What depth should I use for the assignment? · How would you like to receive it? · Should I include references?

AI in biochemistry protein structure genomics proteomics metabolomics biomarker discovery personalized nutrition review

artificial intelligence biochemistry

artificial intelligence protein structure prediction

machine learning metabolomics proteomics

artificial intelligence diabetes cancer metabolism

~/

mkdir -p "/tmp/workspace/ai_biochemistry_assignment/src" "/tmp/workspace/ai_biochemistry_assignment/output" && python3 - <<'PY' from pathlib import Path p=Path('/tmp/workspace/ai_biochemistry_assignment/src/create_assignment.py') content=r'''from docx import Document from docx.shared import Inches, Pt, RGBColor from docx.enum.text import WD_ALIGN_PARAGRAPH from docx.enum.section import WD_SECTION from docx.oxml import OxmlElement from docx.oxml.ns import qn from docx.enum.style import WD_STYLE_TYPE OUT='/tmp/workspace/ai_biochemistry_assignment/output/AI_in_Biochemistry_Assignment.docx' doc=Document() sec=doc.sections[0] sec.top_margin=Inches(.8); sec.bottom_margin=Inches(.75); sec.left_margin=Inches(.85); sec.right_margin=Inches(.85) styles=doc.styles normal=styles['Normal']; normal.font.name='Times New Roman'; normal._element.rPr.rFonts.set(qn('w:eastAsia'),'Times New Roman'); normal.font.size=Pt(12); normal.paragraph_format.line_spacing=1.5; normal.paragraph_format.space_after=Pt(6) for s in ['Title','Heading 1','Heading 2','Heading 3']: st=styles[s]; st.font.name='Times New Roman'; st._element.rPr.rFonts.set(qn('w:eastAsia'),'Times New Roman') styles['Heading 1'].font.size=Pt(16); styles['Heading 1'].font.bold=True; styles['Heading 1'].font.color.rgb=RGBColor(128,0,0); styles['Heading 1'].paragraph_format.space_before=Pt(14); styles['Heading 1'].paragraph_format.space_after=Pt(6) styles['Heading 2'].font.size=Pt(13); styles['Heading 2'].font.bold=True; styles['Heading 2'].font.color.rgb=RGBColor(80,0,0) styles['Heading 3'].font.size=Pt(12); styles['Heading 3'].font.bold=True def page_num(section): footer=section.footer.paragraphs[0]; footer.alignment=WD_ALIGN_PARAGRAPH.CENTER run=footer.add_run('Page '); fld=OxmlElement('w:fldSimple'); fld.set(qn('w:instr'),'PAGE'); footer._p.append(fld) page_num(sec) def p(text='', boldlead=None): par=doc.add_paragraph() if boldlead and text.startswith(boldlead): par.add_run(boldlead).bold=True; par.add_run(text[len(boldlead):]) else: par.add_run(text) return par def bullet(text): doc.add_paragraph(text, style='List Bullet') def heading(t,l=1): doc.add_heading(t,level=l) def citation(text): p(text) # Cover for _ in range(5): doc.add_paragraph() t=doc.add_paragraph(); t.alignment=WD_ALIGN_PARAGRAPH.CENTER; r=t.add_run('ARTIFICIAL INTELLIGENCE IN BIOCHEMISTRY'); r.bold=True; r.font.name='Times New Roman'; r.font.size=Pt(24); r.font.color.rgb=RGBColor(128,0,0) doc.add_paragraph().add_run('\n') x=doc.add_paragraph(); x.alignment=WD_ALIGN_PARAGRAPH.CENTER; x.add_run('Assignment').bold=True; x.runs[0].font.size=Pt(16) for _ in range(8): doc.add_paragraph() for label in ['Submitted by: ______________________________','Roll No.: ____________________________________','Course/Class: _________________________________','Submitted to: _________________________________','Institution: __________________________________','Date: ________________________________________']: q=doc.add_paragraph(label); q.alignment=WD_ALIGN_PARAGRAPH.CENTER doc.add_page_break() heading('Table of Contents') contents=['Abstract','1. Introduction','2. AI in Protein Structure Prediction','3. AI in Enzyme Research','4. AI in Metabolic Pathway Analysis','5. AI in Diabetes and Metabolic Syndrome','6. AI in Cancer Metabolism','7. AI in Genomics','8. AI in Proteomics','9. AI in Metabolomics','10. AI in Biomarker Discovery','11. AI in Personalized Nutrition','12. Cross-cutting Challenges and Ethical Considerations','13. Future Directions','14. Conclusion','References'] for item in contents: p(item) p('Note: Page numbers update automatically in Word when a formal table of contents is inserted or fields are refreshed.') doc.add_page_break() heading('Abstract') p('Artificial intelligence (AI) has become an important set of computational methods for biochemistry because modern biological experiments produce high-dimensional, noisy and interconnected data. Machine learning (ML), deep learning, graph neural networks and generative models can identify patterns in molecular sequences, structures, spectra and clinical measurements that are difficult to detect manually. This assignment reviews AI applications across protein structure prediction, enzyme research, metabolic pathway analysis, diabetes and metabolic syndrome, cancer metabolism, genomics, proteomics, metabolomics, biomarker discovery and personalized nutrition. Landmark protein-structure systems such as AlphaFold demonstrate how deep learning can convert sequence information into useful structural hypotheses. In parallel, AI improves interpretation of mass spectrometry, variant prioritization, multi-omics integration, glucose forecasting and diet-response prediction. However, an accurate model is not automatically clinically useful. Dataset quality, population representation, confounding, external validation, interpretability, privacy and regulatory oversight remain central concerns. AI should therefore be viewed as an assistive scientific method: it generates testable hypotheses, prioritizes experiments and supports decisions, while biochemical experiments and clinical judgement remain necessary for confirmation.') p('Keywords: artificial intelligence; machine learning; biochemistry; multi-omics; protein structure; metabolomics; precision medicine.') heading('1. Introduction') p('Biochemistry explains life through molecules: nucleic acids store information, proteins execute cellular functions, enzymes control reaction rates, and metabolites reflect the chemical state of cells and organisms. Contemporary technologies such as next-generation sequencing, mass spectrometry, high-content imaging and continuous biosensors generate data at a scale that cannot be analysed reliably by simple rules alone. AI refers broadly to computer systems that perform tasks associated with learning, pattern recognition, prediction or decision support. ML is a subset of AI in which algorithms learn associations from examples. Deep learning uses multilayer neural networks, while generative AI learns a distribution from which it can propose new sequences, structures or molecular designs.') p('The relevance of AI to biochemistry comes from the structure of the data. A protein sequence is an ordered string; a metabolic network is a graph; a mass spectrum is a high-dimensional signal; a clinical record is longitudinal and incomplete. Different AI architectures are suited to these forms. Convolutional networks detect local patterns, transformers model long-range dependencies, graph neural networks represent molecular interactions, and probabilistic models quantify uncertainty. The goal is not to replace the biochemical method. The goal is to turn complex observations into hypotheses that can be experimentally checked. Textbook discussions of AI in genomics similarly emphasize its ability to process very large datasets and identify relationships, while warning that biased health datasets and weak safeguards can reproduce inequities and threaten privacy [1].') heading('2. AI in Protein Structure Prediction') heading('2.1 Biochemical importance',2) p('Protein function depends strongly on three-dimensional structure. A sequence alone does not directly reveal an active site, binding surface, conformational switch or disease-causing structural defect. Classical methods such as X-ray crystallography, nuclear magnetic resonance spectroscopy and cryo-electron microscopy remain the reference methods, but can be slow, expensive or technically difficult for flexible proteins and complexes. Computational structure prediction therefore has long been a major aim of structural biochemistry.') heading('2.2 AI approaches and AlphaFold',2) p('Modern deep-learning systems use evolutionary information from multiple sequence alignments, structural templates and physical constraints to infer residue-residue geometry. AlphaFold2 was a watershed because it produced highly accurate predictions for many single-chain proteins in the CASP14 assessment [2]. RoseTTAFold provided an independent deep-learning approach for rapid prediction of protein structures and complexes [3]. The AlphaFold Protein Structure Database subsequently made predicted structures available at proteome scale, increasing access for students and researchers [4]. Models report confidence measures, including predicted local distance difference test scores and predicted aligned error. These values should guide interpretation: a high-confidence core can be useful, whereas a low-confidence region may be intrinsically disordered or simply uncertain.') p('AI models are now used to predict protein complexes, protein-ligand interactions, effects of mutations and alternative conformations. In biochemical research, a predicted fold can suggest catalytic residues, locate a putative cofactor-binding pocket, rationalize a missense variant or support construct design for experimental work. In drug discovery, structures can help prioritize docking, virtual screening and medicinal-chemistry experiments. Recent reviews describe the rapid expansion of AI from static structure prediction towards interaction prediction and protein design [5].') heading('2.3 Limits',2) p('Predicted structures are hypotheses, not direct measurements. Accuracy can be lower for protein complexes, membrane proteins, flexible loops, post-translationally modified proteins, ligand-bound states and multiple conformations. A structure that looks plausible may still give an incorrect functional interpretation. Researchers should inspect confidence scores, compare homologues, test key residues by mutagenesis, and validate important claims with structural or biochemical experiments. In this way AI reduces the search space but does not eliminate the need for evidence.') heading('3. AI in Enzyme Research') heading('3.1 Functional annotation and catalytic prediction',2) p('Enzymes catalyse reactions with high specificity. Yet many newly sequenced proteins remain poorly annotated, and similar sequences can have different substrates. AI can combine sequence embeddings, structural features, taxonomic context and known reactions to predict enzyme commission classes, substrate preferences and catalytic residues. Protein language models learn statistical regularities from millions of sequences and can be fine-tuned for functional prediction. Structural models provide an additional view of active-site shape and cofactor-binding regions.') p('Tools such as DeepEC and related classifiers illustrate how deep learning can assign enzyme functions from sequence data [6]. Such predictions are particularly useful for annotating genomes, identifying orphan enzymes in pathways and selecting candidates for expression. A sensible workflow is: obtain an AI-ranked list, inspect conservation of candidate residues, compare with curated databases, express selected proteins, and confirm activity using substrate assays and kinetic measurements. The final biochemical question remains experimental: does the enzyme catalyse the proposed reaction under defined conditions?') heading('3.2 Enzyme engineering',2) p('Directed evolution traditionally creates variants and screens them experimentally. AI makes this process more efficient by learning sequence-function relationships from measured variants, predicting which mutations may improve activity, stability, selectivity or expression, and proposing diverse candidates for the next round. Generative models can propose sequences compatible with a desired fold or active-site environment. This is relevant to industrial biocatalysis, biodegradation, food processing and synthesis of pharmaceuticals. The benefits are greatest when models are trained on high-quality, assay-specific data and when predicted variants are tested with appropriate controls.') heading('3.3 Reaction and kinetics modelling',2) p('ML can also predict reaction outcomes, estimate enzyme-substrate compatibility and model kinetic behaviour from time-series data. However, kinetic constants depend on pH, temperature, ionic strength, cofactors and assay design. A model trained across heterogeneous assays may learn laboratory artifacts rather than biological rules. Reporting the assay context and uncertainty is therefore essential. AI should support mechanistic enzymology, not substitute for it.') heading('4. AI in Metabolic Pathway Analysis') p('Metabolism is organized as an interconnected reaction network rather than a list of isolated pathways. AI helps reconstruct pathways from genomes, identify missing reactions, infer regulatory links and integrate transcriptomic, proteomic and metabolomic data. Graph-based methods are naturally suited to this problem because metabolites, enzymes and reactions form nodes and edges. Constraint-based metabolic models can be combined with ML to improve predictions of fluxes, growth phenotypes and responses to perturbation.') p('Pathway-guided deep-learning architectures incorporate established pathway databases such as KEGG, Reactome and Gene Ontology into their networks. This can make models more interpretable than unconstrained neural networks because an output can be related to a biological process rather than an anonymous feature [7]. In microbial systems, AI-assisted pathway reconstruction supports identification of biosynthetic gene clusters and metabolic engineering targets. In human disease, it can link changes in genes, proteins and metabolites to altered energy metabolism or inflammation.') p('Important challenges include incomplete pathway annotation, compartmentalization, tissue specificity and causal direction. A metabolite change may be a cause, consequence or compensatory response. The strongest analyses combine AI with isotope tracing, enzyme activity measurements, flux analysis and perturbation experiments. These methods can distinguish association from actual pathway flux.') heading('5. AI in Diabetes and Metabolic Syndrome') p('Diabetes and metabolic syndrome are biochemical disorders involving abnormal glucose regulation, insulin action, lipid metabolism, inflammation and energy balance. Their management produces rich longitudinal data: continuous glucose-monitoring (CGM) traces, insulin doses, meals, activity, sleep, laboratory values and medication history. AI can use these data to forecast glucose, detect impending hypo- or hyperglycaemia, estimate insulin needs and identify clinically meaningful patterns [8].') p('In type 1 diabetes, automated insulin delivery systems combine CGM, an insulin pump and a control algorithm. The algorithm adjusts insulin delivery based on measured and predicted glucose, but users and clinicians remain responsible for safe use, carbohydrate intake, device maintenance and response to alarms. In type 2 diabetes, models can support screening, risk stratification, prediction of complications and individualized lifestyle interventions. AI-based retinal-image analysis is also an example of a metabolic-disease application, although it must be assessed in the intended healthcare setting [9].') p('Metabolic syndrome includes central adiposity, dyslipidaemia, hypertension and dysglycaemia. AI may integrate these correlated measures with genomic and lifestyle information to identify subgroups that respond differently to interventions. Yet prediction can be misleading if it is trained in one population and deployed in another. Models must be externally validated, calibrated, monitored for bias and used alongside clinical guidelines. A glucose forecast is useful only if it improves safe decisions and meaningful patient outcomes.') heading('6. AI in Cancer Metabolism') p('Cancer cells frequently reprogram metabolism to support growth, redox control and survival. Increased glycolysis, altered glutamine use, lipid synthesis, nucleotide production and interactions with the tumour microenvironment are examples. These changes are heterogeneous across cancer types and even within a single tumour. AI can analyse imaging, genomics, transcriptomics, proteomics and metabolomics together to classify metabolic states, nominate therapeutic targets and estimate response to treatment.') p('For example, ML can identify metabolite signatures associated with tumour subtype or prognosis, while pathway-informed models can connect gene-expression changes with metabolic reactions. AI also helps process mass-spectrometry imaging data, which preserves spatial information and can reveal metabolic variation at the tumour margin. Multi-omics approaches may expose vulnerabilities that would not appear in any one data layer [10].') p('However, a model that separates tumour from normal tissue is not proof that a particular metabolic pathway is driving cancer. Validation should include metabolite quantification, isotope-labelled tracing, gene knockdown or pharmacological inhibition, and testing in relevant cellular or animal models. There are also safety concerns: a biomarker may fail across different ethnic groups, stages of disease or laboratory platforms. AI is most persuasive when it leads to a reproducible biological mechanism and a clinically testable intervention.') heading('7. AI in Genomics') p('Genomics is well suited to AI because sequencing datasets contain millions of variants and complex relationships between sequence and phenotype. AI supports base calling, variant calling, annotation, prediction of splicing effects, prioritization of disease variants, gene regulation modelling and interpretation of single-cell data. Deep learning can learn sequence motifs and long-range regulatory relationships that are difficult to express as hand-built rules.') p('DeepSEA demonstrated that deep learning can predict the chromatin effects of noncoding variants from sequence [11]. More recent sequence models such as Enformer use long-range genomic context to predict gene-expression-related signals [12]. In clinical genomics, AI can prioritize candidate variants in rare disease by combining inheritance, phenotype, population frequency, conservation and functional evidence. It can also cluster single cells into cell types and identify dysregulated regulatory programs in disease.') p('Genomic data are personally identifying and may have implications for biological relatives. Consent, secure storage, controlled sharing and clear governance are therefore required. Genomic models can also inherit ancestry bias because reference datasets remain unevenly distributed. AI should never convert a probabilistic variant prediction into a diagnosis without clinical, phenotypic and laboratory correlation. The role of clinical geneticists and genetic counsellors remains essential.') heading('8. AI in Proteomics') p('Proteomics measures the proteins present in a biological sample, their abundance, modifications, interactions and sometimes spatial distribution. Mass spectrometry produces large collections of spectra that must be matched to peptides, quantified across samples and interpreted biologically. ML improves several stages: retention-time prediction, fragmentation-spectrum prediction, peptide identification, false-discovery control and extraction of quantitative signals.') p('Prosit, for example, uses deep learning to predict peptide tandem mass spectra and retention times, improving targeted and data-independent proteomic analyses [13]. Deep learning is also used to predict protein subcellular localization, post-translational modifications and interaction networks. A recent review argues that ML is increasingly being integrated across the whole proteomics experiment rather than used only at a single computational step [14].') p('The biochemical significance of proteomics lies in its proximity to phenotype: proteins execute most cellular functions. Nevertheless, abundance does not always equal activity. Enzyme activity may change through phosphorylation, localization, binding partners or metabolite availability. Good proteomic AI analyses therefore require rigorous sample preparation, batch correction, appropriate multiple-testing control and orthogonal validation such as immunoassay, targeted mass spectrometry or functional assays.') heading('9. AI in Metabolomics') p('Metabolomics measures small molecules that represent the integrated output of genes, proteins, diet, microbiota, medicines and environment. This makes it attractive for disease phenotyping, but also challenging because metabolite signals are chemically diverse, affected by pre-analytical handling and difficult to identify conclusively. AI can assist peak detection, spectral deconvolution, compound annotation, feature selection, classification and integration with other omics layers.') p('Supervised algorithms such as random forests, support-vector machines and neural networks can classify samples when labelled training data are available. Unsupervised methods can find clusters or latent factors. Reviews of ML in metabolomics emphasize its role in diagnostic interpretation while also stressing risks of overfitting in small, high-dimensional datasets [15,16]. Proper workflows separate training and test sets before feature selection, use nested cross-validation, report confidence intervals, and validate on an independent cohort.') p('Metabolomic associations are especially prone to confounding by diet, fasting status, medication, sex, age, renal function and sampling time. AI can detect impressive patterns that disappear when these factors change. Standard operating procedures and transparent reporting are therefore as important as the choice of algorithm. Metabolite identities should be confirmed where possible by authentic standards and accurate mass, retention time and fragmentation evidence.') heading('10. AI in Biomarker Discovery') p('A biomarker is a measurable characteristic that indicates a normal biological process, disease process, exposure or response to an intervention. Biomarkers may be diagnostic, prognostic, predictive, monitoring or safety biomarkers. AI is valuable because candidate biomarker discovery often begins with thousands of genes, proteins or metabolites. ML can rank features, construct multivariable signatures and integrate heterogeneous inputs such as clinical variables, omics results and imaging data.') p('A credible biomarker pipeline has several stages: define the intended use; assemble representative discovery data; prevent data leakage; select features within resampling; lock the model; test it in external cohorts; assess calibration and clinical utility; and verify analytical performance. Area under the receiver-operating characteristic curve alone is insufficient. A test should also show sensitivity, specificity, predictive values, calibration, feasibility and benefit compared with existing practice. The FDA-NIH BEST resource provides a useful framework for biomarker terminology and qualification [17].') p('Multi-omics AI can improve early detection by combining complementary signals, but it increases complexity and cost. Feature importance methods may help interpretation, but they do not prove causation. Before clinical adoption, assays need reproducibility across sites and instruments, and a prospective study should show that use of the biomarker improves a patient-relevant outcome. These requirements prevent attractive discovery results from becoming unusable clinical tests.') heading('11. AI in Personalized Nutrition') p('Personalized nutrition seeks to tailor dietary advice using individual characteristics rather than relying only on population averages. Relevant inputs may include dietary records, anthropometry, blood glucose, lipids, activity, sleep, medical history, microbiome composition, genomics and metabolomics. AI can combine these inputs to predict postprandial glucose responses, nutrient adequacy, adherence barriers and likely response to dietary changes.') p('A notable study by Zeevi and colleagues showed that postprandial glycaemic responses vary substantially between individuals and can be predicted from clinical and microbiome features [18]. The PREDICT 1 study further demonstrated marked individual variability in metabolic responses to standardized meals [19]. These findings support the biochemical idea that diet responses depend on more than food composition alone. Continuous glucose monitoring and mobile applications can provide frequent feedback, although their recommendations must be understandable and safe.') p('Personalized nutrition should not be confused with deterministic “gene-based diets.” Most diet-related traits are polygenic and strongly influenced by environment, culture, access to food and behaviour. Recommendations must remain consistent with established nutritional principles and should consider affordability, allergies, kidney disease, pregnancy, eating disorders and medication interactions. Claims should be supported by controlled trials with clinically relevant outcomes, not merely improved prediction accuracy. Recent reviews note that multi-omics integration has potential for precision nutrition but also faces methodological and implementation challenges [20].') heading('12. Cross-cutting Challenges and Ethical Considerations') heading('12.1 Data quality, bias and generalizability',2) p('AI is only as reliable as its data and evaluation. Common threats include missing values, batch effects, mislabeled samples, narrow recruitment, confounding and data leakage. An algorithm can show excellent internal accuracy while failing in a new laboratory or population. Datasets should represent the people and biological states in which the model will be used. Performance should be reported separately across relevant groups, and calibration should be checked rather than assumed.') heading('12.2 Interpretability and reproducibility',2) p('Biochemical applications often require explanation. A model that predicts a disease state should identify whether the signal reflects a plausible metabolite pattern, a technical artifact or an unrelated confounder. Interpretable or pathway-guided models can aid hypothesis generation, but explanation methods themselves have limits. Reproducibility requires versioned data, documented preprocessing, preregistered analysis where possible, code sharing, independent validation and reporting of negative results.') heading('12.3 Privacy, consent and governance',2) p('Genomic and multi-omics data are sensitive because they can identify individuals and reveal information about relatives. Security measures should include data minimization, access control, encryption, audit trails and explicit consent procedures. Researchers must define whether data may be reused, shared or used to train commercial models. In clinical settings, regulatory assessment should address safety, performance drift, human oversight and accountability.') heading('12.4 Environmental and educational considerations',2) p('Large models consume computational resources and may be inaccessible to laboratories with limited infrastructure. A smaller, transparent model that performs adequately may be preferable to a large opaque system. Education is also necessary: biochemists need enough AI literacy to assess validation, while data scientists need enough biochemical understanding to avoid implausible conclusions.') heading('13. Future Directions') p('The next phase of AI in biochemistry will likely focus on integrated, experimental and causal systems. Foundation models may connect sequences, structures, images, spectra and scientific text. Generative models may propose enzymes, proteins or metabolic routes for experimental testing. Spatial multi-omics will help relate molecular programs to tissue architecture. Digital twins may combine physiology and longitudinal data to simulate response to therapy, although such systems require strong validation before use in patient care.') p('The most useful systems will combine human expertise with transparent uncertainty. Rather than claiming a single correct answer, they should indicate confidence, identify the evidence supporting a prediction and recommend the next discriminating experiment. This approach aligns AI with the iterative logic of biochemistry: observation, hypothesis, experiment, revision and replication.') heading('14. Conclusion') p('AI is reshaping biochemistry by making it possible to analyse molecular data at scales that exceed manual methods. Its most visible achievements include accurate protein-structure prediction, improved interpretation of genomics and mass spectrometry, and multi-omics models for disease and nutrition. Across all applications, the same principle applies: AI generates or prioritizes hypotheses, while experimental validation establishes biochemical truth. Rigorous design, representative datasets, external validation, privacy protection, interpretability and clinical oversight are necessary for responsible translation. Used in this disciplined manner, AI can accelerate discovery while preserving the standards of evidence that make biochemistry reliable.') heading('References') refs=[ '1. Nussbaum RL, McInnes RR, Willard HF. Thompson & Thompson genetics and genomics in medicine. 9th ed. Philadelphia: Elsevier; 2023. p. 442.', '2. Jumper J, Evans R, Pritzel A, Green T, Figurnov M, Ronneberger O, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596(7873):583-9.', '3. Baek M, DiMaio F, Anishchenko I, Dauparas J, Ovchinnikov S, Lee GR, et al. Accurate prediction of protein structures and interactions using a three-track neural network. Science. 2021;373(6557):871-6.', '4. Varadi M, Anyango S, Deshpande M, Nair S, Natassia C, Yordanova G, et al. AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models. Nucleic Acids Res. 2022;50(D1):D439-44.', '5. Heinzinger M, Rost B. Artificial intelligence learns protein prediction. Cold Spring Harb Perspect Biol. 2024;16(10):a041458.', '6. Ryu JY, Kim HU, Lee SY. Deep learning enables high-quality and high-throughput prediction of enzyme commission numbers. Proc Natl Acad Sci U S A. 2019;116(28):13996-4001.', '7. Yu L, Kong J, Zhang X, et al. Pathway-guided architectures for interpretable AI in biological research. Brief Bioinform. 2025;26(6):bbaf559.', '8. Gautier T, Ziegler LB, Gerber MS. Artificial intelligence and diabetes technology: a review. Metabolism. 2021;124:154872.', '9. Abràmoff MD, Lavin PT, Birch M, Shah N, Folk JC. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. NPJ Digit Med. 2018;1:39.', '10. Armitage EG, Southam AD. Monitoring cancer prognosis, diagnosis and treatment efficacy using metabolomics and artificial intelligence. Metabolomics. 2021;17(6):52.', '11. Zhou J, Troyanskaya OG. Predicting effects of noncoding variants with deep learning-based sequence model. Nat Methods. 2015;12(10):931-4.', '12. Avsec Ž, Agarwal V, Visentin D, Ledsam JR, Grabska-Barwinska A, Taylor KR, et al. Effective gene expression prediction from sequence by integrating long-range interactions. Nat Methods. 2021;18(10):1196-203.', '13. Gessulat S, Schmidt T, Zolg DP, Samaras P, Schnatbaum K, Zerweck J, et al. Prosit: proteome-wide prediction of peptide tandem mass spectra by deep learning. Nat Methods. 2019;16(6):509-18.', '14. Neely BA, Dorfer V, Martens L. Toward an integrated machine learning model of a proteomics experiment. J Proteome Res. 2023;22(6):1715-23.', '15. Galal AA, Talal MM, Moustafa A. Applications of machine learning in metabolomics: disease modeling and classification. Front Genet. 2022;13:1017340.', '16. Xu Y, Cao L, Chen Y, et al. Integrating machine learning in metabolomics: a path to enhanced diagnostics and data interpretation. Small Methods. 2024;8(9):e2400305.', '17. FDA-NIH Biomarker Working Group. BEST (Biomarkers, EndpointS, and other Tools) resource [Internet]. Silver Spring (MD): Food and Drug Administration; 2021 [cited 2026 Oct 4]. Available from: https://www.ncbi.nlm.nih.gov/books/NBK338448/', '18. Zeevi D, Korem T, Zmora N, Israeli D, Rothschild D, Weinberger A, et al. Personalized nutrition by prediction of glycemic responses. Cell. 2015;163(5):1079-94.', '19. Berry SE, Valdes AM, Drew DA, Asnicar F, Mazidi M, Wolf J, et al. Human postprandial responses to food and potential for precision nutrition. Nat Med. 2020;26(6):964-73.', '20. Nourazarain A, Vaziri Y. Nutrigenomics meets multi-omics: integrating genetic, metabolic, and microbiome data for personalized nutrition strategies. Genes Nutr. 2025;20:30.' ] for r in refs: p(r) # add header on all except cover for section in doc.sections: header=section.header.paragraphs[0]; header.alignment=WD_ALIGN_PARAGRAPH.CENTER; rr=header.add_run('AI IN BIOCHEMISTRY'); rr.font.name='Times New Roman'; rr.font.size=Pt(9); rr.font.bold=True # first page header different sec.different_first_page_header_footer=True # metadata doc.core_properties.title='Artificial Intelligence in Biochemistry' doc.core_properties.subject='Assignment' doc.core_properties.author='' doc.save(OUT) print(OUT) ''' p.write_text(content) PY python3 "/tmp/workspace/ai_biochemistry_assignment/src/create_assignment.py"

~/ai_biochemistry_assignment/output

python3 - <<'PY' from docx import Document from pathlib import Path path=Path('/tmp/workspace/ai_biochemistry_assignment/output/AI_in_Biochemistry_Assignment.docx') doc=Document(path) print('exists',path.exists(),'bytes',path.stat().st_size) print('paragraphs',len(doc.paragraphs),'tables',len(doc.tables)) print('headings',sum(1 for p in doc.paragraphs if p.style and p.style.name.startswith('Heading'))) print('references',sum(1 for p in doc.paragraphs if p.text[:2].isdigit() and '. ' in p.text)) print('last:',doc.paragraphs[-1].text) PY