Dynamic language model representations for multi-objective reaction optimisation
Joshua W. Sin, David Ming Segura, Bojana Ranković, Siu Lun Chau, Marius D. R. Lutz, Andrea Anelli, Ryan P. Burwood, Kurt Püntener, et al.
cs.LG
Quick verdict
Joint LLM-GP text featurisation speeds multi-objective reaction BO; strong benchmarks and two prospective HTE wins, yet no baseline arm ran prospectively.
Summary
The paper introduces Alice, a multi-objective Bayesian optimisation framework in which textual descriptions of reaction conditions are encoded by a LoRA-fine-tuned T5-base language model, passed through objective-specific projection heads into independent Gaussian processes, and trained end-to-end by joint marginal log-likelihood maximisation, so representations adapt as data accumulate. Benchmarked against Kraken/COSMO-RS descriptor baselines (bases and precursors one-hot) and one-hot encoding within an identical qLogNParEGO loop over 20 random seeds, it reached 90%/95% hypervolume thresholds in fewer iterations: about 13 versus 26 and 24 iterations on a nickel-catalysed Suzuki dataset (384 experiments), about 65 versus 80 on a palladium Suzuki dataset (404 experiments), and about 1.2 versus 2.4 plate iterations in a 96-well virtual HTE benchmark (21,648 conditions). Two prospective campaigns — a palladium-catalysed cyanation (30,000-condition space) and a three-objective asymmetric hydrogenation (8,064 conditions) — used two 96-experiment plates each, giving gram-scale products in 94% isolated yield and 84% yield at 99.6% ee. Key limitations: prospective campaigns had no head-to-head featurisation baselines; cyanation round 2 was manually constrained to the safer cyanide source; comparisons lack significance testing; retrospective benchmarks partly reuse the authors' own published datasets.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Jointly fine-tuned LLM-GP representations learned from textual reaction descriptions reach 90%/95% hypervolume convergence in fewer experiments than Kraken/COSMO-RS descriptor baselines and one-hot encoding across sequential and parallel multi-objective BO benchmarks.
- A single learned text representation spans chemically heterogeneous components (mono- and bidentate phosphines; elemental, organic, and inorganic additives; Ir and Ru chiral catalyst families) without any descriptor engineering.
- Applied prospectively, the framework identified lower-hazard K4[Fe(CN)6]-based cyanation conditions (>99% conversion and selectivity) that scaled to 94% isolated yield at gram scale in two 96-well plates, under 1% of a 30,000-condition space.
- In a three-objective dynamic kinetic hydrogenation over 8,064 conditions, two plates sufficed to find conditions delivering the target syn alcohol in 84% isolated yield and 99.6% ee at 2.0 g scale.
Strengths
True prospective validation with scale-up
Two wet-lab campaigns (Sections 3.1-3.2) run on an automated HTE platform, each closing in two 96-well plates and translating directly to gram scale: 94% isolated yield for the cyanation at 5.0 g and 83.6% yield with 99.6% ee for the hydrogenation at 2.0 g, with full NMR characterisation in the Supplementary Information — rare rigour for an ML-methods paper.
Unified representation where descriptor libraries genuinely fail
The cyanation space spans 40 mono- and bidentate Pd catalysts plus additives from elemental Zn to NEt3 and KOAc, and the hydrogenation spans 32 chiral Ir/Ru complexes with disparate coordination geometries — precisely the heterogeneous component classes the paper argues (Section 1) fragment descriptor spaces.
Consistent benchmark advantage across regimes
Strong reproducibility posture
Honest disclosure of loop interventions and sensitivity
Weaknesses
Attribution gap in prospective claims
Manual mid-campaign intervention in the cyanation
Thin statistical inference
Circularity and dataset reuse in benchmarks
Untested robustness and calibration of the deep kernel
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This arXiv preprint (2609.11790, Roche Basel with EPFL LIAC) introduces Alice, a multi-objective Bayesian optimisation (BO) framework whose reaction representation is learned rather than engineered: textual condition descriptions (component-value pairs such as ligand, solvent, base) are encoded by a LoRA-fine-tuned language model, projected per objective into independent Gaussian processes, and trained end-to-end by joint marginal log-likelihood maximisation (Methods, LLM-based deep kernel GP). Three central claims: (1) dynamic LLM-GP representations reach 90%/95% hypervolume convergence in fewer experiments than descriptor libraries and one-hot encoding across sequential and parallel regimes; (2) two prospective wet-lab campaigns — Pd-catalysed cyanation and three-objective asymmetric hydrogenation — succeeded in two 96-well plates each, validated at gram scale (94% and 84% isolated yield, the latter at 99.6% ee); (3) one representation spans chemically heterogeneous components without feature engineering.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does the Alice framework do?
Alice runs multi-objective Bayesian optimisation of chemical reactions using reaction representations learned from text: each candidate condition is written as a short text description, embedded by a LoRA-fine-tuned T5-base language model, and fed through per-objective projection heads into Gaussian process surrogates that are trained jointly with the language model, so the representation improves as experimental data accumulate.
How many experiments did the prospective campaigns need, and what did they achieve?
Each campaign used two rounds of 96-well plate experiments (192 reactions total): the palladium-catalysed cyanation sampled under 1% of a 30,000-condition space and delivered gram-scale product in 94% isolated yield using the low-toxicity cyanide source K4[Fe(CN)6]; the three-objective asymmetric hydrogenation sampled about 2.4% of 8,064 conditions and delivered the target syn alcohol in 84% isolated yield with 99.6% enantiomeric excess at 2 g scale.
Did the algorithm alone choose the safer cyanide source?
No. In plate 1 the best hits all used the toxic sources NaCN or Zn(CN)2. The authors then manually fixed the cyanide source to the safer K4[Fe(CN)6] for plate 2 and re-optimised within that constraint; the framework then found multiple conditions exceeding 99% conversion and 99% selectivity.
Is the code and data available?
Yes. The code is on GitHub (schwallergroup/alice) under an Apache 2.0 licence, the retrospective benchmark datasets are stated to be open-source, and the complete design spaces for both prospective campaigns are given in Extended Data Figures 7 and 8, with all HTE data and product characterisations in the Supplementary Information.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Extended Data figures (bar charts, seed-percentage panels) were extracted as garbled layout text, so numeric values beyond those stated in prose (e.g., exact % of seeds reaching each threshold) could not be independently read or verified.
- Spectral listings in Supplementary Figures 2-5 are OCR-garbled; the supplementary experimental procedures and yields are, however, fully legible, and no prose content was missing from the extraction.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.