Artificial Intelligence in the Life Sciences

A compound's toxicity can be gauged from the toxicities of the compounds it resembles. We showed that in 2003 for rat oral LD50: the compounds in a reference set of 13,645 that most resemble the query form a consortium and vote on its toxicity. This work returns to that method twenty-three years later, on public data, because the reference set used in 2003 is no longer public. Rebuilt on 11,989 curated rat oral LD50 values, the method answers for 67.0% of the held-out compounds carrying an exact value, and 92.1% of those answers land within one log unit; in 2003 the figures were 94% within one log unit on 57.3% of compounds. We then ran it over 28 further endpoints, each recording whether a compound showed a particular toxicity, from cytochrome inhibition to mutagenicity to clinical toxicity: 149,224 measurements on 34,162 compounds. The method answers only when a compound has close enough neighbors, and on those it correctly identifies a median 72.9% of the compounds that showed the toxicity and 96.5% of those that did not. Separately we asked whether a reverse screen carries information about these endpoints beyond molecular size and lipophilicity. We search a compound against a published index of 27,910 ligands solved inside proteins and count the protein families those sites belong to. A compound whose matches reach no protein family inhibits CYP3A4 2.5% of the time; one whose matches span sixteen or more families inhibits 51.2% of the time. Added to heavy atom count, molecular weight and cLogP, the screen raises the mean area under the ROC curve from 0.72 to 0.76 across the 28 endpoints. The OpenADMET consortium has since measured CYP3A4 inhibition on 2,335 compounds, and the same pattern holds. Compounds that resemble known binders of sixteen or more protein families inhibit CYP3A4 24.6% of the time; compounds whose matches reach no protein family inhibit it 7.3% of the time. The more protein families a compound resembles binders of, the more likely it is to inhibit CYP3A4. This association is strong enough to act on before anything is made. Scoring a library on how far its compounds reach across protein families, together with their size and lipophilicity, and setting aside the riskiest fifth removes two thirds of the androgen receptor binders and about half of the CYP3A4 inhibitors. Scoring needs only a two-dimensional structure, with no docking and no conformer generation at query time. Interactive scoring, the data and the pool are at toxpred.ai. Keywords: toxicity prediction; structural similarity; nearest neighbors; reverse screening; target breadth; promiscuity; LD50; cytochrome P450; virtual library triage

Accurately predicting cellular responses to drugs remains a challenge with the potential to reduce experimental screening costs and accelerate drug discovery. Current computational approaches represent compounds through chemical structures, which carry little information regarding their activity within biological systems. We show that gene expression responses, measured in a reference cell line, can instead serve as transferable representations of chemical perturbations. The proposed framework, BioPert, uses biological representations with a small neural network to predict transcriptional delta responses in cell lines of different lineages. Across the Tahoe-100M and LINCS L1000 datasets, BioPert outperforms molecular fingerprints and embeddings learned from chemical structure, which provide limited improvements over random controls. BioPert's prediction correlations reach 0.79 on Tahoe-100M, corresponding to an improvement of 0.34 over the next-best representation. On LINCS, performance varies with the experimental reproducibility of the test conditions, highlighting the impact of batch effects. Notably, BioPert considerably surpasses predictions obtained by copying the reference response, demonstrating that it captures context-dependent effects. Further analysis of a C32-cobimetinib case shows pathway-level accuracy of the predictions, including when the reference and target responses differ. These results open a new path for chemical perturbation prediction and could ultimately reduce the burden of phenotypic screening.

Background: Recent work has brought large language models (LLMs) into clinical variant interpretation. In most existing approaches, the model is confined to text-level operations, retrieving, extracting, compacting, and summarizing heterogeneous evidence, while the genuine genetic evaluation is delegated to a separate deterministic engine. Other tools employ LLMs to mine functional evidence for a single ACMG criterion, to rerank a classifier's output, or to answer isolated variant questions. Integrative genetic reasoning, by contrast, demands multi-source evidence weighing for each variant, reconciliation of inheritance with observed zygosity and parental genotypes, joint interpretation of variants in the same gene, and the derivation of a case-level conclusion. Moreover, to protect patient privacy, this must be done on hospital premises. Methods: o build an end-to-end variant interpreter that runs on local infrastructure, we benchmarked 13 small language models (SLMs) on 150 multiple-choice genetics questions and selected Qwen3.5-9B. In IVA (Intelligent Variant Analyser), a four-stage pipeline, the model assigns ACMG/AMP criteria, applies de novo, dominant, recessive, and X-linked inheritance logic to trios, and writes a structured report, while deterministic code handles retrieval, phasing, point summation, citation checking, and formatting. We evaluated IVA on synthetic scenarios, 16 solved trios, and 327 solved singleton cases of visual impairment, on a single GPU. Results: On the benchmark, Qwen3.5-9B reached a mean accuracy of 0.81, compared with 0.89 for a 1.6-trillion-parameter cloud model. In the 16 trios, IVA reported every previously identified pathogenic, likely pathogenic, or uncertain-significance variant of interest (sensitivity 100%; precision 39.5%). In the 327 singleton cases, it recovered 93% of the expert-reported causative variants and classified 82% of them as pathogenic or likely pathogenic (precision 43%). Failures fell into four recurrent classes: omission, unverified mechanism, mismatch between reasoning and label, and arithmetic errors. Conclusions: A small language model running on local hardware can serve as the reasoning layer for case-level and trio-based variant interpretation when it is surrounded by deterministic safeguards. Because its outputs are not deterministic, analyst review and prospective multicenter validation remain necessary.

Scientific design requires reasoning in scientific spaces that represent the objects being designed. Building on Scinetics' Chain of Science Tokens, we present chain-of-structures (COS) as a concrete protein-design case study that connects reasoning and generation through explicit structural states. In this technical report, we apply COS to enzyme design, where catalytic function requires precise local geometry within a compatible global fold, making it a demanding test of reasoning from chemical requirements to protein structure. Starting from a reaction specification and target length, COS couples a language reasoning expert with an all-atom generation expert across four rounds: catalytic core, active site, global backbone, and complete sequence and structure. Each reasoning state directly conditions generation, and the resulting structure becomes context for the next decision. Discrete residue constraints and continuous latents connect the two experts, enabling structural supervision of the reasoning interface. On the evaluated benchmark, COS leads the compared baselines on five of six metrics, achieving 57.0% EC3 agreement and a TM-score of 0.729.

Large language model (LLM) agents are increasingly capable of carrying out autonomous computational research, but it remains unclear whether they can develop molecular modeling methods that compete with strong human-developed approaches. Here, we evaluate autonomous method development across four settings: Therapeutics Data Commons (TDC) ADMET tasks, the OpenADMET ExpansionRx Challenge, the activity prediction track of the OpenADMET PXR Induction Challenge, and the Practical Molecular Optimization (PMO) benchmark. Agents followed the prescribed data splits, metrics, and evaluation protocols for each benchmark. On TDC, the best agent-developed score across configurations improved on the leaderboard reference on 18 of 21 tasks and on the combined leaderboard and peer-reviewed reference on 17, while individual Codex configurations improved on the combined reference on 13 tasks. On ExpansionRx, five-agent and single-agent Codex configurations achieved overall scores corresponding to fourth and sixth place relative to the published final leaderboard, with the five-agent configuration reaching top-five performance on seven of nine endpoints. On PXR, five-agent Codex achieved a mean absolute error slightly outperforming the best published challenge entry. On PMO, five-agent Codex developed a molecular optimizer that outperformed existing methods evaluated under the benchmark's standard setup; other methods obtained higher reported scores under modified conditions, including additional molecular pretraining data or a different oracle-call budget. Together, these results show that current LLM agents can autonomously develop high-performing molecular prediction and optimization methods across substantially different drug-discovery settings, in several cases matching or exceeding leading human-developed approaches.

Accurate modelling of biomolecular interactions is fundamental to drug discovery, yet current artificial intelligence (AI) workflows remain fragmented across structure prediction, affinity estimation, molecular design, and experimental decision-making. We introduce AnewDDE, an agentic Drug Discovery Engine that connects these capabilities into a closed-loop system for biomolecular interaction modelling and design. We demonstrate several components of this system: AnewFold delivers superior performance on challenging targets, including antibody-antigen complexes and molecular glues, providing more reliable structural hypotheses together with pocket identification and conformational analysis; AnewDesign achieves a 10.7% success rate in identifying binders with single-digit-nanomolar affinities measured by surface plasmon resonance (SPR) in a representative nanobody discovery campaign, substantially reducing the experimental search burden; AnewAffinity combines accuracy approaching that of free-energy perturbation (FEP) methods with the speed required for high-throughput screening and prioritizes promising transformations; and AnewMind, a large language model (LLM) at the hundred-billion-parameter scale that underwent in-house full-parameter post-training, provides scientific reasoning and absorption, distribution, metabolism, excretion, and toxicity (ADMET) prediction for pharmaceutical research and development (R&D). To systematically evaluate AnewMind, we introduce PharmBench, an internal benchmark designed by pharmaceutical R&D experts to evaluate long-horizon decision-making in drug discovery. Evaluations on ADMET tasks and PharmBench show that AnewMind performs competitively with leading frontier models. Together, these capabilities establish AnewDDE as an integrated engine for addressing challenging therapeutic targets and translating molecular insights into experimentally validated candidates at scale.

Machine learning benchmarks often treat image files as independent observations, although multiple files may represent repeated views or other related samples. Prior work has shown that non-independent partitioning can inflate model performance, but the sensitivity of a biomedical benchmark to progressively increasing train-test relatedness is less often quantified. Here, we quantified how apparent classification performance changes when recoverable sample groups are progressively allowed to cross training, validation, and test partitions. Using source structure and filenames, we reconstructed analysis groups in a microscopy corpus compiled from multiple sources and generated five class stratified partitioning conditions, ranging from an intact group reference to random assignment of individual files. We applied the same training recipe within each model family to 129,173 images spanning 14 classes of blood cells and hematopoietic cell states. For a pretrained Vision Transformer, random assignment of individual files increased the mean test macro F1 from 0.853 to 0.977 and accuracy from 0.851 to 0.990 relative to the intact group reference. Apparent performance increased progressively across the intermediate conditions. The endpoint pattern also appeared with a ResNet-50 backbone, and the largest apparent gains occurred in immature granulocytic and blast classes. A descriptive analysis within the random assignment condition further showed higher scores among test images that had another member of the same group in the training set, although the exposed and unexposed subsets were not exchangeable. The grouping workflow, exact split definitions, code, and compact result files are released as HSD-Bench to support reproducible evaluation of hidden sample dependence.

Hit-to-lead optimization is a multi-objective molecular-design problem in which binding-related scores must be considered together with drug-likeness and synthetic accessibility. We developed a closed-loop language-model medicinal-chemistry agent that combines target and starting-hit context, PubMed retrieval, molecular-property tools, docking, and feedback from previously scored analogues. In a matched computational benchmark comprising six kinase targets, six methods, a 40-candidate budget, and six independent campaigns per method, the agent achieved the highest mean aggregate score under a pre-specified composite objective on four targets. Re-analysis of the same candidate sets with drug-likeness-gated docking and docking-drug-likeness hypervolume produced different target-level winners, showing that evaluation rules can change the comparative interpretation of an optimization campaign. Candidate distributions and component ablations indicated that the agent's advantage was associated primarily with retention of scored-candidate feedback and higher-QED regions rather than with docking alone. An EGFR case study illustrates a scaffold-preserving in-silico optimization trajectory from erlotinib. The study is a computational benchmark and evaluation framework; the prioritized molecules are hypotheses for subsequent experimental testing, not experimentally validated leads.

Agentic systems increasingly coordinate molecular-design tools, but it is unclear which layer of the stack limits outcomes on real projects. We developed MAGI, an open modular agent that authors objectives, launches and monitors optimization, interprets structure activity relationships, and revises its strategy accordingly. MAGI generates molecules either directly through the LLM or by delegating to REINVENT4, with scoring services interchangeable behind a common contract. We tested it across nine retrospective lead-optimization campaigns from three pharmaceutical companies, replayed under fixed temporal cutoffs. Both routes produced valid structures: LLM proposals stayed closer to local chemistry and reached comparable or higher primary activity in fewer operations, whereas REINVENT explored broader chemical space. Whether a campaign met its objective depended on the predictive models, not on the generation route: attainment followed model accuracy on the chemistry proposed, dropping once that chemistry moved outside the model applicability domain. Separately, a blinded evaluation asked whether the MAGI output could pass as expert work: chemists were not able to discriminate agentic proposals from held-out compounds, and judged the SAR reasoning broadly plausible yet incomplete. Together, these results position MAGI as a coordination layer pluggable into existing computational chemistry workflows. The ceiling on real projects, however, remains currently set by scorer applicability rather than by tool orchestration.

Membrane permeability remains a major challenge in cyclic-peptide drug discovery, where optimization often depends on estimating changes relative to measured reference compounds. Existing predictors mainly model absolute permeability, while heterogeneity across literature sources complicates pooled training and comparative evaluation. We present CycDelta, a multimodal residue-graph model trained directly on within-source permeability differences. CycDelta combines directed message passing with pretrained Uni-Mol monomer embeddings, physicochemical descriptors and assay information in a shared encoder for reference and query peptides. To evaluate this comparative formulation under controlled conditions, we establish CycDeltaBench, comprising in-distribution, scaffold-based out-of-distribution, external and zero-shot unseen-assay tests. On the out-of-distribution task, CycDelta achieved Pearson 0.69 versus 0.46 for the strongest baseline. It also led all evaluated methods on two external data sets and achieved zero-shot Spearman 0.70 versus 0.32 for the strongest baseline on an unseen assay. Residue-level attribution relates predictions to monomer contributions, while gradient-guided screening supports prioritization of permeability-enhancing analogue hypotheses. These results position CycDelta as a transferable and interpretable framework for comparative permeability prediction and cyclic-peptide optimization.

Within the ADMET continuum, metabolism and excretion (ME) form a critical bridge between drug exposure established by absorption and distribution and downstream efficacy and toxicity. However, no existing framework for drug ME prediction has simultaneously achieved broad coverage of endpoints and robust data recency and completeness. Here, we developed MetExPred, a multi-view prediction framework that comprises 17 classification endpoints and two regression endpoints, covering the overall drug ME process. The framework combines sequence-based and graph-based molecular representations, while optionally incorporating ESM-2 protein representations when experimentally annotated targets are available. A protein-aware masking strategy enables the same architecture to operate in both molecular-only and target-enhanced settings, and multi-view attention adaptively integrates the available representations. Across the classification tasks, MetExPred achieved mean AUROC, AUPRC and F1 scores of 0.844, 0.734 and 0.697, respectively. For clearance and half-life prediction, the model achieved RMSE values of 0.686 and 0.668. MetExPred showed the strongest average performance among the evaluated baselines, while ablation studies confirmed the complementary contributions of molecular sequence, graph and protein information. These results provide a unified and flexible modeling framework for systematic ME prediction of drug compounds, enabling early-stage virtual screening and pharmacokinetic assessment during lead optimization.

Volumetric microscopy produces datasets too large to share as files and too rich to convey as still images, yet exploring them interactively means installing desktop software or building web infrastructure. Here we present Luxar, an open-source Python framework that compiles multidimensional data into compact archives that any modern browser renders from static file hosting. Fluorescence volumes are represented as mixtures of anisotropic Gaussians (Gaussian splats): the image becomes geometry that the GPU draws directly, with no voxel grid to rebuild. Splats share one scene with points, trajectories and surfaces. Blind-spot cross-validation sets the number of Gaussians without noise-free references. Across 17 volumes from 12 datasets and four modalities, this yields a median 99-fold (6- to 340-fold) size reduction at 26 to 67 dB peak signal-to-noise ratio (PSNR), in minutes per volume on one GPU. We demonstrate Luxar on a 500-timepoint embryogenesis recording, super-resolution localisations, single-cell atlases and lineage tracking, each explorable through a shareable link.

The -Bonacci weighted sequence space {ell}(,) is introduced for [&ge;] 2, 1 [&le;] < {infty}, and > 0). It is proved to be a Banach space which is isometrically isomorphic to {ell}, with an unconditional Schauder basis. The space is also reflexive and uniformly convex. Its strict inclusion chain, continuous dual, and asymptotic truncation estimate of order - / are also established. Based on this framework, Scaled &#119898-BonacciNet is proposed as a lightweight CNN that rescales Fibonacci-Net channels using the dominant &#119898-Bonacci root and employs depth wise separable convolutions. Tribonacci-Net achieved 96.09% accuracy and a 0.9581 macro F1-score on the 15-class Brain MRI dataset using 0.85 million parameters and 0.366 GFLOPs. Pentanacci-Net attained 95.87% accuracy on the 44-class dataset while reducing parameters and FLOPs by 47.10% and 44.48%, respectively. On the CKD ultrasound dataset, Hexanacci-Net achieved 83.75% accuracy and a 0.7941 macro F1-score, with corresponding reductions of 50.37% and 45.10%. These findings support dominant-root scaling as a structured approach to lightweight CNN design.

Protein inverse folding, which infers compatible amino acid sequences from a given backbone structure, is central to rational protein design. Current methods rely largely on sequence recovery rate as the evaluation metric, emphasizing structural compatibility but seldom addressing function. Yet functional activity depends on the proper network of non-covalent interactions-protein-ligand contacts being a prime example. We argue that function-oriented protein design is governed by three complementary constraints: structural, interaction, and evolutionary, none of which are handled in a unified manner by existing pretrained models. To bridge this gap, we introduce zPolaron, a graph neural network built on protein-ligand complexes for functional protein design. zPolaron represents residues and ligand atoms as a heterogeneous graph and jointly encodes structural, interaction, and evolutionary information by fusing physical-interaction-enhanced graph representations with ESM-2 embeddings. Across multiple fitness prediction benchmarks, zPolaron outperforms all baselines, with gains evident across structural, interaction, and evolutionary dimensions. Wet-lab experiments further confirm its utility in functional enzyme screening and design: the optimized CalA6-D205A-S160G mutant achieves a 17.84-fold improvement in activity, while the CalB-I121M mutant yields a 1.91-fold increase. Together, these results show that integrating interaction and evolutionary information fills the missing puzzle of inverse folding, enhancing protein fitness prediction beyond structural compatibility. zPolaron is available at https://github.com/zelixirSH/zPolaron.

Recent advances in deep learning-based structure prediction, including AlphaFold3 and the open-source Boltz model family, have extended biomolecular modeling to joint prediction of protein-ligand, protein-nucleic acid, and multi-chain complexes with binding-affinity estimation. Boltz-2 is among the most feature-complete of these open models, but its practical use requires a local CUDA-capable GPU, command-line execution, and manually authored YAML configuration files, limiting accessibility for researchers without dedicated computational infrastructure. We developed Boltz2-Notebook, a Colab-native interface comprising four integrated stages - automated environment setup, interactive parameter-to-YAML generation, execution management, and automated confidence and affinity visualization - together with a manifest-driven batch mode for multi-target screening. All modelling capabilities are inherited unmodified from Boltz-2; Boltz2-Notebook's contributions are limited to accessibility, input construction, and workflow automation. Independent of the software, we curated a benchmark of 317 protein-ligand pairs (122 proteins, 277 ligands) from BindingDB and predicted binding affinity in triplicate using the Boltz-2 command-line engine on high-performance computing infrastructure. Predicted and experimental pIC50 values showed moderate correlation (Pearson r = 0.609 [95% CI 0.540-0.675]; Spearman {rho} = 0.625; R2 = 0.371; MAE = 0.968 pIC50 units), with high triplicate reproducibility (pairwise r = 0.97) but a systematic compression of the predicted affinity range and no measurable relationship between Boltz-2's self-reported confidence metrics and prediction accuracy. Boltz2-Notebook is freely available as open-source software and provides external, reproducible evidence - including a specific confidence-calibration limitation - relevant to interpreting Boltz-2 affinity predictions responsibly.

We present an automated pipeline for finding therapeutic interventions in channelopathies. It starts from patch clamp recordings of the variant channel, searches over the pharmacologically accessible conductances, and says which currents must change and by how much. The intervention never touches the mutated channel: it compensates by modulating others. The feasibility of this approach is demonstrated through two computational strategies, each addressing a specific capability gap. First, we employ high-fidelity NEURON simulations combined with exploratory search algorithms to model two clinically described variants, and identify "stability windows" in which channel modulations restore the reference firing pattern. The compensating configurations are degenerate: 37 conductance triples fire the same number of action potentials at every one of the 35 injected-current levels, and the 10,780 triples reaching the best intervention value produce only six distinct firing patterns. Second, to address the computational cost of such detailed modeling, we implement a differentiable Hodgkin-Huxley model in PyTorch. Applied here to theoretical mutation screening, this approach trades granular detail for speed, treating the search as a continuous optimization problem that makes very large, non-local searches of the conductance parameter space affordable: the differentiable forward model evaluates 34,000-64,000 candidate parameter sets per second on a single consumer graphics card, against 11.9-16.9 per second for the non-differentiable pipeline on a 72-core node. The gain arises from forward-model throughput feeding a search rather than from the gradient itself, and it is what makes the 32,000,000-point survey of the solution topology reported here affordable. These computational predictions can inform practical drug development and high-throughput screening by naming the currents that must change and by how much.

Molecular structures are abundant, while experimental bioactivity is sparse and distributed across assays. We introduce GRASP, a 93.5M-parameter graph Transformer that learns from these sources in sequence. GRASP first learns molecular structure through replaced-token detection on 1.54 billion ZINC20 molecule presentations, then adapts to sparse bioactivity across 642 ChEMBL assays. Across 23 cluster-held-out OpenADMET endpoints, GRASP achieves a mean MAE of 0.374, compared with 0.383 for CheMeleon, 0.408 for ECFP4-LightGBM, and 0.440 for Chemprop. ChEMBL adaptation reduces mean MAE from 0.407 to 0.374 and improves 21 of 23 endpoints, showing the value of intermediate bioactivity supervision. On 22 TDC ADMET tasks, GRASP gives a better point estimate than CheMeleon on 13 tasks. These results support staged molecular representation learning, where abundant molecular structure and sparse experimental measurements provide complementary supervision at different stages of training.

The emergence of drug resistance in Mycobacterium tuberculosis (M.tb) presents a major challenge to drug discovery, as compounds active against drug-susceptible (DS) isolates may not retain comparable inhibitory activity against drug-resistant (DR) isolates. Here, we developed an AI/ML workflow to independently prioritize small molecules separately against DS and DR isolates of M.tb and predict MIC, MIC90, and MIC99 as distinct quantitative endpoints. Following curation of data retrieved from the ChEMBL repository, molecular-activity datasets comprised 25,645 and 3,749 compounds tested against whole cell DS and DR isolates respectively. Molecular representations were generated using RDKit descriptors and Morgan fingerprints, along with embeddings from ChemBERTa, MolCLR, and Uni-Mol, and evaluated for binary activity classification and endpoint-specific regression. The best-performing DS classifier combined RDKit/Morgan features with ExtraTrees (balanced accuracy 0.782; AUROC 0.864; AUPRC 0.870), while histogram gradient boosting using the same features performed best for DR classification (balanced accuracy 0.762; AUROC 0.836; AUPRC 0.880). Similarly, RDKit/Morgan features combined with tree-based regressors achieved the best overall performance across all six DS/DR regression endpoints, with pooled out-of-fold RMSE values ranging from 0.604 to 0.784 pMIC units. To facilitate broader application, we developed MtbAIM (https://github.com/jasdeep002/MtbAIM), a Google Colab-based prediction platform that enables users to assess and prioritize small molecules against DS and DR isolates of M.tb.

Automated triage of whole-slide cervical cytology images remains constrained by the heavy dependence on single-cell bounding boxes or pixel-level annotations, which introduce significant bottlenecks, high expert overhead, and observer variability. To address these challenges, we present a weakly-supervised, detection-free Multiple Instance Learning (MIL) framework powered by a Gated Attention pooling engine that operates exclusively on slide-level predictions. By treating whole-slide images as bags of local instance patches, the dual-branch gated network dynamically assigns non-linear attention weights to highlight isolated dysplastic cells while suppressing benign background and debris, successfully resolving the "needle-in-a-haystack" problem inherent to high-grade lesions. Evaluated across multi-center cohorts including internal datasets (SIPaKMeD, Herlev, and CRIC) and the unannotated, out-of-distribution Mendeley LBC validation cohort processed via an unsupervised marker-controlled watershed pipeline our approach demonstrates robust generalization without requiring dense instance labels. Furthermore, our architectural comparison reveals a key trade-off: lightweight backbones like MobileNetV2 optimize in-distribution multi-center accuracy (90.96% Accuracy, 0.9800 ROC-AUC), whereas higher-capacity models like Xception ensure superior out-of-distribution robustness (80.43% Balanced Accuracy) under domain shifts. Ultimately, this detection-less pipeline provides a scalable and clinically viable solution to automate slide triage and alleviate cytopathologist shortages in resource-constrained environments.

Accurate molecular property prediction requires effective molecular representations that can describe a molecule from multiple complementary perspectives. Existing deep learning approaches typically use SMILES strings, two-dimensional molecular graphs, or three-dimensional conformations as inputs. These representations capture different aspects of molecular information: SMILES encodes a sequential description, molecular graphs describe atom-bond connectivity, and 3D conformations provide spatial information from atomic coordinates. However, these modalities are often learned separately or merged with low effective fusion operations, which may not sufficiently capture interactions among sequential, topological, and geometric features. In this work, we propose Tri-Modality Cross-Attention (TMCA), a multimodal framework that integrates 1D SMILES, 2D molecular graphs, and 3D molecular conformations for molecular property prediction. TMCA uses pretrained encoders for the 1D and 3D branches, with a SMILES-based Transformer for the 1D branch and a recent conformation-aware pretrained model for the 3D branch. At the same time, a trainable 2D graph encoder is designed to support modality fusion. All the parameters from the pretrained models are frozen to make overall training more efficient. We evaluate TMCA on four datasets (i.e., BBBP, BACE, ClinTox, and HIV) from the MoleculeNet database for classification tasks and compare its performance with state-of-the-art competing approaches. Experimental results show that TMCA achieves the best overall performance, demonstrating the value of multi-modality integration for drug property prediction and the effectiveness of our cross-attention-based fusion strategy. The source code of TMCA is available at: https://github.com/Ay-Zhao/TMCA