
MS39
decanedioyl
Positive
178
3.7
PEG4
PEG4
Negative
193
4.1
Appendix
Table B2 : Explicit-solvent MD for generated EGFR-VHL PROTACs against experimental controls. MS39 is a validated cell-active EGFR-VHL degrader [ Cheng et al., 2020 ] ; PEG4 is a PEG4-linked congener (negative control). Ternary penalty in kcal/mol.
Parameter
Value
Hardware
4 × A6000 NVIDIA RTX 6000 Ada GPU (48 GB VRAM)
Training epochs
5
Optimizer
Adam
Learning rate
1×10−4
Learning rate schedule
Cosine decay with 10% warmup
Minimum learning rate
1×10−5
Appendix
Table E1 : SMILESReDi training configuration.
Component
Configuration
Input
Unpooled token-level embeddings from ESM-2 650M
Convolutional layers
3 initial CNN kernel layers (dimension 384)
Cross-attention layers
4 layers (dimension 2048)
Prediction head
Shared prediction head (dimension 1024)
Appendix
Table E2 : Binding affinity model architecture.
Parameter
Value
Training epochs
50
Learning rate
3.84×10−5
Dropout rate
0.15
CNN layers
3 initial layers (dimension 384)
Cross-attention layers
4 layers (dimension 2048)
Prediction head dimension
1024
Appendix
Table E3 : Binding affinity model training configuration.
Molecular generation develops computational methods to create chemical structures and conformations for discovery, optimization, and structure elucidation.
Biomolecular therapeutics often start from known sequences and require targeted editing to improve multiple properties while satisfying hard biochemical and manufacturability constraints. However, existing generative methods do not jointly support multi-objective optimization, hard feasibility, and sequence editing in discrete, variable-length biological spaces. In this work, we introduce Pareto-Constrained Molecule Editing (pCoMole), a framework built on discrete flow matching that steers a pre-trained Edit Flow toward user-specified preferences while enforcing terminal feasibility. pCoMole defines a feasibility-gated terminal distribution using an augmented Tchebycheff utility and realizes the resulting preference tilt through a Doob-h transform of the underlying edit process. To make this construction practical, we approximate the required harmonic function using short Monte Carlo rollouts over candidate edits, yielding an efficient guided editor with provable preference consistency. We validate pCoMole by shrinking GFP while retaining fluorescence-related properties, shortening diverse Cas9 orthologs while preserving PAM specificity, and compressing peptide binders into short peptidomimetics that optimize seven drug-related properties under hard constraints. In wet lab testing, two 229-residue pCoMole-designed eGFP variants retained clear green fluorescence in BL21 cells after 10 deletions, with either one or two substitutions. Together, pCoMole enables constraint-aware, Pareto-aligned editing of biomolecular sequences in discrete, variable-length spaces.
Tong Chen, Maximilian Holsman, Lin Zhao +1
Department of Computer and Information Science, University of Pennsylvania · Department of Computer Science, Duke University · Department of Bioengineering, University of Pennsylvania
Adapting generative foundation models, in particular diffusion and flow models, to optimize given reward functions (e.g., binding affinity) while satisfying constraints (e.g., molecular synthesizability) is fundamental for their adoption in real-world scientific discovery applications such as molecular design or protein engineering. While recent works have introduced scalable methods for reward-guided fine-tuning of such models via reinforcement learning and control schemes, it remains an open problem how to algorithmically trade-off reward maximization and constraint satisfaction in a reliable and predictable manner. Motivated by this challenge, we first present a rigorous framework for Constrained Generative Optimization, which brings an optimization viewpoint to the introduced adaptation problem and retrieves the relevant task of constrained generation as a sub-case. Then, we introduce Constrained Flow Optimization (CFO), an algorithm that automatically and provably balances reward maximization and constraint satisfaction by reducing the original problem to sequential fine-tuning via established, scalable methods. We provide convergence guarantees for constrained generative optimization and constrained generation via CFO. Ultimately, we present an experimental evaluation of CFO on both synthetic, yet illustrative, settings, and a molecular design task. Across these evaluations, CFO achieves consistent increases in reward while ensuring high constraint satisfaction, showcasing its practical utility for constrained generative optimization.
Sven Gutjahr, Riccardo De Santi, Luca Schaufelberger +2
Department of Computer Science, ETH Zurich · ETH AI Center · NCCR Catalysis, Switzerland +1
Target-conditioned cyclic peptide design requires coordinating interfacial side chains with macrocyclic closure. Existing end-to-end generators concurrently update geometric and chemical variables from noise, assigning residues to disordered coordinates. Meanwhile, predictor-assisted workflows often get stuck in local traps because they lack feedback in the propose-and-filter stage and ignore 3D structural proposals during sequence search. To overcome the drawbacks of these two paradigms, we propose DC-FLOW, a staged generative framework built around Delayed-Coupling Flow Matching. In principle, DC-FLOW postpones sequence and side-chain activation until a target-conditioned backbone takes shape, resolving structural ambiguity before chemical assignment. Then, it feeds delayed activations back into the backbone, achieving co-evolution of backbone geometry and emerging residue identities by coupling their flows. This mechanism enables more reliable macrocyclic closure and favorable interface packing. For predictor-assisted workflows, DC-FLOW supplies 3D proposals with in-trajectory guidance, enabling cascaded test-time scaling to prune redundant backbones and terminate unpromising refinement early. Extensive experiments show that DC-FLOW improves cyclization fidelity, diversity, and interface energetics, substantially increasing the yield of confident cyclic peptide binders under a fixed final-candidate budget. Our code is available at https://github.com/mianqiu00/DCFlow.
Yitian Wang, Minjie Cheng, Yaning Cui +3
Gaoling School of Artificial Intelligence, Renmin University of China · DP Technology · Cyclon Therapeutics +2