Pith. sign in

REVIEW 4 major objections 5 minor 135 references

A human-AI system that walks chemists through reasoned steps produces higher-quality, more diverse drug candidates than a baseline with the same generative model.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 15:00 UTC pith:DYBYJQG3

load-bearing objection A solid HCI systems paper with a genuinely new co-abduction framework and a competent expert study; the 'higher-quality molecules' headline is over-claimed because it rests on the same surrogate model the system helps users optimize. the 4 major comments →

arxiv 2607.18564 v2 pith:DYBYJQG3 submitted 2026-07-20 cs.HC

HALO: Interactive Co-abductive Reasoning in Scientific Hypothesis Generation

Youngseung Jeon , Kat Limqueco , JiaSyuan Chang , Xiang 'Anthony' Chen This is my paper
classification cs.HC
keywords abductive reasoninghuman-AI collaborationhypothesis generationdrug discoveryligand optimizationlarge language modelsscientific discoverycreativity support tools
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that the bottleneck in scientific hypothesis generation is not generating candidates but reasoning over them, and that this reasoning can be operationalized as a human-AI collaboration called co-abduction. It presents HALO, an interface that walks medicinal chemists through three stages: clustering generated molecules by which properties they improve, identifying optimization strategies within each cluster, and synthesizing strategies across clusters. In an expert study with ten medicinal chemists, HALO users rated the system significantly higher on observation, strategy identification, and multi-strategy composition, and they produced molecules that improved more target properties and showed greater structural diversity than those from a baseline without the abductive structure. A sympathetic reader would care because this reframes what AI support for scientific discovery should look like: not more generation, but structured support for the abductive reasoning that turns raw candidates into grounded hypotheses.

Core claim

The paper claims that human-AI collaboration can operationalize abductive reasoning—inference to the most plausible explanation—as a three-stage loop, and that this loop, embodied in the HALO system, measurably improves molecular hypothesis generation. In a within-subjects study with ten medicinal chemists, HALO users rated the system significantly higher on efficient observation, systematic strategy identification, and coherent multi-strategy composition, and the molecules they submitted improved more of four target properties while being more structurally diverse, compared with a baseline that had the same generative model but no abductive structure. The authors also observed "abductive le

What carries the argument

The central object is the co-abduction framework, a decomposition of abductive reasoning into three stages: (1) clustering hypothesis candidates by which target properties they improve or worsen; (2) distilling optimization strategies within each cluster via LLM-generated explanations of fragment–property relationships; and (3) synthesizing strategies across clusters into new candidates. In HALO, this is realized through a clustering module that groups molecules by property-improvement profiles, a strategy module that computes shared scaffolds and R-group fragment libraries and prompts a large language model for actionable strategies, and a synthesis module that recombines intra- and inter-c

Load-bearing premise

The central claim assumes that the number of target properties a molecule improves, as scored by a machine-learning ADMET predictor, is a valid measure of drug-candidate quality; the paper acknowledges that wet-lab validation is needed.

What would settle it

Take the molecules that HALO users submitted, synthesize them, and measure the target properties in vitro or in vivo; if the predicted property improvements (e.g., lower liver toxicity, better solubility) do not reproduce, the paper's quality claim collapses.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If co-abduction works as described, AI tools for scientific discovery should focus less on generating more candidates and more on structuring the reasoning over those candidates.
  • Clustering by property-improvement profiles makes the optimization landscape legible, letting researchers quickly see which properties are hard to improve and which clusters to focus on.
  • LLM-generated, fragment-grounded strategies reduce reliance on manual and heuristic inspection, giving experts evidence-based directions they might not have considered.
  • Recombining strategies across clusters with real-time property updates preserves structural coherence and supports iterative refinement toward multi-property goals.
  • The observed post-insight shift from AI generation to manual editing suggests that AI should be more proactive before an abductive leap and more restrained afterward.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • My inference: the co-abduction loop is a general reasoning scaffold, not a drug-discovery tool; it should transfer to materials discovery, protein design, or any domain where candidates are scored on competing properties and strategies can be recombined.
  • My inference: the paper's 'leap' finding implies an adaptive-interaction design principle—systems could detect when a user has formed a hypothesis and automatically lower AI initiative to avoid interference.
  • My inference: a stronger test would compare HALO against a baseline that shows the same LLM strategies without the clustering step, isolating whether the observed benefit comes from clustering or from strategy content alone.
  • My inference: because quality is measured by a machine-learning property predictor, the magnitude of the quality gain is an upper bound; wet-lab validation could shrink or eliminate it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces co-abduction, a human-AI collaborative framework for abductive reasoning in scientific hypothesis generation, and implements it in HALO, a system for molecular lead optimization in drug discovery. HALO operationalizes three co-abductive stages: clustering generated molecules by property improvement profiles (MolCluster), identifying intra-cluster optimization strategies via an LLM (MolStrategy), and synthesizing strategies across clusters (MolSynthesis). The authors evaluate HALO in a within-subjects study with 10 medicinal chemists, comparing it against a baseline that includes the same AI generation and property scoring infrastructure but omits the three co-abductive components. They report significant advantages for HALO on self-reported measures of observation, strategy identification, and multi-strategy synthesis, and on objective measures of submitted molecule quality (ADMET-AI-improved property count) and diversity (lower Tanimoto similarity). The paper also analyzes 'abductive leap' events logged via an Insight button, finding reduced AI generation and increased manual editing after the first leap.

Significance. If the findings hold, the work is a useful contribution to HCI and AI-assisted scientific discovery: it targets the reasoning process around candidate generation rather than only candidate generation itself, and it provides a concrete system design and a controlled comparison that isolates co-abductive features. Strengths include the within-subjects counterbalanced study, the baseline designed to separate the three stages, the use of both self-report and interaction-log measures, the computational pre-evaluation of the molecule generator, and the authors' explicit acknowledgment of several limitations. The main risk is that the headline 'higher-quality molecules' claim in the abstract and conclusion rests on a quality metric computed with the same ADMET-AI oracle that is embedded in the system's scoring displays, and one of the two RQ2 analyses uses an unpaired test on paired data. These issues are fixable but currently weaken the central quantitative contribution.

major comments (4)
  1. [§6.2.4, Table 1; §7.4; Abstract] RQ2 quality is measured as the number of four ADMET-AI-predicted properties improved relative to the starting lead. ADMET-AI is the same oracle that powers real-time property displays, cluster coloring, and strategy feedback in both HALO and the baseline (§5, §A.5). Participants can thus hill-climb on ADMET-AI's scoring function more effectively in HALO, and the 'higher-quality molecules' claim does not follow if ADMET-AI diverges from wet-lab pharmacology. §7.4 acknowledges this limitation, yet the abstract and conclusion retain unqualified 'higher-quality' wording. Please soften the claim to 'higher predicted ADMET-AI property scores' or provide external validation with an independent oracle or experimental data.
  2. [§6.2.4, Table 1] The quality comparison uses the paired Wilcoxon signed-rank test, which is appropriate for the within-subjects design. The diversity comparison, however, uses the unpaired Wilcoxon rank-sum test, which assumes independent samples. Because the same 10 participants produced both conditions, the paired signed-rank test should be used. Please re-analyze with the correct paired test and report exact p-values and an effect size (e.g., matched-pairs rank-biserial correlation). The reported diversity result may change substantially.
  3. [§6.1.3, Figure 4] Twelve Likert items are analyzed individually at α=0.05 without any correction for multiple comparisons. With N=10, the probability of at least one spurious significant result across 12 tests is non-negligible. Please apply a multiplicity correction (e.g., Holm-Bonferroni) or pre-specify a small set of composite outcomes, and report effect sizes. The consistent unidirectional pattern across all 12 items is reassuring, but the current reporting overstates the strength of evidence for the RQ1 claims.
  4. [§2.3 vs §4.1] Section 2.3 states that the formative study was conducted with 'ten medicinal chemists,' but Section 4.1 reports five participants (Table 4 also lists only five formative-study participants). This is a factual inconsistency that should be corrected; it affects the reader's understanding of the design rationale.
minor comments (5)
  1. [Abstract] The phrase 'we present a co-abduction' should read 'we present co-abduction' (unnecessary article).
  2. [Table 1] The row labeled 'Diversity' reports average pairwise Tanimoto similarity; lower values indicate higher diversity. Consider renaming it 'Average Pairwise Similarity (lower = more diverse)' or reporting diversity as 1−similarity to avoid confusion.
  3. [Figure 5] The legend uses similar shapes/colors for MolCluster, MolStrategy, and MolSynthesis. Please increase visual distinction and consider annotating example rows with participant IDs for readability.
  4. [§6.1.2] 'Both corresponding interfaces' is awkward; rephrase to 'both interfaces' or 'the HALO and baseline interfaces.'
  5. [§A.2] Task B has a high hallucination rate (29.7–39.8% RDKit parsing failures). The paper reports this, but it would be useful to state whether the user study pipeline filtered these outputs and whether the difference in generation reliability between Task A and Task B affected the user experience.

Circularity Check

0 steps flagged

No load-bearing circularity; RQ2's ADMET-AI-based quality metric is an acknowledged surrogate-oracle limitation, not a derivation that reduces to its own inputs.

full rationale

The paper contains no mathematical derivation, no fitted-parameter chain, and no invocation of a uniqueness theorem; the core claims rest on a user study with expert ratings, log analysis, and an oracle-based quality surrogate. The nearest circularity-adjacent point is RQ2: quality is measured by counting how many of four properties improve according to ADMET-AI ("For quality, we evaluated how many of the four properties were improved, using ADMET-AI [103], which is also used in HALO (Δ=after−before)", §6.2.4), and ADMET-AI is also the model behind property displays in both HALO and the baseline (§A.5). This means the headline 'higher-quality molecules' claim is limited by the oracle's fidelity, but it is not a construction-level reduction: no parameter is fitted from the outcome and then reported as a prediction, and the baseline also exposes ADMET-AI scores, so the comparison retains its internal meaning as an interface comparison. The authors explicitly flag this limitation in §7.4: "evaluating the quality of compounds generated by HALO solely using model-based scores and counts is inherently limited. Definitive assessment requires scientific validation through wet-lab experiments." The self-citations in the paper (e.g., [53], [66]) are illustrative and not load-bearing for the central claim. No circular step can be exhibited from the paper's own equations or definitions.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 1 invented entities

The paper introduces no fitted mathematical parameters or physical entities. The hand-chosen thresholds and pipeline heuristics shape the candidate space. The empirical claims rest on domain assumptions: abductive reasoning as the right model of hypothesis generation, ADMET-AI as a surrogate for pharmacological quality, LLM output reliability, and self-report as evidence of reasoning. These are reasonable for a CHI-style study but not independently established.

free parameters (3)
  • MCS coverage threshold = 0.6
    Hand-chosen cutoff in Appendix A.1 for retaining candidates whose maximum common substructure with the source covers at least 60% of heavy atoms; directly shapes the candidate pool users see.
  • Top-k source/strategy selection = top 3 sources; up to 3 cores
    Appendix A.6: candidates are ranked by Pareto criteria and the top three chosen as sources; greedy set-cover selects up to three cores. These heuristics are design choices, not derived from theory.
  • User study protocol constants = 30-min task; 7 submitted molecules; N=10; 12 Likert items
    §6.1: These procedural constants constrain statistical power and generalizability; the 12 questionnaire items are analyzed separately without multiple-comparison correction.
axioms (5)
  • domain assumption Abductive reasoning, with stages observation → pattern identification → hypothesis generation, is an accurate model of scientific hypothesis generation.
    Invoked throughout (§2.3, §4.3); if this cognitive model is wrong, the co-abduction framework loses its foundation.
  • domain assumption ADMET-AI predicted property scores are a valid proxy for real pharmacological properties.
    Quality metric uses ADMET-AI counts (§6.2.4); the authors acknowledge wet-lab validation is needed (§7.4).
  • domain assumption LLM-generated cluster explanations and strategies are chemically plausible enough to support reasoning.
    LLM outputs are post-processed only for SMILES validity (Appendix A.6); no independent chemical verification is reported.
  • domain assumption Participants' Likert ratings and self-reported 'aha' clicks measure actual abductive reasoning.
    RQ1 and RQ3 rely on self-report and log clicks (§6.1.3); no objective cognitive measure is used.
  • domain assumption The fine-tuned Llama generation model produces sufficiently valid, scaffold-preserving candidates during user sessions.
    Offline evaluation shows 0% hallucination for Task A but 29.7–39.8% for Task B (Table 2); in-session validity may vary by task.
invented entities (1)
  • co-abduction framework no independent evidence
    purpose: Conceptual model claiming that humans and AI jointly progress through clustering candidates, identifying strategies, and synthesizing strategies to form hypotheses.
    Central conceptual contribution; supported only by the internal user study and qualitative quotes. The promised transfer to materials/protein design (§7.1) is an untested prediction, so there is no independent falsifiable handle yet.

pith-pipeline@v1.3.0-alltime-deepseek · 26560 in / 14151 out tokens · 151666 ms · 2026-08-01T15:00:26.905962+00:00 · methodology

0 comments
read the original abstract

Scientific discovery is essential yet inefficient, primarily because generating hypotheses within a vast search space hinders breakthroughs. While current AI systems assist in generating new hypothesis candidates, they lack interactive support for the reasoning process by which users develop these outputs into promising hypotheses, resulting in surface-level hypotheses. To address this issue, we present co-abduction, a human-AI collaborative framework for abductive reasoning in scientific hypothesis generation. To operationalize co-abduction, we build HALO, a human-AI collaborative system for molecular hypothesis generation in drug discovery, enabling improved candidate clustering, strategy identification, and multi-strategy synthesis. In expert studies involving 10 medicinal chemists, HALO significantly facilitated abductive reasoning for hypothesis generation -- efficient candidate observation, systematic strategy identification, and coherent multi-strategy composition -- and enabled participants to produce higher-quality, more diverse candidate molecules.

Figures

Figures reproduced from arXiv: 2607.18564 by JiaSyuan Chang, Kat Limqueco, Xiang 'Anthony' Chen, Youngseung Jeon.

Figure 1
Figure 1. Figure 1: HALO is a human-AI collaborative system that assists medicinal chemists in abductive reasoning to generate promising molecular hypotheses in drug discovery. HALO supports abductive reasoning through three stages: (1) Improved-candidate Cluster, experts and AI generate molecular candidates and cluster them by property improvement profiles (green: augmented, red: worsened); (2) Strategy Identification, a lar… view at source ↗
Figure 2
Figure 2. Figure 2: Description of key terminologies for compound [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the HALO workflow: (1) Improved Candidates Cluster, (2) Strategy Identification, and (3) Multi-strategy Synthesis. Improved Candidates Cluster (A, B) generated molecules and organized them into property-based clusters to support efficient observation. Strategy Identification (C) extracts and summarizes intra-cluster optimization strategies from candidate groups. Multi-strategy Synthesis (D, E) … view at source ↗
Figure 4
Figure 4. Figure 4: Bar plots showing the results of the user study. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Each row illustrates a participant’s component [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: (a) Core scaffold preservation rate across constraint [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Overview of the compound generation pipeline. [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Structural similarity distributions between seed and [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Overview of the HALO pipeline. A three-stage workflow: (1) Observation, generate and cluster molecules by properties; [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Two tabs of MolStrategy. The overview tab displays an AI-generated summary of the cluster alongside its identified strengths and weaknesses. A scaffold image shows the shared core structure identified across cluster members, and a top candidates section highlights the highest-ranked molecules based on Pareto evaluation. The candidates tab lists all cluster members in a sortable table, where users can togg… view at source ↗
Figure 11
Figure 11. Figure 11: The baseline interface used for user studies. The Molecule Editor and Molecule Generator appear as pop-up windows [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: The initial lead compounds and the desired properties of Task-A and Task-B. Four representative cases illustrate [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: The prompts used to generate molecules through Llama 3.1 8B-Instruct. [PITH_FULL_IMAGE:figures/full_fig_p022_13.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

135 extracted references · 7 linked inside Pith

  1. [1]

    [n. d.]. Nvida - bionemo. https://www.nvidia.com/en-us/clara/bionemo/. Ac- cessed: 2024-07-31

    2024

  2. [2]

    Atilla Kaan Alkan, Shashwat Sourav, Maja Jablonska, Simone Astarita, Rishabh Chakrabarty, Nikhil Garuda, Pranav Khetarpal, Maciej Pióro, Dimitrios Tanog- lidis, Kartheik G Iyer, et al . 2025. A Survey on Hypothesis Generation for Scientific Discovery in the Era of Large Language Models.arXiv preprint arXiv:2504.05496(2025)

    Pith/arXiv arXiv 2025

  3. [3]

    Dana F AlKharboush, Frank Kozielski, Geoffrey Wells, and Exequiel OJ Porta

  4. [4]

    Shm Garanganao Almeda, JD Zamfirescu-Pereira, Kyu Won Kim, Pradeep Mani Rathnam, and Bjoern Hartmann. 2024. Prompting for discovery: Flexible sense-making for ai art-making with dreamsheets. InProceedings of the 2024 CHI Conference on Human Factors in Computing Systems. 1–17

    2024

  5. [5]

    Saleema Amershi, Maya Cakmak, W Bradley Knox, and Todd Kulesza. 2014. Power to the people: The role of humans in interactive machine learning.AI magazine35, 4 (2014), 105–120

    2014

  6. [6]

    Saleema Amershi, Dan Weld, Mihaela Vorvoreanu, Adam Fourney, Besmira Nushi, Penny Collisson, Jina Suh, Shamsi Iqbal, Paul N Bennett, Kori Inkpen, et al. 2019. Guidelines for human-AI interaction. InProceedings of the 2019 chi conference on human factors in computing systems. 1–13

    2019

  7. [7]

    Ian Arawjo, Chelse Swoopes, Priyan Vaithilingam, Martin Wattenberg, and Elena L Glassman. 2024. Chainforge: A visual toolkit for prompt engineering and llm hypothesis testing. InProceedings of the 2024 CHI Conference on Human Factors in Computing Systems. 1–18

    2024

  8. [8]

    Sara Barrena and Jaime Nubiola. 2019. Abduction: The Logic of Creativity. In The Bloomsbury Companion to Contemporary Peircean Semiotics. Bloomsbury, 185–203

    2019

  9. [9]

    Berman et al

    Helen M. Berman et al. 2000. The Protein Data Bank.Nucleic Acids Research28, 1 (2000), 235–242

    2000

  10. [10]

    G Richard Bickerton, Gaia V Paolini, Jérémy Besnard, Sorel Muresan, and An- drew L Hopkins. 2012. Quantifying the chemical beauty of drugs.Nature chemistry4, 2 (2012), 90–98

    2012

  11. [11]

    Camille Bilodeau, Wengong Jin, Tommi Jaakkola, Regina Barzilay, and Klavs F Jensen. 2022. Generative models for molecular discovery: Recent advances and challenges.Wiley Interdisciplinary Reviews: Computational Molecular Science12, 5 (2022), e1608

    2022

  12. [12]

    Jesse D Bloom, Sy T Labthavikul, Christopher R Otey, and Frances H Arnold

  13. [13]

    Paul E Blower Jr, Kevin P Cross, Michael A Fligner, Glenn J Myatt, Joseph S Verducci, and Chihae Yang. 2004. Systematic analysis of large screening sets in drug discovery.Current Drug Discovery Technologies1, 1 (2004), 37–47

    2004

  14. [14]

    Keith T Butler, Daniel W Davies, Hugh Cartwright, Olexandr Isayev, and Aron Walsh. 2018. Machine learning for molecular and materials science.Nature559, 7715 (2018), 547–555

    2018

  15. [15]

    Wesley Carpenter. 2019. The aha! moment: The science behind creative insights. InToward super-creativity-improving creativity in humans, machines, and human- machine collaborations. IntechOpen

    2019

  16. [16]

    Fengming Chen, Ranran Zhao, Xingxing Han, Huan Li, and Zhishu Tang. 2026. Harnessing AI to fuse phenotypic signatures for drug target identification: progress in computational modeling.Briefings in Bioinformatics27, 1 (2026), bbag045

    2026

  17. [17]

    Minjun Chen, Jürgen Borlak, and Weida Tong. 2013. High lipophilicity and high daily dose of oral medications are associated with significant risk for drug-induced liver injury.Hepatology58, 1 (2013), 388–396

    2013

  18. [18]

    Minjun Chen, Ayako Suzuki, Jürgen Borlak, Raúl J Andrade, and M Isabel Lucena

  19. [19]

    Ziqi Chen, Martin Renqiang Min, Srinivasan Parthasarathy, and Xia Ning. 2021. A deep generative model for molecule optimization via one fragment modifica- tion.Nature machine intelligence3, 12 (2021), 1040–1049

    2021

  20. [20]

    Lydia B Chilton, Juho Kim, Paul André, Felicia Cordeiro, James A Landay, Daniel S Weld, Steven P Dow, Robert C Miller, and Haoqi Zhang. 2014. Frenzy: collaborative data organization for creating conference sessions. InProceedings of the SIGCHI Conference on Human Factors in Computing Systems. 1255–1264

    2014

  21. [21]

    Lydia B Chilton, Greg Little, Darren Edge, Daniel S Weld, and James A Landay

  22. [22]

    José A Cisneros, Michael J Robertson, Brandon Q Mercado, and William L Jorgensen. 2017. Systematic study of effects of structural modifications on the aqueous solubility of drug-like molecules.ACS medicinal chemistry letters8, 1 (2017), 124–127

    2017

  23. [23]

    Victoria Clarke, Virginia Braun, and Nikki Hayfield. 2015. Thematic analysis. Qualitative Psychology: A Practical Guide to Research Methods(2015), 222–248

    2015

  24. [24]

    Kai-Hendrik Cohrs, Emiliano Diaz, Vasileios Sitokonstantinou, Gherardo Varando, and Gustau Camps-Valls. 2025. Large language models for causal hypothesis generation in science.Machine Learning: Science and Technology6, 1 (2025), 013001

    2025

  25. [25]

    The UniProt Consortium. 2023. UniProt: the Universal Protein Knowledgebase in 2023.Nucleic Acids Research51, D1 (2023), D523–D531

    2023

  26. [26]

    Gabriele Corso, Hannes Stärk, Bowen Jing, Regina Barzilay, and Tommi Jaakkola

  27. [27]

    Stefano Curtarolo et al . 2012. AFLOW: an automatic framework for high- throughput materials discovery.Computational Materials Science58 (2012), 218–226

    2012

  28. [28]

    Daniel W Davies, Keith T Butler, Adam J Jackson, Andrew Morris, Jarvist M Frost, Jonathan M Skelton, and Aron Walsh. 2016. Computational screening of all stoichiometric inorganic materials.Chem1, 4 (2016), 617–627

    2016

  29. [29]

    Vishal Dey, Xiao Hu, and Xia Ning. 2025. GeLLM 3O: Generalizing Large Language Models for Multi-property Molecule Optimization.arXiv preprint arXiv:2502.13398(2025)

    Pith/arXiv arXiv 2025

  30. [30]

    Vishal Dey, Xiao Hu, and Xia Ning. 2025. Large language models for controllable multi-property multi-objective molecule optimization.molecules1, 331,586 (2025), 433–166

    2025

  31. [31]

    The Diagram is like Guardrails

    Zijian Ding, Michelle Brachman, Joel Chan, and Werner Geyer. 2025. " The Diagram is like Guardrails": Structuring GenAI-assisted Hypotheses Exploration with an Interactive Shared Representation. InProceedings of the 2025 Conference on Creativity and Cognition. 606–625

    2025

  32. [32]

    Chris Donahue and Yewon Kim. 2025. AMUSE: AI-Integrated Songwriting Workflows. InProceedings of the CHI Conference on Human Factors in Computing Systems

    2025

  33. [33]

    Jurgen Drews. 2000. Drug discovery: a historical perspective.science287, 5460 (2000), 1960–1964

    2000

  34. [34]

    David TF Dryden, Andrew R Thomson, and John H White. 2008. How much of protein sequence space has been explored by life on Earth?Journal of The Royal Society Interface5, 25 (2008), 953–956

    2008

  35. [35]

    Kevin Dunbar. 1993. Concept discovery in a scientific domain.Cognitive science 17, 3 (1993), 397–434

    1993

  36. [36]

    Kevin Dunbar. 1995. How scientists really reason: Scientific reasoning in real- world laboratories. (1995)

    1995

  37. [37]

    Anna Fedor, Eörs Szathmáry, and Michael Öllinger. 2015. Problem solving stages in the five square problem.Frontiers in psychology6 (2015), 1050

    2015

  38. [38]

    Jacob G Foster, Andrey Rzhetsky, and James A Evans. 2015. Tradition and innovation in scientists’ research strategies.American sociological review80, 5 (2015), 875–908

    2015

  39. [39]

    Jie Gao, Yuchen Guo, Gionnieve Lim, Tianqin Zhang, Zheng Zhang, Toby Jia-Jun Li, and Simon Tangi Perrault. 2024. CollabCoder: A Lower-barrier, Rigorous Workflow for Inductive Collaborative Qualitative Analysis with Large Language Models. InProceedings of the 2024 CHI Conference on Human Factors in Computing Systems. 1–29

    2024

  40. [40]

    Massimo Garbuio and Nidthida Lin. 2021. Innovative idea generation in prob- lem finding: Abductive reasoning, cognitive impediments, and the promise of artificial intelligence.Journal of Product Innovation Management38, 6 (2021), 701–725

    2021

  41. [41]

    Anna Gaulton et al. 2012. ChEMBL: a large-scale bioactivity database for drug discovery.Nucleic Acids Research40, D1 (2012), D1100–D1107

    2012

  42. [42]

    Katy Ilonka Gero, Chelse Swoopes, Ziwei Gu, Jonathan K Kummerfeld, and Elena L Glassman. 2024. Supporting sensemaking of large language model outputs at scale. InProceedings of the 2024 CHI Conference on Human Factors in Computing Systems. 1–21

    2024

  43. [43]

    Elena L Glassman, Jeremy Scott, Rishabh Singh, Philip J Guo, and Robert C Miller

  44. [44]

    Ziwei Gu, Joyce Zhou, Ning-Er Lei, Jonathan K Kummerfeld, Mahmood Jasim, Narges Mahyar, and Elena L Glassman. 2025. AbstractExplorer: Leveraging Structure-Mapping Theory to Enhance Comparative Close Reading at Scale. In Proceedings of the 38th Annual ACM Symposium on User Interface Software and Technology. 1–25

    2025

  45. [45]

    Zhangxu He, Weiguang Yang, Feifei Yang, Jingyu Zhang, and Liying Ma. 2024. In- novative medicinal chemistry strategies for enhancing drug solubility.European Journal of Medicinal Chemistry279 (2024), 116842

    2024

  46. [46]

    Izumi V Hinkson, Benjamin Madej, and Eric A Stahlberg. 2020. Accelerating therapeutics for opportunities in medicine: a paradigm shift in drug discovery. Frontiers in pharmacology11 (2020), 770

    2020

  47. [47]

    David Hitchcock. 1979. Deductive and inductive: Types of validity, not types of argument.Informal Logic2, 3 (1979)

    1979

  48. [48]

    Samuel C Hoffman, Vijil Chenthamarakshan, Kahini Wadhawan, Pin-Yu Chen, and Payel Das. 2022. Optimizing molecules using efficient queries from property evaluations.Nature Machine Intelligence4, 1 (2022), 21–31. Conference acronym ’XX, June 03–05, 2025, Woodstock, NY Jeon et al

    2022

  49. [49]

    OverCode: Visualizing variation in student solutions to programming problems at scale.ACM Transactions on Computer-Human Interaction (TOCHI) 22, 2 (2015), 1–35

    2015

  50. [50]

    Ross Irwin, Spyridon Dimitriadis, Jiazhen He, and Esben Jannik Bjerrum. 2022. Chemformer: a pre-trained transformer for computational chemistry.Machine Learning: Science and Technology3, 1 (2022), 015022

    2022

  51. [51]

    Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand Ceder, et al. 2013. Commentary: The Materials Project: A materials genome approach to accelerating materials innovation.APL materials1, 1 (2013)

    2013

  52. [52]

    Pouria Jandaghi, Hamed S Najafabadi, Andrea S Bauer, Andreas I Papadakis, Matteo Fassan, Anita Hall, Anie Monast, Magnus von Knebel Doeberitz, John P Neoptolemos, Eithne Costello, et al. 2016. Expression of DRD2 is increased in human pancreatic ductal adenocarcinoma and inhibitors slow tumor growth in mice.Gastroenterology151, 6 (2016), 1218–1231

    2016

  53. [53]

    Youngseung Jeon, Matthew K Hong, Yan-Ying Chen, Kalani Murakami, Jonathan Q Li, Xiang Anthony Chen, and Matthew Klenk. 2024. Weaving ml with human aesthetic assessments to augment design space exploration: An automotive wheel design case study. InExtended Abstracts of the CHI Conference on Human Factors in Computing Systems. 1–10

    2024

  54. [54]

    Tingting Jiang, Zhumo Sun, Shiting Fu, and Yan Lv. 2024. Human-AI inter- action research agenda: A user-centered perspective.Data and Information Management8, 4 (2024), 100078

    2024

  55. [55]

    Tom Hosack, Djamil Damry, and Sujata Biswas. 2023. Drug-induced liver injury: a comprehensive review.Therapeutic advances in gastroenterology16 (2023), 17562848231163410

    2023

  56. [56]

    Hyeonsu B Kang, Nouran Soliman, Matt Latzke, Joseph Chee Chang, and Jonathan Bragg. 2023. ComLittee: Literature Discovery with Personal Elected Author Committees. InProceedings of the 2023 CHI Conference on Human Factors in Computing Systems. 1–20

    2023

  57. [57]

    Abdul Karim, Matthew Lee, Thomas Balle, and Abdul Sattar. 2021. CardioTox net: a robust predictor for hERG channel blockade based on deep learning meta-feature ensembles.Journal of Cheminformatics13, 1 (2021), 60

    2021

  58. [58]

    Nabin Khanal, Chun Meng Yu, Jui-Cheng Chiu, Anav Chaudhary, Ziyue Zhang, Kakani Katija, and Angus G Forbes. 2024. FathomGPT: A natural language interface for interactively exploring ocean science data. InProceedings of the 37th Annual ACM Symposium on User Interface Software and Technology. 1–15

    2024

  59. [59]

    Juho Kim, Haoqi Zhang, Paul André, Lydia B Chilton, Wendy Mackay, Michel Beaudouin-Lafon, Robert C Miller, and Steven P Dow. 2013. Cobi: A community- informed conference scheduling tool. InProceedings of the 26th annual ACM symposium on User interface software and technology. 173–182

    2013

  60. [60]

    Sunghwan Kim et al. 2021. PubChem in 2021: new data content and improved web interfaces.Nucleic Acids Research49, D1 (2021), D1388–D1395

    2021

  61. [61]

    John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. 2021. Highly accurate protein structure prediction with AlphaFold.nature596, 7873 (2021), 583–589

    2021

  62. [62]

    David Klahr and Kevin Dunbar. 1988. Dual space search during scientific reasoning.Cognitive science12, 1 (1988), 1–48

    1988

  63. [63]

    Paris Koloveas, Serafeim Chatzopoulos, Dionysis Diamantis, Christos Try- fonopoulos, and Thanasis Vergoulis. 2025. Accelerating Scientific Discovery with Multi-Document Summarization of Impact-Ranked Papers.arXiv preprint arXiv:2508.03962(2025)

    Pith/arXiv arXiv 2025

  64. [64]

    Ehud Kroll and Lauri Koskela. 2017. Studying design abduction in the context of novelty. InDS 87-7 Proceedings of the 21st International Conference on Engineering Design (ICED 17) Vol 7: Design Theory and Research Methodology, Vancouver, Canada, 21-25.08. 2017. The Design Society, 61–70

    2017

  65. [65]

    Yoonjoo Lee, Hyeonsu B Kang, Matt Latzke, Juho Kim, Jonathan Bragg, Joseph Chee Chang, and Pao Siangliulue. 2024. PaperWeaver: Enriching Topi- cal Paper Alerts by Contextualizing Recommended Papers with User-collected Papers. InProceedings of the CHI Conference on Human Factors in Computing Systems. 1–19

    2024

  66. [66]

    Ziwen Li, Xiang Chen, and Youngseung Jeon. 2025. GraPPI: a retrieve-divide- solve GraphRAG framework for large-scale protein-protein interaction explo- ration. InFindings of the Association for Computational Linguistics: NAACL 2025. 3635–3648

    2025

  67. [67]

    Douglas B Kitchen, Hélène Decornez, John R Furr, and Jürgen Bajorath. 2004. Docking and scoring in virtual screening for drug discovery: methods and applications.Nature reviews Drug discovery3, 11 (2004), 935–949

    2004

  68. [68]

    Yiren Liu, Pranav Sharma, Mehul Oswal, Haijun Xia, and Yun Huang. 2025. PersonaFlow: Designing LLM-Simulated Expert Perspectives for Enhanced Research Ideation. InProceedings of the 2025 ACM Designing Interactive Systems Conference. 1–29

    2025

  69. [69]

    Ricardo Macarron, Martyn N Banks, Dejan Bojanic, David J Burns, Dragan A Cirovic, Tina Garyantes, Darren VS Green, Robert P Hertzberg, William P Janzen, Jeff W Paslay, et al. 2011. Impact of high-throughput screening in biomedical research.Nature reviews Drug discovery10, 3 (2011), 188–195

    2011

  70. [70]

    Gerald Maggiora, Martin Vogt, Dagmar Stumpfe, and Jurgen Bajorath. 2014. Molecular similarity in medicinal chemistry: miniperspective.Journal of medic- inal chemistry57, 8 (2014), 3186–3204

    2014

  71. [71]

    Lorenzo Magnani. 2009. Creative abduction and hypothesis withdrawal. In Models of discovery and creativity. Springer, 95–126

    2009

  72. [72]

    Nikolas Martelaro, David Chuan-En Lin, and Hyeonsu B. Kang. 2025. Inkspire: A Sketch-Driven AI Tool for Design Ideation. InProceedings of the CHI Conference on Human Factors in Computing Systems

    2025

  73. [73]

    Junyong Lin, Lu Dai, Ruiqian Han, Yijie Sui, Ruilin Wang, Xingliang Sun, Qinglin Wu, Min Feng, Hao Liu, and Hui Xiong. 2025. ScIRGen: Synthesize Realistic and Large-Scale RAG Dataset for Scientific Research. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 5619–5630

    2025

  74. [74]

    Gordon E Moore et al . 1965. Cramming more components onto integrated circuits

    1965

  75. [75]

    David F Nippa, Kenneth Atz, Yannick Stenzhorn, Alex T Müller, Andreas Tosstorff, Jörg Benz, Hayley Binch, Markus Bürkler, Achi Haider, Dominik Heer, et al. 2025. Expediting hit-to-lead progression in drug discovery through reac- tion prediction and multi-dimensional optimization.Nature Communications (2025)

    2025

  76. [76]

    Stellan Ohlsson. 1992. Information-processing explanations of insight and related phenomena.Advances in the psychology of thinking1 (1992), 1–44

    1992

  77. [77]

    Charles O’Neill, Tirthankar Ghosal, Roberta Răileanu, Mike Walmsley, Thang Bui, Kevin Schawinski, and Ioana Ciucă. 2025. Sparks of science: Hypothesis generation using structured paper data.arXiv preprint arXiv:2504.12976(2025)

    Pith/arXiv arXiv 2025

  78. [78]

    Antonio J Osuna-Mascaró and Alice MI Auersperg. 2021. Current understanding of the “insight” phenomenon across disciplines.Frontiers in Psychology12 (2021), 791398

    2021

  79. [79]

    Kristin McEuen, Jürgen Borlak, Weida Tong, and Minjun Chen. 2017. Associa- tions of drug lipophilicity and extent of metabolism with drug-induced liver injury.International journal of molecular sciences18, 7 (2017), 1335

    2017

  80. [80]

    1934.Collected papers of charles sanders peirce

    Charles Sanders Peirce. 1934.Collected papers of charles sanders peirce. Vol. 5. Harvard University Press

    1934

Showing first 80 references.