Fingolimod Hydrochloride

CAMBRIDGE, England & BOSTON--( BUSINESS WIRE )--Cycle Pharmaceuticals, Inc. (Cycle) announces today completion of the acquisition of Banner Life Sciences, LLC (Banner) which includes its BAFIERTAM ® (monomethyl fumarate) product for the treatment of relapsing forms of multiple sclerosis (MS), in adults, approved by the US Food and Drug Administration (FDA). MS affects an estimated 1 million people in the US with around 200 new cases being diagnosed each week. 1 Approximately 85% of diagnoses are the relapsing-remitting form of the condition and, without treatment, about 50% of those progress to the more serious secondary-progressive form within 10 years. 1 BAFIERTAM will become Cycle’s second branded product for the treatment of MS alongside TASCENSO ODT ® (fingolimod). As with TASCENSO, BAFIERTAM patients will benefit from Cycle’s industry leading hub program, Cycle Vita™*. BAFIERTAM patients will continue to be supported by Banner’s established network during a transition period. The purchase of Banner was funded from cash on hand and demonstrates the capability of Cycle to reinvest in products where it can provide the highest quality medicines and support to patients who need it most. This acquisition comes at a time of increasing attention to the long-standing concern about the substandard quality of generic medicines in the neurology community 2 , as most recently documented in MS by Professor Darin Okuda (UT Southwestern Medical Center, Dallas, TX). 3 “ We are delighted to strengthen Cycle’s offer to the MS community by adding BAFIERTAM to our MS product portfolio. We will be able to support more patients via our best-in-class patient support hub Cycle Vita*, as well as giving patients full confidence in the medicines they are using to manage their MS, ” says James Harrison – CEO of Cycle. Ondra LLP served as exclusive financial advisor to Cycle, Goodwin Procter LLP acted as legal advisor. Leerink Partners served as exclusive financial advisor to Banner and Skadden, Arps, Slate, Meagher & Flom LLP acted as legal advisor. To find out more about BAFIERTAM, please visit www.bafiertam.com . To find out more about TASCENSO ODT, please visit www.tascenso.com . To find out more about Cycle Vita, please visit www.cyclevita.life or call 888-360-8482. About BAFIERTAM ® (monomethyl fumarate) BAFIERTAM is a prescription medicine used to treat relapsing forms of multiple sclerosis (MS), to include clinically isolated syndrome, relapsing-remitting disease, and active secondary progressive disease, in adults. It is not known if BAFIERTAM is safe and effective in children. SELECTED SAFETY INFORMATION CONTRAINDICATIONS BAFIERTAM is contraindicated in patients with known hypersensitivity to monomethyl fumarate, dimethyl fumarate, diroximel fumarate, or to any of the excipients of BAFIERTAM. BAFIERTAM should not be taken with dimethyl fumarate or diroximel fumarate. WARNINGS AND PRECAUTIONS Anaphylaxis and Angioedema - BAFIERTAM can cause anaphylaxis and angioedema after the first dose or at any time during treatment. Signs and symptoms in patients taking dimethyl fumarate (the prodrug of BAFIERTAM) have included difficulty breathing, urticaria, and swelling of the throat and tongue. Patients should be instructed to discontinue BAFIERTAM and seek immediate medical care should they experience signs and symptoms of anaphylaxis or angioedema. Progressive Multifocal Leukoencephalopathy - Progressive multifocal leukoencephalopathy (PML) has occurred in patients with MS treated with dimethyl fumarate (the prodrug of BAFIERTAM). PML is an opportunistic viral infection of the brain caused by the JC virus (JCV) that typically only occurs in patients who are immunocompromised, and that usually leads to death or severe disability. A fatal case of PML occurred in a patient who received dimethyl fumarate for 4 years while enrolled in a clinical trial. During the clinical trial, the patient experienced prolonged lymphopenia (lymphocyte counts predominantly <0.5x10 9 /L for 3.5 years) while taking dimethyl fumarate. The patient had no other identified systemic medical conditions resulting in compromised immune system function and had not previously been treated with natalizumab, which has a known association with PML. The patient was also not taking any immunosuppressive or immunomodulatory medications concomitantly. PML has also occurred in patients taking dimethyl fumarate in the postmarketing setting in the presence of lymphopenia (<0.9x10 9 /L). While the role of lymphopenia in these cases is uncertain, the PML cases have occurred predominantly in patients with lymphocyte counts <0.8x10 9 /L persisting for more than 6 months. At the first sign or symptom suggestive of PML, withhold BAFIERTAM and perform an appropriate diagnostic evaluation. Typical symptoms associated with PML are diverse, progress over days to weeks, and include progressive weakness on one side of the body or clumsiness of limbs, disturbance of vision, and changes in thinking, memory, and orientation leading to confusion and personality changes. Magnetic resonance imaging (MRI) findings may be apparent before clinical signs or symptoms. Cases of PML, diagnosed based on MRI findings and the detection of JCV DNA in the cerebrospinal fluid in the absence of clinical signs or symptoms specific to PML, have been reported in patients treated with other MS medications associated with PML. Many of these patients subsequently became symptomatic with PML. Therefore, monitoring with MRI for signs that may be consistent with PML may be useful, and any suspicious findings should lead to further investigation to allow for an early diagnosis of PML, if present. Lower PML-related mortality and morbidity have been reported following discontinuation of another MS medication associated with PML in patients with PML who were initially asymptomatic compared to patients with PML who had characteristic clinical signs and symptoms at diagnosis. It is not known whether these differences are due to early detection and discontinuation of MS treatment or due to differences in disease in these patients. Herpes Zoster and Other Serious Opportunistic Infections - Serious cases of herpes zoster have occurred with dimethyl fumarate (the prodrug of BAFIERTAM), including disseminated herpes zoster, herpes zoster ophthalmicus, herpes zoster meningoencephalitis, and herpes zoster meningomyelitis. These events may occur at any time during treatment. Monitor patients on BAFIERTAM for signs and symptoms of herpes zoster. If herpes zoster occurs, appropriate treatment for herpes zoster should be administered. Other serious opportunistic infections have occurred with dimethyl fumarate, including cases of serious viral (herpes simplex virus, West Nile virus, cytomegalovirus), fungal (Candida and Aspergillus), and bacterial (Nocardia, Listeria monocytogenes, Mycobacterium tuberculosis) infections. These infections have been reported in patients with reduced absolute lymphocyte counts (ALC) as well as in patients with normal ALC. These infections have affected the brain, meninges, spinal cord, gastrointestinal tract, lungs, skin, eye, and ear. Patients with symptoms and signs consistent with any of these infections should undergo prompt diagnostic evaluation and receive appropriate treatment. Consider withholding BAFIERTAM treatment in patients with herpes zoster or other serious infections until the infection has resolved. Lymphopenia - BAFIERTAM may decrease lymphocyte counts. In the MS placebo-controlled trials with dimethyl fumarate (the prodrug of BAFIERTAM), mean lymphocyte counts decreased by approximately 30% during the first year of treatment with dimethyl fumarate and then remained stable. Four weeks after stopping dimethyl fumarate, mean lymphocyte counts increased, but did not return to baseline. Six percent (6%) of dimethyl fumarate patients and <1% of placebo patients experienced lymphocyte counts <0.5x10 9 /L (lower limit of normal 0.91x10 9 /L). The incidence of infections (60% vs 58%) and serious infections (2% vs 2%) was similar in patients treated with dimethyl fumarate or placebo, respectively. There was no increased incidence of serious infections observed in patients with lymphocyte counts <0.8x10 9 /L or <0.5x10 9 /L in controlled trials, although one patient in an extension study developed PML in the setting of prolonged lymphopenia (lymphocyte counts predominantly <0.5x10 9 /L for 3.5 years). In controlled and uncontrolled clinical trials with dimethyl fumarate, 2% of patients experienced lymphocyte counts <0.5 x 10 9 /L for at least six months, and in this group, the majority of lymphocyte counts remained <0.5x10 9 /L with continued therapy. In these patients with prolonged, severe lymphopenia, the median time for lymphocyte counts to return to normal after discontinuing dimethyl fumarate was 96.0 weeks. In these controlled and uncontrolled clinical studies, among patients who did not experience prolonged, severe lymphopenia during treatment, the median times for lymphocyte counts to return to normal after discontinuing dimethyl fumarate were as follows: 4.3 weeks in patients with mild lymphopenia (lymphocyte count ≥0.8x10 9 /L) at discontinuation, 10.0 weeks in patients with moderate lymphopenia (lymphocyte count 0.5 to <0.8x10 9 /L) at discontinuation, and 16.7 weeks in patients with severe lymphopenia (lymphocyte count <0.5x10 9 /L) at discontinuation. Neither BAFIERTAM nor dimethyl fumarate have been studied in patients with preexisting low lymphocyte counts. Obtain a CBC, including lymphocyte count, before initiating treatment with BAFIERTAM, 6 months after starting treatment, and then every 6 to 12 months thereafter, and as clinically indicated. Consider interruption of BAFIERTAM in patients with lymphocyte counts less than 0.5 x 10 9 /L persisting for more than six months. Given the potential for delayed recovery of lymphocyte counts, continue to obtain lymphocyte counts until their recovery if BAFIERTAM is discontinued or interrupted because of lymphopenia. Consider withholding treatment from patients with serious infections until resolution. Decisions about whether or not to restart BAFIERTAM should be individualized based on clinical circumstances. Liver Injury - Clinically significant cases of liver injury have been reported in patients treated with dimethyl fumarate (the prodrug of BAFIERTAM) in the postmarketing setting. The onset has ranged from a few days to several months after initiation of treatment with dimethyl fumarate. Signs and symptoms of liver injury, including elevation of serum aminotransferases to greater than 5-fold the upper limit of normal and elevation of total bilirubin to greater than 2-fold the upper limit of normal have been observed. These abnormalities resolved upon treatment discontinuation. Some cases required hospitalization. None of the reported cases resulted in liver failure, liver transplant, or death. However, the combination of new serum aminotransferase elevations with increased levels of bilirubin caused by drug-induced hepatocellular injury is an important predictor of serious liver injury that may lead to acute liver failure, liver transplant, or death in some patients. Elevations of hepatic transaminases (most no greater than 3 times the upper limit of normal) were observed during controlled trials with dimethyl fumarate. Obtain serum aminotransferase, alkaline phosphatase (ALP), and total bilirubin levels prior to treatment with BAFIERTAM and during treatment, as clinically indicated. Discontinue BAFIERTAM if clinically significant liver injury induced by BAFIERTAM is suspected. Flushing - BAFIERTAM may cause flushing (e.g., warmth, redness, itching, and/or burning sensation). In clinical trials of dimethyl fumarate (the prodrug of BAFIERTAM), 40% of dimethyl fumarate-treated patients experienced flushing. Flushing symptoms generally began soon after initiating dimethyl fumarate and usually improved or resolved over time. In the majority of patients who experienced flushing, it was mild or moderate in severity. Three percent (3%) of patients discontinued dimethyl fumarate for flushing, and <1% had serious flushing symptoms that were not life-threatening but led to hospitalization. Studies with dimethyl fumarate show that administration of non-enteric coated aspirin (up to a dose of 325 mg) 30 minutes prior to dosing may reduce the incidence or severity of flushing. In the BAFIERTAM studies, the presence of food did not impact the incidence of flushing. Serious Gastrointestinal Reactions - Serious gastrointestinal reactions, including perforation, ulceration, hemorrhage, and obstruction, some with fatal outcomes, have been reported in the postmarketing setting with the use of fumaric acid esters, including dimethyl fumarate, with or without concomitant aspirin use. The majority of these events have occurred within 6 months of fumaric acid ester treatment initiation. In controlled clinical trials, the incidence of serious gastrointestinal adverse events was 1% in patients treated with dimethyl fumarate; these events, none of which were fatal, included vomiting (0.3%) and abdominal pain (0.3%). Monitor patients, promptly evaluate, and discontinue BAFIERTAM for new or worsening severe gastrointestinal signs and symptoms. DRUG INTERACTIONS Concomitant Dimethyl Fumarate or Diroximel Fumarate - Both dimethyl fumarate and diroximel fumarate are metabolized to monomethyl fumarate. Therefore, BAFIERTAM is contraindicated in patients currently taking dimethyl fumarate or diroximel fumarate. BAFIERTAM may be initiated the day following discontinuation of either of these drugs. USE IN SPECIFIC POPULATIONS Pregnancy - There are no adequate data on the developmental risk associated with the use of BAFIERTAM in pregnant women. Available data from a pregnancy registry for dimethyl fumarate (the prodrug of BAFIERTAM), observational studies, and pharmacovigilance with dimethyl fumarate use in pregnant women have not indicated an increased risk of major birth defects, miscarriage, or other adverse maternal or fetal outcomes. Most of the reported exposures to dimethyl fumarate occurred during the first trimester of pregnancy (see Data). In animals, adverse effects on offspring survival, growth, sexual maturation, and neurobehavioral function were observed when dimethyl fumarate (DMF) was administered during pregnancy and lactation at clinically relevant doses (see Data). The background risk of major birth defects and miscarriage for the indicated population is unknown. All pregnancies have a background risk of birth defect, loss, or other adverse outcomes. In the U.S. general population, the estimated background risk of major birth defects and miscarriage in clinically recognized pregnancies is 2-4% and 15-20%, respectively. Lactation - There are no data on the presence of DMF or MMF in human milk. The effects on the breastfed infant and on milk production are unknown. Patients should call their healthcare provider if they are breastfeeding or plan to breastfeed because it is not known if BAFIERTAM passes into breast milk. Pediatric Use - Safety and effectiveness in pediatric patients have not been established. Geriatric Use - Clinical studies of dimethyl fumarate (the prodrug of BAFIERTAM) and of BAFIERTAM did not include sufficient numbers of patients aged 65 and over to determine whether they respond differently from younger patients. Please click here for Important Safety Information and full Prescribing Information , including Patient Information for BAFIERTAM. About TASCENSO ODT ® INDICATIONS TASCENSO ODT is a prescription medicine used to treat relapsing forms of multiple sclerosis (MS), to include clinically isolated syndrome, relapsing-remitting disease, and active secondary progressive disease, in adults and children 10 years of age and older. CONTRAINDICATIONS Do not take TASCENSO ODT if: In the last 6 months you experienced heart attack, unstable angina, stroke or mini-stroke (transient ischemic attack or TIA), or certain types of heart failure. You have an irregular or abnormal heartbeat (arrhythmia), including a heart finding called prolonged QT as seen on an ECG, or if you take medicines that change your heart rhythm. You are allergic to fingolimod or any of the other ingredients. IMPORTANT SAFETY INFORMATION Warnings and Precautions TASCENSO ODT may cause serious side effects. TASCENSO ODT can slow your heart rate, therefore you will be monitored after the first dose. There is an increased risk of serious infections, some of which can be life threatening; contact your doctor if you develop any symptoms of infections. You may experience shortness of breath and increased blood pressure. Progressive Multifocal Leukoencephalopathy (PML), a rare brain infection, has occurred in patients taking fingolimod; call your doctor immediately if you experience symptoms including weakness on one side of the body or clumsiness of limbs, disturbance of vision, and changes in thinking, memory, and orientation leading to confusion and personality changes. If PML is confirmed, treatment should be stopped. Your doctor will monitor you for the development of Immune Reconstitution Inflammatory Syndrome (PML-IRIS), which has been reported in patients who stopped treatment after developing PML. Macular edema, a vision problem that can cause some of the same vision symptoms as an MS attack may occur; call your doctor immediately if you have any vision changes. TASCENSO ODT may cause liver damage; call your doctor immediately if you experience unexplained nausea, abdominal pain, fatigue, anorexia, jaundice and/or dark urine. Rare cases of Posterior Reversible Encephalopathy Syndrome in adults have been reported with fingolimod; call your doctor immediately if you experience sudden onset of severe headaches, altered mental status, visual disturbances or seizures. TASCENSO ODT may cause harm to your unborn baby; speak to your doctor before becoming pregnant. Stopping TASCENSO ODT may cause worsening of your MS symptoms, consult a doctor before stopping treatment. Malignancies have been associated with fingolimod treatment; limit exposure to sunlight and call your doctor if you have any changes in skin appearance. You may experience a hypersensitivity reaction; call your doctor if you experience rash, urticaria and angioedema. TASCENSO ODT remains in the blood and can continue to have effects up to two months after treatment. Adverse Reactions The most common side effects with fingolimod are headache, abnormal liver tests, diarrhea, cough, flu, inflammation of the sinuses, back pain, stomach pain, and pain in arms and legs. If you have any new or worsening negative side effects whilst taking TASCENSO ODT you should immediately call you doctor. Tell your doctor about all your medical conditions and all medicines you take, before starting treatment. If you take too much TASCENSO ODT immediately call your doctor or go to the nearest hospital emergency room. For more detailed information, please refer to the full Prescribing Information at www.tascenso.com/PI . For more detailed information, please refer to the Medication Guide - www.tascenso.com/MG . To report SUSPECTED ADVERSE REACTIONS, contact Cycle Pharmaceuticals Ltd at 1-888-533-1625 or the FDA at: 1-800-FDA-1088 or www.fda.gov/medwatch . References Healthline. (2022). Multiple Sclerosis: Facts, Statistics, and You. [online] Available via: https://www.healthline.com/health/multiple-sclerosis/facts-statistics-infographic#risk-factors [Accessed Jan 13 2025]. American Academy of Neurology. (2001). Substitution of Generic Drugs May Cause Problems for Epilepsy Patients. [online] Available via: https://www.aan.com/PressRoom/home/PressRelease/115 [Accessed Jan 13 2025]. Okuda DT, et al. (2024). Clinical and radiological implications of subpotent generic fingolimod in multiple sclerosis: a case series. Therapeutic Advances in Neurological Disorders. 2024;17, 17562864241300047. Available at: https://doi/10.1177/17562864241300047 [Accessed: Jan 13 2025]. *Some areas of support may not be accessible to all patients. Eligibility criteria may apply to ensure compliance with all applicable federal and state requirements, and benefits may be limited to commercially insured patients only. For more detailed information about eligibility, terms and conditions, please contact the Cycle Vita team at 888-360-8482 or visit www.cyclevita.life . BAFIERTAM ® is a registered trademark of Banner Life Sciences, LLC. TASCENSO ODT ® is a registered trademark of Handa Neuroscience, LLC. Cycle Vita™ is a trademark of Cycle Pharmaceuticals Limited. ©2025 Cycle Pharmaceuticals Limited. All rights reserved. About Cycle Pharmaceuticals Cycle Pharmaceuticals was founded in 2012 with the sole aim of delivering drug treatments and product support to the underserved rare disease patient community. Cycle focuses on rare metabolic, immunological, and neurological genetic conditions. Within neurological conditions, we focus on multiple sclerosis. Cycle is headquartered in Cambridge, UK and has offices in Boston, Massachusetts. For more information, please visit www.cyclepharma.com and follow us on X , LinkedIn and Facebook .

今天为大家介绍的是一篇华东理工大学唐赟教授团队在《Briefings in Bioinformatics》上发表的文章。本文通过结合迁移学习和基于提示的文本预测策略，提出了MetaPredictor模型，能够提供更全面和准确的药物代谢预测。unsetunset一、文章摘要unsetunset 代谢过程可以将药物转化为具有不同性质的代谢物，这可能影响其功效和安全性。因此，研究药物候选物的代谢命运对药物发现具有重大意义。虽然已经开发了计算方法来预测药物代谢物，但大多数方法面临两个主要障碍：由于代谢转化规则或特定酶家族的限制导致模型泛化能力不足，以及预测假阳性率过高。 在此，作者提出了MetaPredictor，这是一种无规则、端到端和基于提示工程的方法，将小分子（包括药物）可能的人体代谢物预测作为序列翻译问题来处理。作者创新性地将提示工程引入深度语言模型，以丰富领域知识并指导决策。结果表明，使用指定代谢位点（SoMs）的提示可以引导模型提出更准确的代谢物预测，与基线模型相比，召回率提高了30.4%，假阳性减少了16.8%。 作者还利用迁移学习策略来解决代谢数据有限的问题。为了适应自动或非专家预测，MetaPredictor被设计为两阶段模式，包括自动识别SoMs，然后进行代谢物预测。与四种现有的药物代谢物预测工具相比，本文的方法在主要酶家族上表现相当，并且具有更好的泛化能力，可以额外识别由不太常见的酶催化的代谢物。 结果表明，通过有效结合迁移学习和基于提示的学习策略，MetaPredictor能够提供更全面和准确的药物代谢预测。unsetunset二、研究背景unsetunset药物代谢 药物代谢通过各种药物代谢酶催化的化学修饰改变药物分子。产生的代谢物可能具有与原始母体分子不同的物理化学、药理学甚至毒理学特性。药物代谢通常分为两个阶段，这些阶段通常在肝脏中进行。 I.相代谢通常涉及药物极性基团的引入或暴露，主要由细胞色素P450（CYP450）酶家族、其他氧化还原酶和水解酶催化。Ⅱ相代谢由转移酶介导，作用是将某些内源性小分子与药物结合，使其易于从体内排出。 尽管外源物质通常通过代谢被解毒和失活，但药物代谢可能会降低药效并引起安全问题。毒性大多由I.相引发，较少见于II.相代谢。除此之外，药物代谢可能导致药物-药物相互作用并影响生物利用度。因此，分析代谢过程对有效开发药物具有重要的指导意义。 传统上，药物代谢的研究需要使用复杂的分析技术，这些技术既耗费资源又耗费人力。现有计算方法的不足 为了辅助实验评估，已经开发了药物代谢预测的计算方法。这些方法中的一些在识别分子内被代谢转化修饰的原子（即代谢位点，SoMs）方面非常有效，如SMARTCyp 、FAME2 、SOMP 和Xenosite ，但它们几乎专门针对CYP同工酶。正确识别SoMs可以帮助推断代谢物结构或建议分子可能被合理设计的位置。 与体外SoM预测相比，从母体化合物推断代谢物结构的计算任务更加困难。目前这项任务的方法主要由基于规则的方法主导，如SyGMa 、Biotransformer 和GLORYx。基于规则的方法面临以下挑战： 首先，它们的应用领域受到转化规则覆盖范围的限制，这些规则是由专家手动编译的。当底物与规则模式不匹配时，它们可能无法生成预测。 其次，规则数量的增加可能导致低精确度表现，因为假阳性的数量增加。为了减少假阳性，一些方法基于统计分析对预测的代谢物进行排序。另一种尝试是构建机器学习模型来识别底物特异性，作为应用规则的初步步骤。然而，大多数机器学习模型仅针对Ⅰ相代谢设计。新的启发 近年来，用于化学反应预测的人工智能（AI）方法取得了显著进展[15]。一个代表性的工作是分子转换器，它将反应预测问题视为机器翻译任务。受此启发，Litsa等人提出了MetaTrans，这是一种基于转换器的深度学习方法，可以直接将母体分子转换为代谢物，绕过了提取代谢转化规则的过程。然而，MetaTrans的精确度仍然相对较低，这是代谢产物预测的一个常见挑战。 基于提示的学习是自然语言处理（NLP）中一个新兴的范式，为增强与AI系统的交互提供了机会。在NLP中，提示是提供给模型的基于文本的额外指令或上下文，以帮助它更好地理解任务并生成预期输出。Thakkar等人通过在分子的SMILES字符串中描述断开位点，将这种提示适应于逆合成预测的背景，并实现了预测准确性提高高达39%。 受此启发，作者提出了一种名为MetaPredictor的基于提示的方法，用于预测小分子的人体代谢物。本研究通过在母体分子的SMILES字符串中嵌入特定注释来引入提示。这些提示指定了代谢位点（SoMs），用于引导基于转换器的语言模型将母体分子转换为相应的代谢物。本研究还在人体代谢转化上验证了这种新方案。结果显示，与基线模型相比，性能提高了30.4%，假阳性减少了16.8%，这证实了SoMs提示可以帮助深度语言模型聚焦于关键区域，并对小分子的代谢物做出更准确的预测。unsetunset三、研究材料与方法unsetunsetMetaPredictor概览 MetaPredictor是一种无规则、端到端、基于提示的工具，用于预测小分子的人体代谢物。作者使用SMILES序列来表征分子，因此代谢物预测任务可以被视为一个序列翻译问题。作者使用seq to seq的Transformer架构进行分子语言建模。 MetaPredictor由两个模块组成：SoM识别器和基于提示的代谢物预测器。它被设计为一个两阶段方案；SoM识别器经过训练，可以自动标记小分子的SoMs，使基于提示的代谢物预测器能够兼容自动或非专家的全局预测。具体来说，作者使用SoM识别器自动标记感兴趣的母体分子的潜在SoMs，然后使用基于提示的代谢物预测器进行代谢物推断。基于提示学习的交互性质使得有效整合外部知识（包括人工提示输入）成为可能，从而实现特定酶或特定SoMs的局部代谢物预测。 作者的新方法结合了迁移学习和基于提示的学习。两个transformer模型首先在通用化学反应数据集上进行预训练，随后在代谢反应数据集上进行微调。此外，作者使用集成策略生成多样化的预测，以适应代谢预测任务。MetaPredictor的完整工作流程如图1所示。图1.MetaPredictor整体框架概览数据的收集与处理2.1化学反应数据集 用于预训练模型的数据集源自Lowe关于挖掘化学反应数据的工作，这个数据集已被广泛应用于化学反应的正向预测和逆合成分析。经过去重和筛选，最终得到了约120万个具有单一产物的训练实例。每个化学反应中的组分都使用规范化的SMILES表示，并且移除了试剂。2.2代谢反应数据集 代谢反应数据集由用规范化SMILES表示的母体分子和人体代谢物对组成。为了获得一个覆盖范围广泛的人体代谢数据集，从开放获取数据库、文献和Lee的《外源物质代谢途径手册》中收集了经实验验证且结构可用的外源和内源化合物的人体代谢物。这些开放获取数据库包括人体代谢组数据库（5.0版）、Recon3D（3.01版）、MetaCyc中的HumanCyc（23.0版）、DrugBank（5.1.10版）和BioTransformer的反应数据库（MetXBioDB）。这些代谢物是由单步酶促反应产生的。对于多步代谢转化的药物，母体分子可以是药物或药物代谢物。本研究将分解反应拆分为两个不同的训练实例，对于标记为可逆的反应，作者保留了两个方向的反应。图2.代谢数据来源 关于内源化合物的代谢反应，保留了最大公共子结构的原子数超过母体分子原子数40%的配对。这个过程可以保证代谢物与母体分子保持显著程度的结构相似性，并过滤掉相关性较低的代谢物。最后，作者合并了来自各种来源的代谢反应并去除了重复项。2.3 测试集和验证集 验证集是由Litsa等人精心挑选的，用于选择微调模型的超参数和监督模型训练，而测试集则用于评估模型的预测性能。具体来说，验证集包含了来自DrugBank的96个母体化合物的代谢反应，这些反应不仅由CYP450酶催化，还包括非CYP450酶催化的反应。测试集源自GLORY方法开发者和DrugBank手动整理的数据集，从中精心挑选了135种药物及其283个已识别的人体代谢物，以创建一个在代谢酶类型上更加多样化的测试集。考虑到母体分子可能存在不同的代谢过程，本研究确保了具有相同母体分子的实例不会被分到不同的数据分区（训练、验证、测试）中。2.4 提示工程 在本研究中，提示（prompts）是自动提取的，用于替代手动标记来构建模型。首先，使用RXNMapper对化学反应数据集和代谢反应数据集中的所有反应进行原子映射，以找出在反应过程中哪些原子的原子环境发生了改变。原子环境包括该原子及与之相连的所有键。对于给定的原子映射代谢反应SMILES，母体分子和代谢物之间环境不同的原子在母体分子的SMILES中被标记，并作为对应于代谢位点的提示。类似地，对于给定的原子映射化学反应SMILES，反应性原子在反应物SMILES中被标记为提示表示。需要说明的是，提示是通过使用SMARTS表示法[:1]引入的，其中''代表任意原子。在提取提示后，原子映射信息被移除。模型的构建 作者处理了不同的数据集用于预训练和微调，以确保它们与代谢位点（SoM）识别器和基于提示的代谢物预测器的源序列和目标序列相一致。对于基于提示的代谢物预测器，本研究在通用化学反应数据集上进行了预训练，其中反应物的SMILES被标记了反应性原子，然后在代谢反应数据集上进行微调，其中母体分子的SMILES被标记了代谢位点（SoMs）。至于SoM识别器，本研究也使用了迁移学习策略来获取关于原子反应性的通用化学知识。为了使SoM识别器能够自动标记SoM，源序列是母体分子的SMILES，目标序列是带有SoM标记的母体分子SMILES。 所有模型都使用监督学习和seq-to-seq Transformer架构，该架构部署在OpenNMT-py库2.3.0版本中。作责研究中采用的迁移学习策略是将预训练模型作为微调模型的起点。对于预训练模型的参数，基于分子transformer进行了微小的调整。关于微调模型，作者尝试了包括SMILES增强策略和批量大小在内的各种参数，并根据验证集的准确性选择模型。模型的优化 考虑到药物可能通过不同的酶形成不同的代谢物这一事实，本研究基于标准束搜索算法构建了集成模型，以从综合角度推断多个可能的序列。4.1束搜索 束搜索算法是一种基于启发式的流行搜索算法，它探索所有可能的字符并维持k个最可能的序列。通过应用束搜索算法，代谢位点（SoM）识别器可以为给定的母体化合物生成几个潜在的代谢位点预测，而基于提示的代谢物预测器可以为带有SoM提示的药物预测多个代谢物。通过调整束大小，可以改变生成预测的数量。为了在扩大搜索空间和获得最佳预测结果之间取得平衡，本研究尝试计算所有微调模型在不同束大小下在验证代谢集上的预测性能。实践表明，3到15之间的束大小能够在精确度和召回率之间提供适当的折中。4.2集成策略 为了减轻单一模型性能可能存在的偏差，研究中经常使用集成策略来获得更好的泛化性能。在本研究中，作者采用了通过平均来自几个在不同策略下训练的解码器的预测分布的方法。为了在最大化正确预测率和保持低假阳性率之间取得平衡，作者通过组合不同的单个模型构建了集成模型，并在验证集上评估了它们的模型性能。最终，作者为代谢位点（SoM）识别器选择了四个单个模型，为基于提示的代谢物预测器选择了五个单个模型。作为补充说明，由于对某些预测进行了过滤，单个模型和束大小为k的集成模型的输出大小被限制为最多k个。模型评估 MetaPredictor的性能评估主要依赖于其预测代谢位点（SoMs）和代谢物的准确性，以及预测的假阳性率。准确性主要通过计算召回率来评估，即模型正确识别的参考代谢物数量与参考代谢物总数的比率。精确度表示假阳性的百分比，通过正确识别的参考代谢物数量与模型输出大小的比例来计算。本研究还计算了每个模型至少正确识别一个、至少一半和全部SoMs或代谢物的输入分子百分比，以观察药物检索的范围。 为了评估算法对解码序列进行排序的有效性，本研究在模型生成N个预测时计算了top-N指标。对于基于提示的代谢物预测器的评估，作者比较了预测代谢物和参考代谢物之间的指纹相似性。如果预测代谢物和参考之间基于指纹的Tanimoto系数等于1，则认为预测是正确的。unsetunset四、研究结果与讨论unsetunset代谢数据分析 用于训练模型的人类代谢反应数据集包含14,782对独特的母体分子和代谢物对。为了更好地理解这个代谢反应数据集，作者从代谢酶的EC分类（EC-level 1）分布角度分析了代谢反应。尽管数据集中有相当一部分没有提供代谢酶信息，但如图2A所示，所有酶类都在标记的对中有所涵盖。在酶的分布中，氧化还原酶（EC1）、水解酶（EC3）和转移酶（EC2）是最突出的类别。前两者主要负责催化I期药物代谢，而后者在II期药物代谢中占主导地位。 作者使用降维算法TMAP对代谢反应数据集进行了可视化。如图3B所示，每个点代表一个基于反应指纹RXNFP计算的相似性的代谢反应。通过EC分类号对TMAP进行颜色编码，这些酶家族形成了相对离散的反应簇，可以观察到大多数未标记的代谢反应可能是由氧化还原酶、转移酶和水解酶催化的。上述分析表明，代谢反应数据集涵盖了全谱的酶，并如预期的那样偏向于药物代谢领域最常见的催化酶，从而为本研究的模型提供了关于代谢酶范围和特异性的训练基础。图3.代谢数据分析（a）数据类型分类（b）TMAP降维可视化模型评估2.1和基准模型比较 首先，通过并列展示四种模型在代谢物测试集上的表现来评估迁移学习和集成策略的有效性，如图4所示。由于MetaPredictor的两个模块处理不同的预测任务，作者基于每个模型在生成五个预测时识别参考SoMs或代谢物的能力进行了单独比较。图4的结果强调了迁移学习和集成策略的重要性。 具体而言，组成集成模型的单个微调模型的平均性能相比预训练模型和仅在代谢反应数据上训练的模型显示出明显的改进。以基于提示的代谢物预测器为例，采用迁移学习导致召回率平均增加37.65%，精确度平均提高6.2%。此外，预训练模型的预测中无效SMILES的比例较低，但不合理SMILES的比例最高，这表明虽然预训练模型理解一般化学反应规则和SMILES语言的语法，但缺乏代谢转化的必要专业知识。另外，仅在代谢数据上训练的模型在预测准确性方面表现不佳，预测了最多的无效SMILES，但不合理SMILES的数量比预训练模型少。这反映了由于数据量相对较小，尽管获得了一些代谢转化的知识，但一般化学训练仍然不足。 此外，可以看到集成模型不仅实现了更广泛的药物检索范围，而且相比平均单个模型，还实现了更高的召回率和更低的假阳性。具体来说，SoM识别器的集成模型将召回率提高了8%，精确度提高了2.6%。这表明集成策略是一种有效的方法，可以增强输出多样性并减少假阳性的发生，同时不增加模型的输出大小。 为了评估引入指定SoMs的提示在改善代谢物预测任务中的有效性，作者使用相同的数据集和方法，但不引入提示信息来训练基线模型。具体来说，基线模型的源序列是母体化合物的SMILES，不包含用于标记发生代谢转化的原子的SMARTS符号[*:1]。基于提示的代谢物预测器可以使用额外的输入提示来指导代谢物转换，而基线模型仅基于训练数据集中代谢转化的潜在概率分布生成代谢物预测。然后，作者比较了基线模型和基于提示的代谢物预测器在代谢测试集上的性能，结果如图5所示。 在两个模型的输出大小相当的所有情况下，基于提示的代谢物预测器都实现了优于基线模型的性能指标。具体来说，实施引导提示导致召回率平均增加30.4%，精确度平均增加16.8%。这表明引入提示学习确实可以指导模型生成更准确的代谢物预测，同时减少假阳性的发生。图4.在TOP5上的模型评估结果图5.输出大小相同时，各部分模型的比较结果2.2和其他工具比较 本研究通过与四种现有的药物代谢预测工具（GLORYx、SyGMa、BioTransformer和MetaTrans）进行比较，评估了MetaPredictor的性能。评估基于单步反应生成的代谢物，使用指纹相似性进行评估，并比较了top-N（N = 5, 10, 15, 12）的预测结果。表1.同类模型比较结果 如表1所示，MetaPredictor展现了出色的预测性能和排序能力，对数据集有更大的覆盖范围。这种优势可能源于MetaPredictor不依赖于化合物和代谢规则的精确匹配进行预测，而基于规则的方法则需要。MetaPredictor的改进归因于引入了提示学习和扩大了训练数据集。图6.（a）各种类型酶预测正确的数量（b）对不同分子量大小预测的召回率（c）不同SOM分子的预测召回率 图6A展示了各方法在不同酶家族（氧化酶、转移酶、水解酶）方面的性能。MetaPredictor在识别与氧化反应和II期代谢相关的代谢物方面表现优异，并能正确预测更多的水解酶代谢物。此外，MetaPredictor还能预测一些不常见酶催化的代谢物，如芬戈莫德、苯肾上腺素和森尼定的代谢过程。 图6B反映了不同方法在各分子量范围内的性能。MetaPredictor在所有范围内都显示出有竞争力的召回率，特别是在<200 Da和350-500 Da范围内表现突出。图5C则展示了各方法对不同SoM数量化合物的预测能力。MetaPredictor对含1-5个SoM的化合物显示出良好的召回率，尤其是对单SoM化合物的预测最为准确。2.3具有困难的案例 为了评估MetaPredictor的能力和局限性，作者分析了代谢测试集中的预测结果。模型错误预测与参考代谢物之间的最大平均相似度为0.73，图7展示了几个典型案例。 1.有时，预测与参考代谢物的差异仅在于一个非反应性原子，如DBMET00112的案例（图7，案例1）。模型正确预测了羟基化反应位点，但错误地替换了一个非反应性基团。 2.在某些情况下，错误可能源于参考代谢物本身，如噻氯匹定和甲氧氯普胺的案例（图7，案例2和3）。文献表明，这些药物可能有多种代谢途径，而本研究的模型成功预测了其中一些未被收录在数据库中的途径。 3.对于连续氧化反应，模型有时会预测中间产物而非最终代谢物，如TAK-438的案例（图7，案例4）。模型预测了醛，而参考代谢物是相应的羧酸。 4.模型在处理多步骤、多位点转化的代谢物时面临挑战，如莫尔西多明和安非他酮的案例（图7，案例5和6）。这是因为模型主要基于单步代谢反应进行训练。 尽管存在这些不完全匹配的情况，作责的分析表明，模型预测仍为药物代谢研究提供了有价值的见解。在许多案例中，模型成功识别了反应类型、代谢位点，甚至代谢反应的中间体。这些结果突显了MetaPredictor在药物代谢预测领域的潜力，同时也指出了未来改进的方向。图7.具有挑战的案例2.4 模型注意力机制分析 通过注意力权重的可视化，可以在一定程度上看到模型是如何学习代谢转化的。图8展示了模型在预测母体药物分子的代谢物时，对其SMILES中的标记分配的注意力权重。较高的注意力权重表明模型认为这些特定标记对于做出预测更为重要。 从图8A可以看出，基于提示的代谢物预测器对[CH2:1]标记分配了较高的注意力权重（颜色更深）。这表明模型聚焦于与母体药物分子的代谢位点（SoM）相对应的关键标记，从而预测出正确的参考代谢物序列。 相反，图8B中的热图显示了一个更分散的注意力模式，对未参与代谢转化的原子分配了较高的权重。这可能揭示了为什么基线模型由于缺乏有效识别准确代谢物预测所需关键区域的能力而产生了不正确的预测。 这种比较进一步突显了引入基于提示学习的优势。通过提供SoM提示，模型可以更好地被引导关注对代谢转化至关重要的标记，从而生成更准确的预测。图8.（a）基于提示的模型正确预测的分子（b）基线预测模型的错误预测分子unsetunset五、总结与展望unsetunset 本研究提出了一种名为MetaPredictor的基于提示的学习方法，用于预测人体内小分子的代谢物。MetaPredictor由SoM识别器和基于提示的代谢物预测器两个transformer模型组成，旨在实现工作流程自动化，扩大应用范围。 首次将SoM提示与深度语言模型结合，丰富领域知识并引导准确的代谢物预测。在top-5预测中，MetaPredictor实现了89%的召回率和22.6%的精确度，比基线模型性能提高约30%，假阳性减少16%。这证明了基于提示学习的有效性，部分解决了代谢物预测中精确度下降的问题。本研究还利用迁移学习获取化学反应的通用知识，解决人体代谢反应数据有限的问题。 尽管未专门针对药物代谢数据训练，MetaPredictor仍显示出与特定药物方法相当或更好的性能。它能预测非常见酶催化的代谢物，扩大了模型的预测范围和泛化能力。MetaPredictor在各种分子量范围内表现稳定，能有效处理不同SoM数量的化合物。 MetaPredictor的推理方式类似人类专家，使用基于提示的语言提高了可解释性。与无规则方法相比，它提供更全面的药物代谢分析，并允许整合外部知识。这种"人在回路"的方法提高了预测准确性，为药物发现领域的深度学习模型提供了新的知识引入范式。unsetunset论文链接unsetunset https://doi-org.libproxy1.nus.edu.sg/10.1093/bib/bbae374unsetunset代码链接unsetunset https://github.com/zhukeyun/Meta-Predictor 投稿人：周艺凌 责任编辑：许燕红