Conditional Chemical Language Models are Versatile Tools in Drug Discovery

Emmanuel Noutahi; Lu Zhu

REVIEW 4 major objections 4 minor 23 references

Conditional Chemical Language Models are Versatile Tools in Drug Discovery

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read SAFE-T shows that one conditional likelihood over fragment-based molecules can unify scoring and design across drug discovery tasks.

desk verdict SAFE-T is a useful conditional generation/scoring system, but its headline zero-shot PMO claims lack overlap analysis and need re-evaluation on disjoint data. read the letter →

arxiv 2507.10273 v1 pith:CWS3F3J7 submitted 2025-07-14 cs.LG q-bio.BM

Lu Zhu , Emmanuel Noutahi This is my paper · ORCID

classification cs.LGq-bio.BM

keywords conditionalgenerationchemicallanguagemodelsdrug-targetinteractionpredictionvirtualscreeningactivitycliffzero-shotlearningSAFErepresentationmolecularoptimization

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that a generative chemical language model, trained once on molecule–target–mechanism pairs, can replace task-specific scoring functions across early drug discovery. Its central idea is to model the conditional likelihood of a fragment-based molecular string given a biological prompt such as a protein target or mechanism of action, and to reuse that single learned distribution for scoring, ranking, and generation. The authors show zero-shot results on virtual screening, drug-target interaction, activity cliff, and goal-directed generation benchmarks, with the generative results reaching the top of the PMO benchmark without any per-task optimization. The payoff, if true, is that one model can run large-scale screening at high speed and generate lead-like molecules while also explaining its decisions at fragment level.

What carries the argument

The load-bearing object is SAFE-T, an autoregressive transformer over SAFE strings — a fragment-based line notation in which a molecule is a sequence of chemically connected fragments — with biological context tokens prepended to the sequence. Its engine is the factorization $p(x \mid c) = \prod_t p(x_t \mid x_{<t}, c)$, so that scoring and generation are literally the same operation. Three training stages of pretraining on unlabeled chemistry, fine-tuning on MoAT-DB with random context masking, and DPO (direct preference optimization) calibration on ACNet pairs shape this distribution, and likelihood normalization schemes — population-based $l_{\text{pop}}$ for prioritization, null-condition $l_{\text{null}}$ for classification — make scores comparable across contexts.

What would settle it

Compute the overlap between the MoAT-DB training pairs and every active in the PMO benchmark for DRD2, JNK3, and GSK-3$\beta$; if a large fraction appear in training and the PMO AUC drops substantially when they are removed, the zero-shot generalization claim fails. A complementary test is to measure scaffold novelty of SAFE-T's top-ranked generated molecules against ChEMBL actives for those targets.

Watch

Extended reading notes

Core claim

SAFE-T is a transformer that learns $p(x \mid c)$, the probability of a molecule written as a SAFE fragment sequence given a biological context $c = (\text{target family}, \text{protein target}, \text{mechanism of action})$. From this one distribution, the paper derives every task: ranking a compound library is ordering by $p(x \mid c)$; predicting a target is Bayesian inversion $p(c \mid x) \propto p(x \mid c)\,p(c)$; an activity cliff is a large log-likelihood gap between near-identical molecules; and goal-directed design is sampling from $p(x \mid c)$ with a prompt. The model is trained in three stages — structure-only pretraining, biological-context fine-tuning on a consolidated dataset of 1.15M compound–target–MoA pairs, and preference calibration on activity-cliff pairs — and then evaluated without further adaptation. The authors report that this zero-shot recipe matches or beats specialized optimizers on DRD2, JNK3, and GSK-3$\beta$, enriches actives on LIT-PCBA, and yields fragment attributions consistent with known structure–activity relationships for the JAK inhibitors tofacitinib and ruxolitinib.

Load-bearing premise

The zero-shot results assume that the benchmark actives — especially the DRD2, JNK3, and GSK-3$\beta$ targets in the PMO benchmark — were not memorized during training, so high scores reflect generalization to unseen chemistry rather than recall of training pairs.

Editorial extensions

If this is right

A single trained model could replace structure-based docking and task-specific scoring functions for early hit finding, because likelihood ranking runs at roughly 800 molecules per second on one GPU.
Goal-directed generation becomes prompt engineering rather than reward engineering: new design objectives need no fine-tuning or reinforcement learning.
Activity cliff detection and fragment-level attribution emerge from the same distribution, so chemists get interpretable structure–activity relationship hypotheses alongside ranked candidate lists.
The same conditional likelihood can be reported across predictive and generative benchmarks, simplifying evaluation of generalist drug-discovery models.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the zero-shot results survive overlap checks, conditional likelihood itself is a candidate universal fitness function; one direction the authors leave open is using $p(x \mid c)$ for target deconvolution, inverting the model to propose the most likely biological context of an uncharacterized active molecule.
The paper hints at polypharmacology via selectivity profiling but does not train on multi-target prompts; an immediate testable extension is prompting with a panel of targets and checking whether sampled molecules balance affinities across the panel.
The 50-shot adaptation result for unseen targets could be stress-tested against docking for a target with very few known actives, since the authors only compare zero-shot and 50-shot settings internally.
Because MoAT-DB is built from active-biased public sources, the model's inverted predictions may inherit the same bias; a useful probe is whether $l_{\text{null}}$ scores suppress known decoys better than actives on DUD-E after removing training overlap.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things to know about this one. First, SAFE-T is a genuinely useful system: it conditions a SAFE fragment language model on protein family, target, and mechanism-of-action tokens, then uses that single conditional distribution for both scoring and generation. The code and data are public, generation validity is high (>95%), and the reported inference speed (~800 molecules/s on an A100) is an order of magnitude above docking baselines. Second, the headline claim—state-of-the-art zero-shot PMO results on DRD2, JNK3, and GSK-3β—is not supported as written. The paper never reports overlap analysis between those targets' actives and MoAT-DB training data, even though DRD2, JNK3, and GSK-3β are among the most annotated targets in ChEMBL34, which MoAT-DB is built from. That omission makes the near-perfect AUCs (0.9950, 0.9927, 0.9948 with near-zero std) look more like recall of memorized actives than generalization. Notably, the paper's own unseen-target experiments show much more variable zero-shot performance (e.g., BCDIN3D ROC-AUC 0.401, SH2D1A EF@1% 0.0), which is hard to square with a model that nails three well-known targets out of the box.

What the paper does well: the unified scoring/generation interface is a solid contribution, the fragment attribution on Tofacitinib and Ruxolitinib matches known SAR, and the DPO-based activity cliff result on ACNet (0.947 vs 0.581 before tuning) is striking if the train/test split really excludes overlapping compounds. The authors also deserve credit for disclosing the DUD-E and GPCR-Bench overlap in Table 10, though those benchmarks should not be described as zero-shot in the abstract when 74.8% and 94.6% of actives appear in training.

The soft spots, in order of severity: (1) missing PMO overlap analysis—needs to be addressed head-on; (2) normalization strategies (lnull vs lpop) are chosen per task after seeing test results—should be pre-specified or validated on held-out data; (3) ACNet DPO split details are too sparse to fully rule out leakage. All are fixable with more analysis, not new science.

Bottom line: this is a serious paper that deserves a real referee, but the current evaluation overreaches. I'd send it to review and require the authors to report overlap stats for every target used in zero-shot claims, rerun key results after removing training actives or with novel targets, and pre-register the normalization choice. If those checks pass, it's a strong systems paper.

Referee Report

4 major / 4 minor

Summary. This paper introduces SAFE-T, a transformer-based conditional generative model over SAFE fragment sequences. Each molecule is scored as p_theta(x | c) with biological context c = (family, target, MoA) tokens, and training proceeds in three stages: unconditional pretraining on ~11.4M molecules, fine-tuning on a newly assembled MoAT-DB of ~1.15M compound-target-MoA records, and DPO calibration on preference pairs derived from the ACNet benchmark. The authors claim that the resulting model supports zero-shot molecular optimization (PMO DRD2/JNK3/GSK-3beta), virtual screening (LIT-PCBA, DUD-E, GPCR-Bench), target-family and DTI prediction (DAVIS, KIBA, MoAT-DB held-out splits), activity cliff detection (ACNet), and fragment-level SAR attribution, with GPU inference throughput of about 800 molecules per second. The headline result is Table 2, where zero-shot conditional sampling is reported to outperform optimization-based methods on all three PMO tasks.

Significance. If the zero-shot claims are valid, this is a significant contribution: one conditional distribution would unify scoring and generation, remove the need for engineered reward functions or structural information, and provide a fast screening tool with interpretable fragment attributions. The paper is also commendable for releasing code and data-preparation scripts, for documenting the construction of MoAT-DB in detail (Appendix A.1), and for explicitly quantifying benchmark overlap for DUD-E, GPCR-Bench, and LIT-PCBA (Table 10), which is more transparent than typical practice in this area. The inference-speed result (Table 11) is a concrete practical strength. However, the significance is conditional: the central zero-shot optimization claim rests on an overlap analysis that is not provided, and several predictive evaluation choices appear to have been made after seeing the benchmarks, so the demonstrated generality of the framework is not yet established.

major comments (4)

[Section 4.3, Table 2; Appendix A.1] The PMO zero-shot claim lacks a training-overlap analysis for its headline targets. Section 4.3 (Table 2) reports state-of-the-art PMO AUC values (DRD2 0.9950, JNK3 0.9927, GSK-3beta 0.9948) for zero-shot conditional sampling, but the manuscript nowhere reports whether actives for these three targets appear in MoAT-DB. Since MoAT-DB is built from ChEMBL34 among other sources and DRD2, JNK3, and GSK-3beta are among the most heavily annotated ChEMBL targets, the near-perfect AUCs with standard deviations of 0.0000-0.0003 could largely reflect recall of memorized training actives rather than generalization; the comparison against RL and search baselines that start from unseeded distributions would then be unfair. Table 10 gives overlap statistics only for DUD-E, GPCR-Bench, and LIT-PCBA. Please add target-level overlap percentages for the PMO actives, report performance after removing all training actives from the generated or scored sets, and state the exact oracle versions used so the reader can see how much of the reported AUC comes from molecules present in MoAT-DB.
[Appendix A.6, Tables 10 and 12] Tables 10 and 12 in Appendix A.6 report SAFE-T as top-scoring on DUD-E (EF@1% 58.66) and GPCR-Bench (EF@1% 57.53) while Table 10 shows that 74.81% of DUD-E actives and 94.62% of GPCR-Bench actives are present in MoAT-DB. The accompanying caveat that the overlap 'should be considered when interpreting the results' does not repair the comparison because the baselines are not evaluated on the non-overlapping subset, and the main-text and abstract claims of 'comparable to or better than existing approaches' are made without this qualifier. Please report metrics restricted to the non-overlapping actives, or exclude overlapping compound-target pairs entirely, for every benchmark rather than only for LIT-PCBA.
[Sections 4.4-4.5; Appendix A.5, Table 9] The normalization strategy is selected per task after observing benchmark results, which compromises the predictive evaluations. Section 4.4 reports target-family prediction with lnull and Section 4.5 reports LIT-PCBA with lpop, but Appendix A.5 (Table 9) shows that these choices were made by comparing l, lprior, lpop, lpop+prior, lnull, and lnull+prior on the same Dval/DMol/DMoA tasks and explicitly recommends different strategies for different tasks. If the normalization is chosen after inspecting results on the evaluation sets, the reported predictive scores are optimistic estimates of zero-shot performance. Please pre-specify the normalization per task, or report all normalization variants for each benchmark and correct for the resulting multiple comparisons.
[Section 4.7; Appendix A.2.3] The abstract lists ACNet among the 'zero-shot evaluations,' but the activity cliff result in Section 4.7 (SAFE-Tpost ROC-AUC 0.947) is obtained with a model trained by DPO on preference pairs derived from the ACNet benchmark (Appendix A.2.3). Excluding training compounds from the new random splits addresses molecule-level leakage only; the target-MoA contexts and the preference supervision still come from the same benchmark, and the jump from 0.581 (SAFE-T) to 0.947 (SAFE-Tpost) is more naturally read as task-specific calibration than as zero-shot generalization. Please evaluate the preference-tuned model on a benchmark not used for DPO, or hold out entire target-MoA combinations during DPO training.

minor comments (4)

[Introduction; Section 4.4] The introduction contains the typo 'abscence' ('absence'), and Section 4.4 writes 'DA VIS' with an unintended space; the benchmark should be 'DAVIS'.
[Table 1] The Table 1 header shows 'SAS ↑', but the synthetic accessibility score is defined in Table 13 as a score where lower is better; the arrow is misleading and should be corrected.
[Table 1] Several rows in Table 1 omit standard deviations (for example, the 'SAFE (baseline)' de novo row has no +/- values for Int.Div or Validity), which makes it difficult to compare variability across methods and contexts.
[Table 10] The header 'Overlap with MoA T-DB' contains a typo; it should read 'Overlap with MoAT-DB'.

Circularity Check

2 steps flagged · score 4.0 of 10

PMO 'zero-shot' top-10 AUCs may be memorized ChEMBL34 actives, and the ACNet cliff-detection result is fit on the same benchmark yet the abstract calls it zero-shot.

fitted input called prediction [Section 4.3 (Table 2); Appendix A.1; Table 10]
"SAFE-T employs a purely generative strategy driven by conditional prompts, requiring no optimization for specific tasks. Remarkably, as shown in Table 2, this zero-shot methodology achieves state-of-the-art performance on all three benchmark tasks ... We introduce MoAT-DB ... aggregating ... ChEMBL34 ... Overlap with MoAT-DB: DUD-E active compounds 74.81%, GPCR-Bench active compounds 94.62%."

MoAT-DB, the fine-tuning corpus, is compiled from ChEMBL34, and the PMO targets DRD2, JNK3, and GSK-3β are standard, heavily annotated ChEMBL targets. The paper provides overlap statistics exposing 74.81% and 94.62% active-compound overlap with MoAT-DB for DUD-E and GPCR-Bench, respectively, but reports no analogous check for the PMO targets. The PMO metric is the AUC of the top-10 oracle scores over molecules sampled from pθ(x|c), and pθ(x|c) was trained on MoAT-DB pairs. If the PMO actives are in MoAT-DB, the 'zero-shot' top-10 scores reduce to ranking memorized training actives, not generalization, and the comparison against optimization baselines that do not start from those actives is statistically forced.
fitted input called prediction [Abstract; Section 4.7 (Table 6); Appendix A.2.3]
"In comprehensive zero-shot evaluations across predictive (LIT-PCBA, DA VIS, KIBA, ACNet) and generative (DRUG, PMO) benchmarks ... For preference tuning, we use pairwise tuples derived from the ACNet benchmark ... We evaluated SAFE-T on activity cliff detection using the ACNet benchmark ... preference-tuned SAFE-Tpost achieves performance (0.947) comparable to the best supervised baselines."

The activity-cliff detector is explicitly obtained by Direct Preference Optimization on ACNet preference tuples, and then the resulting model is evaluated on the ACNet benchmark. The abstract labels ACNet as a zero-shot evaluation, but the 0.947 AUC of SAFE-Tpost is the outcome of fitting the model to the same benchmark's preference data. This is a fitted input being renamed as a capability of the conditional generative model: the reported 'prediction' is forced by the training pairs rather than emerging from the zero-shot conditional distribution pθ(x|c).

full rationale

The mathematical core of the paper is standard: pθ(x|c) is a causal likelihood, p(c|x) ∝ pθ(x|c)p(c) is Bayes inversion, and the fragment attribution ψfi compares likelihoods of counterfactuals. I find no equation-level self-definition in those derivations. The circularity burden is in the evaluation claims. First, the central generative result — zero-shot state-of-the-art PMO AUCs on DRD2, JNK3, and GSK-3β — omits any overlap analysis against MoAT-DB, even though the paper demonstrates for other benchmarks that overlap with MoAT-DB reaches 74.8–94.6% of active compounds. Without an overlap or retraining analysis, those near-perfect AUCs may be recall of ChEMBL34 actives, making the 'zero-shot' label and the comparison to optimization baselines unfair. Second, the ACNet cliff-detection result is produced by preference-tuning on ACNet pairs yet the abstract lists ACNet among zero-shot benchmarks; fitting to a benchmark and then reporting the benchmark result is not a self-contained demonstration. These issues are benchmark-circularity problems rather than algebraic circularity, so I do not assign a score of 6 or higher, but they do affect the paper's headline claims and justify a score of 4.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the quality of MoAT-DB annotations, the validity of posterior inference, SAFE fragment semantics, and the assumption that benchmark actives are absent from training. No new physical entities are introduced; SAFE-T is an architecture, not a postulated entity.

free parameters (3)

Normalization strategy per task (lpop, lnull, lprior) = chosen post hoc; e.g., lpop for prioritization, lnull for DTI
Appendix A.5 reports task-specific selection after comparing variants; this is a model-selection choice that determines the reported scores and is not fixed a priori.
Activity cliff threshold delta = not specified; chosen via percentile/outlier heuristics
A.3.2 defines a cliff as an absolute log-likelihood difference exceeding delta, with delta set empirically; the activity cliff results depend on this unstated constant.
DPO temperature beta and ACNet preference split = not reported
A.2.3 says beta controls the preference margin but no value is given; SAFE-Tpost results depend on it, and overlap between DPO training pairs and ACNet evaluation splits is not fully specified.

assumptions (5)

domain assumption MoAT-DB annotations (target, family, MoA) are accurate and sufficiently standardized for conditional training.
A.1 manually standardizes 22 MoA categories, labels 85% of entries as missing, and retains one protein class per target, which introduces noise; predictive performance depends on this.
standard math Bayes rule p(c|x) proportional to p_theta(x|c) p(c) with p(c) estimated from training data is a valid inference scheme for DTI and MoA prediction.
A.3.1 invokes posterior inference; prior frequencies from MoAT-DB may bias rare targets.
domain assumption SAFE/BRICS fragment decomposition preserves structure-activity signal and supports fragment-level attribution.
A.1 and A.4 use BRICS decomposition and counterfactual replacement; there is no independent validation that SAFE fragments correspond to pharmacophoric units.
domain assumption Benchmark targets in PMO and LIT-PCBA are not substantially represented in the training set, so zero-shot scores measure generalization.
Implicit in Sections 4.3 and 4.5; Table 10 reports overlap only for DUD-E, GPCR-Bench, and LIT-PCBA, not PMO; if actives are in MoAT-DB, results are inflated.
domain assumption Oracle scores in PMO accurately reflect drug-target activity and are comparable to training labels.
Section 4.3 uses PMO oracles as ground truth for goal-directed optimization; no benchmark overlap analysis is provided for these targets.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Conditional Chemical Language Models are Versatile Tools in Drug Discovery." pith.science (2026). https://pith.science/paper/CWS3F3J7

@misc{pith2026250710273,
  author       = {Pith},
  title        = {Pith review of: Conditional Chemical Language Models are Versatile Tools in Drug Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CWS3F3J7}},
  note         = {Machine review of arXiv:2507.10273}
}

read the original abstract

Generative chemical language models (CLMs) have demonstrated strong capabilities in molecular design, yet their impact in drug discovery remains limited by the absence of reliable reward signals and the lack of interpretability in their outputs. We present SAFE-T, a generalist chemical modeling framework that conditions on biological context -- such as protein targets or mechanisms of action -- to prioritize and design molecules without relying on structural information or engineered scoring functions. SAFE-T models the conditional likelihood of fragment-based molecular sequences given a biological prompt, enabling principled scoring of molecules across tasks such as virtual screening, drug-target interaction prediction, and activity cliff detection. Moreover, it supports goal-directed generation by sampling from this learned distribution, aligning molecular design with biological objectives. In comprehensive zero-shot evaluations across predictive (LIT-PCBA, DAVIS, KIBA, ACNet) and generative (DRUG, PMO) benchmarks, SAFE-T consistently achieves performance comparable to or better than existing approaches while being significantly faster. Fragment-level attribution further reveals that SAFE-T captures known structure-activity relationships, supporting interpretable and biologically grounded design. Together with its computational efficiency, these results demonstrate that conditional generative CLMs can unify scoring and generation to accelerate early-stage drug discovery.

Figures

Figures reproduced from arXiv: 2507.10273 by the authors.

**Figure 2.** Normalized likelihood distributions for decoys and binders on selected target from the [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

**Figure 3.** Fragment attribution analysis for Tofacitinib, a JAK inhibitor. Each point represents a [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

Figures from the paper (9 more)

**Figure 4.** Figure 4: TMAP visualization of chemical space coverage in MoAT-DB, colored by target family. The [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]

**Figure 5.** Figure 5: Distribution of unique compounds and scaffolds across target families in MoAT-DB, illustrating [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]

**Figure 6.** Figure 6: Example of sampled molecules under the de novo and fragment-constrained design given the biological context [TK protein kinase]/[CHEMBL203]/[inhibitor]. The shared core structures are [2*]C#C.[14*]OCCOC for linker design and [1*]c1cccc(Nc2ncnc3cc([2*])c([3*])cc23)c1 fo…

**Figure 7.** Figure 7: Likelihood distributions under a Tyrosine Protein Kinase inhibitor prompt for the Enamine [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]

**Figure 8.** Figure 8: SAFE-T vs GEAM fragment-level interpretability on Ruxolitinib (a JAK2 inhibitor). [PITH_FULL_IMAGE:figures/full_fig_p024_8.png]

**Figure 9.** Figure 9: Counterfactual space around Tofacitinib (JAK1 inhibitor). ∆ represents the difference in [PITH_FULL_IMAGE:figures/full_fig_p025_9.png]

**Figure 10.** Figure 10: Distribution of generative design performance as a function of SAFE-T model size and [PITH_FULL_IMAGE:figures/full_fig_p029_10.png]

**Figure 11.** Figure 11: Performance distribution of SAFE-T models in molecule prioritization on the LIT-PCBA [PITH_FULL_IMAGE:figures/full_fig_p030_11.png]

**Figure 12.** Figure 12: Performance of SAFE-T models on activity cliff prediction across 185 Target-MoA [PITH_FULL_IMAGE:figures/full_fig_p031_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 12 canonical work pages

[3]

C., and Murray, C

Jhoti, H., Williams, G., Rees, D. C., and Murray, C. W. The ’rule of three’ for fragment-based drug discovery: where are we now? Nature Reviews Drug Discovery 2013 12:8, 12:644–644, 7

work page 2013
[8]

A unified approach to interpreting model predictions

Lundberg, S. A unified approach to interpreting model predictions. arXiv preprint arXiv:1705.07874,

arXiv
[9]

Mesbahi, Y. E. and Noutahi, E. Safe setup for generative molecular design. arXiv preprint arXiv:2410.20232,

work page Pith review arXiv
[10]

Fame: fragment-based conditional molecular generation for phenotypic drug discovery

Pham, T.-H., Xie, L., and Zhang, P. Fame: fragment-based conditional molecular generation for phenotypic drug discovery. In Proceedings of the 2022 SIAM International Conference on Data Mining (SDM), pp. 720–728. SIAM,

work page 2022
[12]

and Belinkov, Y

Vig, J. and Belinkov, Y. Analyzing the structure of attention in a transformer language model. arXiv preprint arXiv:1906.04284,

arXiv 1906
[15]

doi: 10.1021/ACS.JMEDCHEM.3C02070/SUPPL FILE/JM3C02070 SI 001.PDF

ISSN 15204804. doi: 10.1021/ACS.JMEDCHEM.3C02070/SUPPL FILE/JM3C02070 SI 001.PDF. Wu, J.-N., Wang, T., Chen, Y., Tang, L.-J., Wu, H.-L., and Yu, R.-Q. t-smiles: a fragment-based molecular representation framework for de novo ligand design. Nature Communications, 15(1):4993, 2024a. Wu, K., Xia, Y., Deng, P., Liu, R., Zhang, Y., Guo, H., Cui, Y., Pei, Q., W...

work page doi:10.1021/acs.jmedchem.3c02070/suppl
[16]

J., Blackshaw, J., Corbett, S., Veij, M

Zdrazil, B., Felix, E., Hunter, F., Manners, E. J., Blackshaw, J., Corbett, S., Veij, M. de, Ioannidis, H., Lopez, D. M., Mosquera, J. F., Magarinos, M. P., Bosc, N., Arcila, R., Kizil¨ oren, T., Gaulton, A., Bento, A. P., Adasme, M. F., Monecke, P., Landrum, G. A., and Leach, A. R. The chembl database in 2023: a drug discovery platform spanning multiple ...

work page 2023
[17]

doi: 10.1093/NAR/GKAD1004

ISSN 0305-1048. doi: 10.1093/NAR/GKAD1004. Zhang, X., Gao, H., Wang, H., Chen, Z., Zhang, Z., Chen, X., Li, Y., Qi, Y., and Wang, R. Planet: a multi-objective graph neural network model for protein–ligand binding affinity prediction. Journal of Chemical Information and Modeling, 64(7):2205–2220, 2023a. Zhang, Z., Zhao, B., Xie, A., Bian, Y., and Zhou, S. ...

arXiv

Show all 23 references

[18]

doi: 10.1093/NAR/GKAD751

ISSN 1362-4962. doi: 10.1093/NAR/GKAD751. 17 Conditional CLMs are Versatile Tools in Drug Discovery A Appendix A.1 MoAT-DB: A New Dataset of Target–Compound–MoA Relationships We introduce MoAT-DB, a new dataset ( D) consolidating Target–Compound–MoA information from four major...

doi
[19]

For context- conditioned fine-tuning, we use MoAT-DB (see Appendix A.1)

with all unique molecules from ChEMBL. For context- conditioned fine-tuning, we use MoAT-DB (see Appendix A.1). For preference tuning, we use pairwise tuples derived from the ACNet benchmark (Zhang et al., 2023b). All molecules are converted from SMILES to SAFE using the BRICS...

2008
[20]

A.2.1 Pre-training on Chemical Data The first training stage involves pre-training a generative model over SAFE strings with biological context tokens masked

All training runs use up to four NVIDIA H100 GPUs with 32 CPU cores per job and a maximum runtime of 96 hours per training stage. A.2.1 Pre-training on Chemical Data The first training stage involves pre-training a generative model over SAFE strings with biological context tok...

1953
[21]

(left image) assigns the highest attribution score to the cyclopentyl group, which has been shown in the literature to be non-essential for JAK2 inhibitory activity. (b) SAFE- T (right image) correctly identifies more relevant fragments (e.g the central pyrrolopyrimidine core)...

2014
[23]

With MoA

and DUD- E (Mysinger et al., 2012). Each benchmark presents distinct challenges and validation scenarios. DUD-E provides computationally generated decoys designed to challenge molecular docking programs while maintaining similar physical property distributions to known actives...

2021
[2004]

Explainability techniques for chemical language models

H¨ odl, S., Robinson, W., Bachrach, Y., Huck, W., and Kachman, T. Explainability techniques for chemical language models. arXiv preprint arXiv:2305.16192,

arXiv
[2010]

In Figure 9, we illustrate a counterfactual exploration around Tofacitinib, a known Janus kinase (JAK) inhibitor

and enables hypothesis generation for lead optimization. In Figure 9, we illustrate a counterfactual exploration around Tofacitinib, a known Janus kinase (JAK) inhibitor. Fragment replacements are sampled either from the true distribution pθ(· |x\fi , ctrue) or a random or unc...

2016
[2013]

doi: 10.1038/nrd3926-c1

ISSN 1474-1784. doi: 10.1038/nrd3926-c1. Jiang, Y., Zhang, G., You, J., Zhang, H., Yao, R., Xie, H., Zhang, L., Xia, Z., Dai, M., Wu, Y., et al. Pocketflow is a data-and-knowledge-driven structure-based molecular generative model. Nature Machine Intelligence, 6(3):326–337,

doi
[2017]

F., Zhang, F., Pilgrim, R., Matias, Y., Barral, J., Fleet, D., and Azizi, S

Wang, E., Schmidgall, S., Jaeger, P. F., Zhang, F., Pilgrim, R., Matias, Y., Barral, J., Fleet, D., and Azizi, S. Txgemma: Efficient and agentic llms for therapeutics. arXiv preprint arXiv:2504.06196,

arXiv
[2018]

doi: 10.1093/NAR/GKX1037

ISSN 1362-4962. doi: 10.1093/NAR/GKX1037. Wognum, C., Ash, J. R., Aldeghi, M., Rodr ´ ıguez-P´ erez, R., Fang, C., Cheng, A. C., Price, D. J., Clevert, D.-A., Engkvist, O., and Walters, W. P. A call for an industry-led initiative to critically assess machine learning for real-...

doi
[2020]

Lee, S., Lee, S., Kawaguchi, K., and Hwang, S. J. Drug discovery with dynamic goal-aware fragments. arXiv preprint arXiv:2310.00841,

arXiv
[2021]

Inseq: An inter- pretability toolkit for sequence generation models

Sarti, G., Feldhus, N., Sickert, L., Van Der Wal, O., Nissim, M., and Bisazza, A. Inseq: An inter- pretability toolkit for sequence generation models. arXiv preprint arXiv:2302.13942,

arXiv
[2022]

URL https://arxiv.org/abs/2206.12411v2

ISSN 10495258. URL https://arxiv.org/abs/2206.12411v2. G´ omez-Bombarelli, R., Wei, J. N., Duvenaud, D., Hern´ andez-Lobato, J. M., S´ anchez-Lengeling, B., Sheberla, D., Aguilera-Iparraguirre, J., Hirzel, T. D., Adams, R. P., and Aspuru-Guzik, A. Automatic chemical design usi...

arXiv
[2023]

P., Liu, M., Reidenbach, D., Peng, Y., Paliwal, S., Nie, W., and Vahdat, A

Lee, S., Kreis, K., Veccham, S. P., Liu, M., Reidenbach, D., Peng, Y., Paliwal, S., Nie, W., and Vahdat, A. Genmol: A drug discovery generalist with discrete diffusion. arXiv preprint arXiv:2501.06158,

arXiv
[2024]

Generating 3d molecules for target protein binding

Liu, M., Luo, Y., Uchino, K., Maruhashi, K., and Ji, S. Generating 3d molecules for target protein binding. arXiv preprint arXiv:2204.09410,

arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [3]

C., and Murray, C

Jhoti, H., Williams, G., Rees, D. C., and Murray, C. W. The ’rule of three’ for fragment-based drug discovery: where are we now? Nature Reviews Drug Discovery 2013 12:8, 12:644–644, 7

work page 2013

[2] [8]

A unified approach to interpreting model predictions

Lundberg, S. A unified approach to interpreting model predictions. arXiv preprint arXiv:1705.07874,

arXiv

[3] [9]

Mesbahi, Y. E. and Noutahi, E. Safe setup for generative molecular design. arXiv preprint arXiv:2410.20232,

work page Pith review arXiv

[4] [10]

Fame: fragment-based conditional molecular generation for phenotypic drug discovery

Pham, T.-H., Xie, L., and Zhang, P. Fame: fragment-based conditional molecular generation for phenotypic drug discovery. In Proceedings of the 2022 SIAM International Conference on Data Mining (SDM), pp. 720–728. SIAM,

work page 2022

[5] [12]

and Belinkov, Y

Vig, J. and Belinkov, Y. Analyzing the structure of attention in a transformer language model. arXiv preprint arXiv:1906.04284,

arXiv 1906

[6] [15]

doi: 10.1021/ACS.JMEDCHEM.3C02070/SUPPL FILE/JM3C02070 SI 001.PDF

ISSN 15204804. doi: 10.1021/ACS.JMEDCHEM.3C02070/SUPPL FILE/JM3C02070 SI 001.PDF. Wu, J.-N., Wang, T., Chen, Y., Tang, L.-J., Wu, H.-L., and Yu, R.-Q. t-smiles: a fragment-based molecular representation framework for de novo ligand design. Nature Communications, 15(1):4993, 2024a. Wu, K., Xia, Y., Deng, P., Liu, R., Zhang, Y., Guo, H., Cui, Y., Pei, Q., W...

work page doi:10.1021/acs.jmedchem.3c02070/suppl

[7] [16]

J., Blackshaw, J., Corbett, S., Veij, M

Zdrazil, B., Felix, E., Hunter, F., Manners, E. J., Blackshaw, J., Corbett, S., Veij, M. de, Ioannidis, H., Lopez, D. M., Mosquera, J. F., Magarinos, M. P., Bosc, N., Arcila, R., Kizil¨ oren, T., Gaulton, A., Bento, A. P., Adasme, M. F., Monecke, P., Landrum, G. A., and Leach, A. R. The chembl database in 2023: a drug discovery platform spanning multiple ...

work page 2023

[8] [17]

doi: 10.1093/NAR/GKAD1004

ISSN 0305-1048. doi: 10.1093/NAR/GKAD1004. Zhang, X., Gao, H., Wang, H., Chen, Z., Zhang, Z., Chen, X., Li, Y., Qi, Y., and Wang, R. Planet: a multi-objective graph neural network model for protein–ligand binding affinity prediction. Journal of Chemical Information and Modeling, 64(7):2205–2220, 2023a. Zhang, Z., Zhao, B., Xie, A., Bian, Y., and Zhou, S. ...

arXiv