PaperlayerView exact revision on arXiv

Source-linked AI summary

PairIT: Autoregressive Transformers for Low-Data Molecule Optimization

Hannah Lawrence, Bodhi P. Vani, Ji Won Park, Natasa Tagasovska, Samuel Don Stanton, Andrew Martin Watkins, Michael Maser, Ewa M. Nowara

arXiv:2610.19123v1cs.LGcs.AIphysics.chem-ph

TL;DR

Seeded molecular optimization must improve properties while keeping candidates close to their starting molecules, including when labeled data are scarce. PairIT trains an autoregressive SMILES transformer on structurally similar molecule pairs and uses prompts for property control. The authors report that a compact model can support low-data optimization, with directed-pair pretraining and early saturation among the findings.

  • Problem

    Seeded optimization requires local molecular changes that improve properties while preserving seed similarity, but low-data settings and few evaluations of seeded biological-property optimization remain under-addressed.

    Source: §1 · p. 2

  • Method

    PairIT trains an autoregressive SMILES transformer on structurally similar seed–target pairs, with structured prompts for property, direction, and approximate change magnitude.

    Source: §1 · p. 2

  • Results

    Directed-pair pretraining outperforms pretraining on individual molecules, and useful paired-pretraining benefits emerge before a full epoch while additional pair data yield diminishing returns.

    Source: §1 · p. 3

  • Takeaways & Limitations

    PairIT combines matched-pair supervision, prompt control, optional iterative re-seeding or RL, and pretraining to support seeded optimization in low-data settings.

    Source: §5 · p. 9

    The work addresses inner-loop candidate generation rather than experimental validation, and oracle-dependent components inherit uncertainty from imperfect property models and distribution shift.

    Source: §5 · p. 10

Abstract

from arXiv · show

Generating novel molecules with desired properties is a central challenge in drug discovery. Although foundational work focused on de novo generation, practical use cases require a more complex setting: conditioning on an initial seed molecule, optimizing multiple properties, and operating in data-sparse regimes. We introduce PairIT, a flexible training framework for autoregressive transformers over SMILES sequences that enables seeded generation by learning from pairs of structurally similar molecules. By leveraging matched pairs, the model learns to generate local, chemically meaningful edits that improve a property of the seed molecule. Prompt tokens extend this framework by specifying the target property, desired direction, and approximate magnitude of change. Multiple prompt tokens can be concatenated to enable multi-property optimization at inference time. For properties where pair supervision alone is insufficient, reinforcement learning post-training can further improve optimization, providing a tunable trade-off between proximity to the seed and the optimization objective. PairIT naturally supports pretraining on a large dataset and fine-tuning on a smaller dataset, which is critical for performance in data-sparse regimes. Surprisingly, we observe diminishing returns to the scale of the pretraining dataset. We demonstrate empirically on MoleculeNet and Therapeutics Data Commons datasets that, with appropriate paired-data curation, a standard transformer model of modest size (68M parameters) can achieve robust seeded molecular optimization in low-data regimes.

1 Introduction

PairIT targets seeded molecular optimization in data-sparse settings, where candidates must improve desired properties while remaining structurally close to a known molecule. It trains an autoregressive model on directed pairs and adds prompt control, pretraining, and optional post-training strategies for this setting.

1 Introduction

  • The target setting is scarce-label inner-loop discovery: propose local edits that improve properties while preserving enough seed structure to retain biological function.Datasets in early-stage discovery often contain only hundreds to thousands of labeled molecules; local search also risks near-reconstruction and overfitting.Source: §1 · pp. 1–2
  • PairIT learns local property-improving edits by training a decoder-only autoregressive transformer on matched pairs of structurally similar molecules.Iterative re-seeding extends optimization at inference time, while RL post-training is considered when pair supervision alone is insufficient.Source: §1 · p. 2
  • Structured prompts specify property, direction, and approximate change magnitude, and multiple prompts support multi-property control without constructing a new pair set for each combination.This addresses the sparsity of pairs that simultaneously improve multiple properties in small datasets.Source: §1 · p. 3
  • Directed-pair pretraining outperforms single-molecule pretraining, with benefits emerging before one full epoch and diminishing returns from additional pair data.Repeated molecular coverage across pairs helps explain early saturation.Source: §1 · p. 3
  • PairIT is a compact 68M-parameter matched-pair autoregressive model designed for low-data seeded optimization and competitive with larger or more elaborate baselines.The paper also evaluates property improvement, novelty, and fidelity to the seed against diverse baselines under limited-data conditions.Source: §1 · p. 3

2 Related Work

Related work spans matched-pair molecular optimization, sequence and graph translation, large language models, reinforcement learning, and property-guided search. PairIT is positioned as a compact autoregressive alternative that uses structured controls and small downstream pair sets.

2 Related Work

  • Matched-pair methods are closest to PairIT, with prior work spanning non-autoregressive pair prediction, conditional density estimation, multi-objective ranking, and recursive inference.PropEn approximates the property-gradient direction; related extensions include conditional density estimation and non-domination ranking during pair construction.Source: §2 · p. 3
  • GeLLM4O-C instruction-tunes 7–8B-parameter language models on approximately 256K molecule pairs, while PairIT uses a 68M-parameter autoregressive SMILES transformer and 88–662 downstream pairs per task.PairIT uses structured control tokens rather than natural-language instructions.Source: §2 · pp. 3–4
  • Molecular optimization alternatives include reward-driven reinforcement learning, preference optimization, latent-space Bayesian search, and gradient-based steering.Property-model-guided methods could in principle be combined with PairIT, whereas vanilla PairIT learns its objective from paired data without a property oracle in the generation loop.Source: §2 · p. 4

3 Proposed Method: PairIT

PairIT frames seeded molecular generation as producing similar molecules with improved properties, using directed matched pairs to train a conditional autoregressive model over target SMILES. Its extensions support iterative refinement, reward-based post-training, and explicit property conditioning.

3 Proposed Method: PairIT

  • PairIT seeks candidates that meet a chosen similarity threshold to a seed while improving its property, training on directed matched seed–target pairs.The formulation allows discrete, continuous, or mixed-valued properties.Source: §3 · p. 4

3.1 PairIT objective and interpretation

  • The conditional next-token objective fits the distribution of local, property-improved neighbors, with pair construction encoding structural locality and improvement direction.The conditional model can represent multiple plausible improved targets for a seed.Source: §3.1 · p. 4

3.2 PairIT implementation and variants

  • PairIT uses a 68M-parameter decoder-only transformer; the tested model-size ablation found little benefit from scaling across 11M–200M parameters.Seed and target SMILES are concatenated with a <sep> token, and generation is conditioned on the seed.Source: §3.2 · p. 4
  • PairIT-Beam generates 8 candidates per seed in each of three rounds, retaining the top 4 as seeds and yielding up to 72 candidates per original seed.Vanilla PairIT samples property-improved candidates directly from the matched-pair model.Source: §3.2 · p. 5
  • When pair supervision is insufficient for property improvement, PairIT can be post-trained with a property-model reward and seed-similarity regularization.The paper uses Group Relative Policy Optimization (GRPO).Source: §3.1 · p. 4, §3.2 · p. 5

3.3 Prompt-conditioning

  • Single-property prompts specify the property identity and a signed delta bin, encoding the desired direction and approximate change magnitude.Prompting changes the conditioning sequence while retaining the same autoregressive training objective.Source: §3.3 · p. 5
  • Multiple prompt blocks can specify multiple properties; property dropout supports specifying subsets of trained properties at inference.A generic target slot can be reassigned to a downstream property during fine-tuning.Source: §3.3 · p. 5

4 Experimental Results

PairIT is evaluated on low-data molecular benchmarks using curated molecular pairs, with candidates ranked and independently scored for seeded optimization. Results show competitive single-property optimization, diminishing returns from paired pretraining, and prompt control over property-change direction and magnitude, though joint bioactivity control is less reliable.

4 Experimental Results

  • Evaluation uses 88–662 matched training pairs per MoleculeNet task and retains five candidates per seed after ranking, with final results independently scored by MiniMol.The benchmarks include BACE, BBBP, and HIV; hERG results are reported separately.Source: §4 · p. 5

4.1 Single-property optimization from pair supervision.

  • PairIT and PairIT-Beam are competitive with the strongest baselines on BACE and substantially outperform them on BBBP; on HIV, PairIT-RL reaches 82.6% VIN versus the non-learning MMP baseline’s 40.4%.Without RL, PairIT and most baselines are around or below the MMP baseline on HIV; the authors treat PairIT-RL as optional because it directly optimizes a learned property model.Source: §4.1 · pp. 6–7, Figure 2 · p. 6

4.2 Pretraining ablations and training dynamics

  • Without ZINC-250K pretraining, VIN drops substantially across all three tasks; molecule-only pretraining also underperforms directed-pair pretraining.The authors associate pretraining with learning SMILES grammar and directed pairs with learning seed-conditioned local transformations.Source: §4.2 · p. 7
  • Useful paired-pretraining signal saturates before one epoch, and larger paired datasets yield rapid initial gains followed by diminishing downstream improvement.The authors note that a pair epoch is not a full pass over unique molecules, since molecules recur across pairs.Source: §4.2 · p. 7, Figure 4 · p. 7
  • Pretraining validation loss correlates with ZINC-250K generation quality and downstream BACE VIN, so checkpoints are selected by validation loss rather than fixed training duration.A similar relationship is reported for zero-shot logP optimization on MoleculeNet.Source: §4.2 · p. 7

4.3 Prompt-conditioned property control

  • On all three MoleculeNet datasets, larger requested QED increases generally produce larger realized changes, supporting graded control despite imperfect bin matching.The evaluation fine-tunes with both pair orderings and sweeps 16 signed delta bins at inference.Source: §4.3 · p. 8, Figure 5 · p. 8
  • For two-property prompting, realized changes scale with requested changes on ZINC-250K, while MoleculeNet results exceed randomized prompts most consistently for low-correlation descriptor pairs and QED+logP.Joint control involving bioactivity is less reliable, which the authors relate to binary supervision and few examples of each joint configuration.Source: §4.3 · pp. 8–9, Table 1 · p. 8
  • With all 10 properties prompted, mean direction adherence across the nine non-activity properties is 0.739–0.773, versus 0.433–0.559 under randomized prompts; performance varies with the included objectives.Restricting prompts to smaller property subsets can improve performance.Source: §4.3 · p. 9, Table 2 · p. 9

5 Discussion

PairIT offers several controls for seeded molecular optimization, from pair-supervised edits to iterative re-seeding, reinforcement learning, and prompts. Its practical use depends on training-pair and objective design, while experimental validation and synthesizability remain outside its scope.

5 Discussion

  • PairIT-Beam iteratively re-seeds high-scoring generations, while PairIT-RL uses a property oracle but can be sensitive to its errors and reward hacking under weak similarity constraints.Prompt conditioning complements these approaches by specifying the direction and magnitude of changes, including for multi-property optimization.Source: §5 · p. 9
  • Sampling temperature changes the property-improvement versus seed-similarity trade-off without retraining, while PairIT-RL’s similarity-reward weight directly controls that balance.The distance function, paired-data construction, prompted properties, and reward function all shape the optimization problem.Source: §5 · p. 9
  • ZINC-250K pretraining improves reliable generation in low-data settings, directed-pair pretraining outperforms individual-molecule pretraining, and useful paired pretraining saturates quickly.More diverse pretraining pairs also change the locality of learned edits, making early stopping an important part of the training recipe.Source: §5 · p. 9
  • PairIT addresses the inner loop of scoring candidates but lacks experimental validation, does not address synthesizability, and requires expert review and safety screening before synthesis.Oracle-dependent components inherit uncertainty from imperfect property oracles and distribution shift; multi-property control is less reliable for conflicting or weakly represented changes.Source: §5 · p. 10

A Related work comparison

  • Table 3 compares molecular generation and optimization methods across dimensions relevant to this setting, including whether methods use iterative property-model-guided search.The table defines “property model-free” as not using iterative property model-guided search during generation.Source: A Related work comparison · p. 16, Table 3 · p. 16

A.1 Selected de novo generation related work

  • De novo property-conditioned generation spans diffusion, flow-matching, walk-jump-style methods, and SMILES language models, with property control typically supplied through activations or inputs.The cited examples include equivariant and non-equivariant approaches, as well as models using augmentation.Source: §A.1 · p. 16

B.1 Model architecture and training hyperparameters

PairIT uses regex-based SMILES pre-tokenization and a fixed WordLevel vocabulary; the appendix identifies the architecture and training hyperparameters for its 68M model.

B.1 Model architecture and training hyperparameters

  • SMILES are split into chemically meaningful units with a regex-based pre-tokenizer, then mapped to a fixed 1,225-token WordLevel vocabulary without BPE or learned merges.The vocabulary includes <cls>, <pad>, <eos>, <unk>, <mask>, and <sep>.Source: §B.1 · p. 16
  • Experiments ran on multi-GPU nodes with NVIDIA B200 GPUs, and pretraining took up to 1.5 hours per run.Source: §B.1 · p. 16
  • The appendix reports architecture and training hyperparameters for the 68M model used in the main experiments.The supplied table caption does not specify the individual settings.Source: Table 4 · p. 17

B.2 Model-size ablation

Model scaling yielded little to no generation-quality improvement beyond the medium scale, while the inference, pairing, and evaluation procedures characterize important trade-offs and scope.

B.2 Model-size ablation

  • The 68M model performs comparably to larger models, while scaling brings little to no gains across key metrics and the largest models show increased memorization.The authors use 68M as a practical default because larger models show overfitting without meaningful generation-quality benefits.Source: §B.2 · p. 17, Figure 6 · p. 18

B.3 Inference-temperature sweep

  • Inference temperatures from 0.1–1.0 trade seed similarity and validity against novelty and property improvement, with 0.5 used across datasets.Lower temperatures favor near-repetition of the seed; higher temperatures tend to improve the target property while reducing validity and fidelity.Source: §B.3 · p. 18, Figure 7 · p. 18, Figure 8 · p. 19

B.4 Pair construction details

  • ZINC-250K uses fingerprint-similar pairs directed toward higher logP, whereas MoleculeNet uses localized matched molecular pairs and filters out pairs with heavy-atom ratios below 0.6.For signed-prompt MoleculeNet models, both pair orderings are included; hERG instead uses Tanimoto-neighbor pairs because MMPs are sparse, yielding 76 usable pairs after filtering.Source: §B.4 · p. 20, Table 5 · p. 20

B.5 Evaluation metrics

  • Evaluation metrics include seed similarity, nearest-neighbor similarity, distribution distance, fragment and scaffold agreement, and reference-set coverage.Higher SimSeed indicates greater seed fidelity; lower FCD indicates closer reference-distribution matching, while higher fragment/scaffold cosine and coverage indicate closer composition or broader reference support.Source: §B.5 · p. 21

B.5.1 Property-oracle agreement

  • Random-forest and MiniMol prediction agreement is higher on benchmark molecules than generated molecules, with the weakest generated-molecule agreement on HIV.For hERG, classification agreement should be interpreted alongside correlation metrics because the benchmark test set is class-imbalanced.Source: §B.5.1 · p. 21, Table 6 · p. 21

B.6 Baseline methods

  • The baseline suite includes matched-pair enumeration, de novo GraphDiT sampling, an instruction-tuned LLM, Bayesian optimization, and graph-based generation methods.Baseline-specific implementation details vary; for example, GraphDiT samples from random noise without an input-molecule condition, while MMP-PROP enumerates property-ranked edits.Source: §B.6 · pp. 21–23

C Theoretical Comparison to DPO

The theoretical comparison contrasts PairIT’s tokenwise supervision with DPO’s sequence-level preference objective. Under stated assumptions, the PairIT supervision proxy grows with sequence difference, while DPO’s pair-level gradient can shrink as the preference margin grows.

C Theoretical Comparison to DPO

  • Because PairIT’s token-level parameter gradients can partially cancel, the analysis compares DPO’s parameter gradient with an aggregate tokenwise supervision proxy instead of the full PairIT parameter-gradient norm.Source: C Theoretical Comparison to DPO · p. 24
  • PairIT applies autoregressive cross-entropy to the target tokens of the matched molecule, conditioned on the seed and control tokens.The analysis compares this distributed tokenwise supervision with DPO’s single sequence-level sigmoid margin term.Source: C Theoretical Comparison to DPO · p. 24
  • The PairIT bound assumes the number of edited target positions grows at least linearly with token edit distance and that edited-token probabilities are not already arbitrarily close to one.The comparison also states assumptions on bounded DPO score gradients and margin growth with pair difference.Source: C Theoretical Comparison to DPO · pp. 24–25
  • PairIT’s aggregate tokenwise supervision proxy grows at least linearly with sequence difference, while under a margin-growth assumption DPO’s pair-level gradient becomes exponentially small as the preference margin grows.The authors conclude that PairIT is less prone to this particular sequence-level saturation effect.Source: C Theoretical Comparison to DPO · p. 25

D.1 Detailed MoleculeNet results

The section reports seeded-generation metrics and comparisons across MoleculeNet, including VIN–similarity trade-offs and failure categories. It also provides distribution-level evaluations and additional generations.

D.1 Detailed MoleculeNet results

  • Tables 7–9 report BACE, BBBP, and HIV seeded-improvement metrics after deduplication and selection of the best five of 64 generations per seed.Figure 9 aggregates VIN, SimSeed, and QED; SNN is computed over novel, valid generations only.Source: Figure 9 · p. 26, §D.1 · p. 26, Table 7 · p. 27

D.2 Distribution-level results

  • Distribution-level metrics compare generated molecules with the class-1 training reference set because the test set is too small for stable estimates.Tables 10–12 give the BACE, BBBP, and HIV results.Source: §D.2 · p. 28, Table 10 · p. 28, Table 11 · p. 28

D.3 VIN–Similarity trade-offs

  • Generative methods produce seed similarities well above random-reference levels on BACE, BBBP, and HIV.Mean Tanimoto similarity for random molecule/random active references is 0.158/0.159 for BACE, 0.109/0.101 for BBBP, and 0.093/0.093 for HIV.Source: §D.3 · p. 28, Figure 10 · p. 29, Figure 11 · p. 29

D.4 Failure-mode analysis

  • The failure-mode figures divide retained generations into VIN successes and categories explaining why the remaining generations fail VIN.Each generation is assigned once using the priority order in the figure legend.Source: §D.4 · p. 30, Figure 13 · p. 30, Figure 14 · p. 31

D.5 Additional generation examples

  • Additional examples show PairIT and PairIT-Beam generations on BACE, BBBP, and HIV, highlighting predicted-active molecules and seed duplicates.Figure 16 marks generations with predicted class probability above 0.5 in green and seed duplicates in red.Source: §D.5 · p. 32, Figure 16 · p. 32

D.6 Additional low-data evaluation on hERG

The hERG evaluation tests PairIT under especially sparse pair supervision. PairIT and PairIT-Beam achieve substantial VIN, while the results also show that VIN does not by itself ensure localized edits.

D.6 Additional low-data evaluation on hERG

  • Only 76 unique direction-improving pairs remain from 2,865 initial training pairs after deduplication and ε-filtering; MMPs yield just 41 blocker-to-non-blocker pairs.The resulting usable supervision is approximately 0.18 pairs per training molecule.Source: §D.6 · p. 33
  • 72.9% VIN for PairIT and 79.8% VIN for PairIT-Beam; PairIT-Beam performs on par with the strongest baselines despite sparse pair supervision.PairIT-RL does not improve over SFT on this task.Source: §D.6 · p. 33, Table 13 · p. 34
  • High VIN does not guarantee localized molecular improvement: HierG2G, DDSBM-SB, and PairIT-Beam achieve high VIN with SimSeed < 0.2.For hERG, improvement means changing from blocker to non-blocker; the highest-VIN methods have relatively low SimSeed.Source: §D.6 · p. 33

E.1 Paired vs. single-molecule vs. no pretraining

The ablations compare pretraining objectives and data choices, then examine training dynamics, validation-loss signals, filtering, logP optimization, and zero-shot transfer. The results highlight the importance of paired pretraining and trade-offs between improvement and seed similarity.

E.1 Paired vs. single-molecule vs. no pretraining

  • Single-molecule pretraining is less effective than similar-molecule pair pretraining; more diverse ZINC pairs modestly raise seed similarity, while performance is not monotonic in neighbor count.Uniform single-molecule pretraining can yield seed-agnostic outputs, while small-k pairing can leave too few unique candidates to fill the top five.Source: §E.1 · p. 34
  • Without ZINC pretraining, VIN collapses across MoleculeNet tasks, and invalid SMILES are the main failure mode; the comparison uses the same architecture and fine-tuning data.The no-pretraining model is trained from random initialization using MoleculeNet directed pairs.Source: §E.1 · p. 35, Figure 17 · p. 35, Figure 18 · p. 35

E.2 Training dynamics and molecular coverage

  • Figure 19 compares overfitting behavior and unique molecules encountered at overfitting across paired datasets with and without property filtering and unconditional single-molecule data.Dotted lines mark early stopping at minimum validation loss.Source: Figure 19 · p. 36

E.3 Validation loss and downstream performance

  • ZINC pretraining validation loss is reported as correlated with both generation metrics and downstream fine-tuned BACE performance.Figure 21 tracks checkpoints pretrained on ZINC pair datasets of varying sizes and fine-tuned on the full BACE dataset.Source: Figure 20 · p. 36, Figure 21 · p. 37

E.4 Property filtering ablation

  • More aggressive property-filtering cutoffs induce correspondingly stronger property changes in generated molecules, even without prompt conditioning.Source: §E.4 · p. 37, Figure 22 · p. 37

E.5 ZINC-250K logP optimization

  • Increasing ε raises the Improved rate but lowers seed similarity; PairIT has its strongest joint valid–novel–improved performance at intermediate ε.Figure 23 sweeps pair-construction thresholds for ZINC-250K logP optimization.Source: §E.5 · p. 38, Figure 23 · p. 38

E.6 Zero-shot logP generalization

  • A model pretrained for ZINC logP optimization transfers zero-shot to MoleculeNet logP optimization, with this ability tracking pretraining validation loss.Source: Figure 24 · p. 38

F.1 Reward and training details

PairIT-RL combines gated penalties with property-improvement and seed-similarity rewards, adding prompt adherence for multi-property settings. The section also reports a similarity-regularization trade-off and presents exploratory multi-property RL examples.

F.1 Reward and training details

  • Invalid, copied, or memorized outputs receive fixed penalties; an optional similarity gate penalizes generations below the seed-similarity threshold before positive reward is computed.For molecules passing the gates, the reward combines predicted property change, an active-class bonus, and a Tanimoto similarity term.Source: §F.1 · p. 39
  • The similarity weight and optional gate address seed proximity in chemical space, which the KL penalty alone does not directly enforce.The text identifies wsim = 5 and τsim = 0.2 as an operating point tied to how far from the seed the property model remains reliable.Source: §F.1 · p. 39

F.2 Similarity regularization and reward hacking.

  • Without a similarity penalty, reward optimization collapses toward property-model-fooling attractors: BBBP VIN falls to 0% and HIV VIN to 22%.Setting wsim = 5 recovers seed similarity while maintaining high activity; a hard Tanimoto < 0.2 penalty works similarly, trading some activity for greater similarity.Source: §F.2 · p. 39

F.3 Exploratory multi-property RL

  • The multi-property RL variant combines the activity reward with a prompt-adherence term and is presented as exploratory rather than part of the primary PairIT-RL evaluation.Qualitative examples appear in Appendix F.4.Source: §F.3 · pp. 39–40

F.4.1 Single-property RL

  • Figures 25–30 show single-property RL examples for two seeds each on BACE, BBBP, and HIV; the BACE captions distinguish novel successes, memorized outputs, and seed copies.For BACE, green marks valid, novel, improved outputs; orange marks training-memorized outputs and red marks seed copies.Source: Figure 25 · p. 40, Figure 26 · p. 41, Figure 27 · p. 41

F.4.2 Multi-property RL

G Prompt-conditioned property control

This appendix describes signed-delta prompts and evaluates single- and multi-property control, including settings with correlated and weakly correlated descriptors.

G.1 Prompt tokens

  • Prompts encode a property identifier and a signed-delta bin that jointly specify the requested direction and approximate magnitude; the seed property value is omitted.Generation begins after the prompt tokens and seed SMILES are supplied as context.Source: §G.1 · p. 50

G.2 Single-property control

  • Realized QED and logP changes scale with prompted magnitude, indicating coarse magnitude control in single-property tests.QED prompts request increases; logP prompts cover both directions, and error bars show standard deviation across five generations per seed per bin.Source: Figure 37 · p. 50

G.3 Two-property control on ZINC-250K

  • Under two-property conditioning on ZINC-250K, larger requested edits produce larger realized changes for selected low-correlation pairs, with greater dispersion at large deltas and saturation for bounded properties such as QED.Each panel compares prompted absolute delta with realized absolute delta; the dashed diagonal marks perfect agreement.Source: Figure 38 · p. 51

G.4 Two-property control on MoleculeNet

  • Combined VIN is 12–64% for selected low-correlation pairs on MoleculeNet and 19–42% for the harder joint QED+logP setting.Combined VIN requires valid, novel outputs to satisfy every nonzero requested direction.Source: Figure 39 · p. 52, Figure 40 · p. 53

G.5 Many-property control

  • One model prompts nine RDKit properties plus a target slot, which is reassigned to dataset-specific bioactivity during fine-tuning while all ten property blocks remain included.During pretraining, the target slot mirrors a randomly chosen real property, and property dropout supports conditioning on subsets at inference.Source: §G.5 · p. 54
  • Full multi-property context helps BBBP, where descriptors correlate with bioactivity; generations decrease TPSA by 7.09 and HBD by 0.354, both negatively correlated with BBB permeability.The authors interpret these changes as suggesting that the model learns the structure–activity relationship through multi-property prompting.Source: §G.5 · pp. 54–55
  • The 84-subset sweep spans 27.9 pp on BBBP and 19.5 pp on HIV; HIV’s best subset reaches 25.0% VIN versus 5.5% for its worst.Consistently poor subsets include rotatable bonds and fraction sp3 carbons, while strong subsets vary by dataset.Source: Table 20 · p. 54, §G.5 · p. 55
  • 25.5% VIN on BACE and 19.8% on HIV under the logP+SA+QED subset exceed the full 10-property prompt’s 21.7% and 12.3%, while BBBP falls from 83.0% to 79.3%.The subset therefore improves VIN on BACE and HIV while modestly reducing it on BBBP.Source: Table 20 · p. 54, §G.5 · p. 54
Loading 2610.19123v1…