# Molecular Property Prediction

Intology extended its autonomous discovery system Locus from language-model post-training to molecular discovery, comparing it at matched compute against prior state-of-the-art methods from BioHub and Boltz and against frontier LLM agents Codex and Claude Code on molecular property prediction, small-molecule potency prediction and antibody design. On Boltz-2 adaptation across six targets, Locus reached a mean Spearman correlation of 0.622, against 0.225 for the Boltz-2.1 API, 0.450 for Claude Code and 0.569 for Codex. In antibody design, Locus reached comparable binding scores with 1.5–2.8× fewer generated candidates, but those scores are OpenDDE ipSAE proxies and still need wet-lab validation.

Intology extended its autonomous discovery system Locus from language-model post-training to molecular discovery, comparing it at matched compute against prior state-of-the-art methods from BioHub and Boltz and against frontier LLM agents Codex and Claude Code on molecular property prediction, small-molecule potency prediction and antibody design. On Boltz-2 adaptation across six targets, Locus reached a mean Spearman correlation of 0.622, against 0.225 for the Boltz-2.1 API, 0.450 for Claude Code and 0.569 for Codex. In antibody design, Locus reached comparable binding scores with 1.5–2.8× fewer generated candidates, but those scores are OpenDDE ipSAE proxies and still need wet-lab validation.

Large language models can now construct and keep revising scientific principles during a discovery run, rather than leaving that translation to human experts. Doing it by hand needs domain expertise, time and effort, and bottlenecks discovery. Principia compares related inputs and their observed outcomes to write causal hypotheses, clusters similar inputs to induce recurring principles, then revises them with new observations. Across molecular optimization, protein engineering, gene perturbation and molecular property prediction, it beat baselines using internal knowledge, experience summaries or web retrieval; on five molecular optimization tasks it averaged AUC@10 0.279 versus 0.250 for the experience-summary baseline.

Large language models can now construct and keep revising scientific principles during a discovery run, rather than leaving that translation to human experts. Doing it by hand needs domain expertise, time and effort, and bottlenecks discovery. Principia compares related inputs and their observed outcomes to write causal hypotheses, clusters similar inputs to induce recurring principles, then revises them with new observations. Across molecular optimization, protein engineering, gene perturbation and molecular property prediction, it beat baselines using internal knowledge, experience summaries or web retrieval; on five molecular optimization tasks it averaged AUC@10 0.279 versus 0.250 for the experience-summary baseline.

The study evaluates nine LLM variants across three families (GPT-4.1, GPT-5, Gemini 2.5) on three MoleculeNet datasets (Delaney solubility, Lipophilicity, QM7 atomization energy) using a systematic blinding procedure that iteratively reduces available information, complemented by 0-, 60-, and 1000-shot in-context sample sizes, in order to determine whether molecular property prediction reflects genuine in-context regression or verbatim retrieval of memorized target values, and it adds positive and negative controls for the memorization experiments, structural reference baselines for the multi-shot experiments, and bootstrap confidence intervals for all results; it finds no evidence of verbatim retrieval on these legacy benchmarks and shows that blinding exposes conflicts between pre-traine

The study evaluates nine LLM variants across three families (GPT-4.1, GPT-5, Gemini 2.5) on three MoleculeNet datasets (Delaney solubility, Lipophilicity, QM7 atomization energy) using a systematic blinding procedure that iteratively reduces available information, complemented by 0-, 60-, and 1000-shot in-context sample sizes, in order to determine whether molecular property prediction reflects genuine in-context regression or verbatim retrieval of memorized target values, and it adds positive and negative controls for the memorization experiments, structural reference baselines for the multi-shot experiments, and bootstrap confidence intervals for all results; it finds no evidence of verbatim retrieval on these legacy benchmarks and shows that blinding exposes conflicts between pre-traine

This study introduces a new avenue for foundation model pretraining: supervised pretraining on low-noise, calculable molecular descriptors to obtain rich, highly transferable molecular representations, demonstrated with CheMeleon, an O(10M) parameter foundation model; across 58 benchmark data sets spanning properties relevant to small-molecule drug discovery and sourced from the industry-led Polaris benchmarking initiative, CheMeleon enables directed message-passing neural networks to finally exceed classical methods in the low-data regime, outperforms classical baselines such as Random Forest on molecular fingerprints and descriptors as well as existing foundation models under rigorous statistical comparisons, and the model and pretraining framework are open-sourced.

This study introduces a new avenue for foundation model pretraining: supervised pretraining on low-noise, calculable molecular descriptors to obtain rich, highly transferable molecular representations, demonstrated with CheMeleon, an O(10M) parameter foundation model; across 58 benchmark data sets spanning properties relevant to small-molecule drug discovery and sourced from the industry-led Polaris benchmarking initiative, CheMeleon enables directed message-passing neural networks to finally exceed classical methods in the low-data regime, outperforms classical baselines such as Random Forest on molecular fingerprints and descriptors as well as existing foundation models under rigorous statistical comparisons, and the model and pretraining framework are open-sourced.

Using UV absorption wavelength (λmax) prediction as a testbed across 18,415 solute-solvent pairs (after the Greenman/Song duplicate-handling protocol) with stratified 5-fold cross-validation and two external data sets totaling over 40,000 molecules, this work compares five models spanning four architecture families (Random Forest, XGBoost, a directed message-passing GNN, a BiGRU, and a pretrained Transformer) and finds that Random Forest is statistically indistinguishable from the best deep learning model for screening within known chemical space (RMSE 31.50 ± 1.47 vs 33.15 ± 3.27 nm, p = 0.16) while training in 15 min on a CPU, whereas on 16 novel UV-absorbing compounds the GNN (MAE 26.2 nm), Transformer (26.3 nm), and BiGRU (28.6 nm) all outperform RF (38.

Using UV absorption wavelength (λmax) prediction as a testbed across 18,415 solute-solvent pairs (after the Greenman/Song duplicate-handling protocol) with stratified 5-fold cross-validation and two external data sets totaling over 40,000 molecules, this work compares five models spanning four architecture families (Random Forest, XGBoost, a directed message-passing GNN, a BiGRU, and a pretrained Transformer) and finds that Random Forest is statistically indistinguishable from the best deep learning model for screening within known chemical space (RMSE 31.50 ± 1.47 vs 33.15 ± 3.27 nm, p = 0.16) while training in 15 min on a CPU, whereas on 16 novel UV-absorbing compounds the GNN (MAE 26.2 nm), Transformer (26.3 nm), and BiGRU (28.6 nm) all outperform RF (38.

This study comprehensively benchmarked tabular foundation models (TFMs) based on in-context learning for small organic molecule property prediction, comparing several TFMs with multiple machine learning methods across 11 data sets (regression, random and structure-aware splits, up to 10,000 molecules each), and found that TFMs consistently outperform XGBoost, CatBoost, multilayer perceptrons, and other descriptor-based methods even with careful hyperparameter selection for the baselines, achieve accuracy on par with or better than graph-based methods including those pretrained on chemical data, with Uni-Mol2 slightly outperforming TFMs in some experiments, while retrieval (selecting the 500 closest neighbors by Tanimoto similarity for each test molecule) yields further improvement on relat

This study comprehensively benchmarked tabular foundation models (TFMs) based on in-context learning for small organic molecule property prediction, comparing several TFMs with multiple machine learning methods across 11 data sets (regression, random and structure-aware splits, up to 10,000 molecules each), and found that TFMs consistently outperform XGBoost, CatBoost, multilayer perceptrons, and other descriptor-based methods even with careful hyperparameter selection for the baselines, achieve accuracy on par with or better than graph-based methods including those pretrained on chemical data, with Uni-Mol2 slightly outperforming TFMs in some experiments, while retrieval (selecting the 500 closest neighbors by Tanimoto similarity for each test molecule) yields further improvement on relat