Добавил:
kiopkiopkiop18@yandex.ru t.me/Prokururor I Вовсе не секретарь, но почту проверяю Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз: Предмет: Файл:
Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_5364_Библиотеки_им_академика_М_И_Перельмана.pdf
Скачиваний:
0
Добавлен:
02.09.2026
Размер:
21 Мб
Скачать
11 Ultra-Large-Scale Virtual Screening 305
Our Giga-Scale VS Experience
We started our rst venture into billion-scale VS in 2021. On the surface, the
project seemed simple: We knew how to run ligand prepar ation and docking
pipelines for millions of compounds. All we needed to do was to think a little
bigger this time! We quickly realized that generating 3D conformers for over
1.5 billion compounds and docking them to a protein target on an HPC cluster
was not as simple as that. HPCs are shared environments and rely on fair
sharing of computational resources among all users, so we had to ensure that
our study would not disrupt anyone elses work [33]. Tools were not designed
for handling the huge input, data sizes began to surge, and temporary les
started clogging up the storage space and straining the le system.
The lessons learned in this endeavor went into the preparation of this
section to h elp others avoid similar pitfalls [23, 33].
3.1 Implications of Dataset Size
When working with enumerated chemical libraries, compounds will c ommonly be obtained in 1D (SMILES) or 2D (sdf) representation. On the million scale, storing this data is usually straightforward. However, ultra-large enumerated libraries begin to require a considerable amount of disk space: For example, SMILES strings of
1.56 billion lead-like compounds used in our work had a total size of 89 GB, and 10 billion SMILES consistently were reported to take up around 1 TB of disk space [4, 23].
To speed up the download process and make the le(s) easier to handle, ultra­large libraries are usually offered in a compressed format (e.g., zipped) and in multiple parts. If the compression was performed on individual parts, and depending on the intended use of the library, you may not need to store the fully uncompressed data on your disk. Nonetheless, keep in mind that uncompressed data will be several­fold larger than the compressed les, depending on the data format and compression algorithm used. For instance, the compressed version of the 1.75 billion SMILES of SAVI takes up 440 GB, while the fully uncompressed data would require 4.4 TB of storage space. The 95 million compounds of CHIPMUNK in compressed sdf format stand at 109 GB and inate to 588 GB when uncompressed. Thu s, while one may simply obtain, decompress, and work with smaller libraries, it is advisable to carefully consider the amount of required disk space when handling ultra-large libraries [8]. Some space can often be saved by limiting decompression to parts that are actively being worked on.
Many VS approaches rely on the availability of 3D conformations for the small molecules. When obtaining ultra-large libraries in 1D or 2D representation, 3D conformers rst need to be generated. This step, unsurprisingly, is heavily affected by the library size and requires careful consideration. First, the number of structures will increase during conformer generation. Therefore, the storage needs and the time
306 I. Pöhner et al.
required to process the library in subsequent steps will grow. How many conformers per molecule to generate is ultimately method- and case-dependent, but, as a rule of thumb for ultra-large approaches, one should strive for the lowest acceptable number of conformers for the chosen method.
For example, the limited conformer generation with a maximum number of eight tautomers and four stereoisomers per tautomer reported by Sivula et al. [23]inflated the number of structures in a 1.56 billion lead-like SMILES-based library around 2.4 times. In this study, the resulting 3D conformer databases in the proprietary Schrödinger Phase database format were 10.3 TB while compressed, and around
22.1 TB when uncompressed. These sizes illustrate the massive increase in disk space requirement when converting from 1D (89 GB) to 3D (22.1 TB). Similarly, Michino et al. [21] converted a version of Enamine REAL with 1.12 billion compounds into binary 3D structure les for their GPU-based shape-screening approach. In their case, they generated 404 les of 15 GB each, totaling 6 TB. Thus, although the exact difference is highly format-dependent, even the generation of a single 3D structure per SMILES string would already signicantly increase the data size, and the problem is aggravated by the compound number ination during conformer generation.
Structure ination therefore needs to be borne in mind when planning 3D ligand or receptor-based screening campaigns. In addition to the 2.4-fold ination reported by Sivula et al. [23], Gentile et al. [34] report conformer generation to inate their
12.3 billion input SMILES by over three-fold to around 38.7 billion unique struc­tures. Thus, the number of structures to process in subsequent steps of the VS workow can easily double or triple and more structures inevitably cause increased time- and resource requirements.
A nal size-related consideration affects the 3D conformer generation and all the following calculation steps: Many tools will rst read all input molecules from one le into memory. This can become a limiting factor when working with ultra-large libraries, as the le sizes may exceed the available memory. Thus, it is essential to assess the memory requirements of different calculation steps, and the dataset needs to be split into appropriate batches to reect the available resources.
3.2 Where and Ho w? Guidelines for Planning and Running
Ultra-Large VS Campaigns
In light of the limited throughput and the massive disk space and memory require­ments discussed above, it is hardly surprising that most ultra-large-scale screening campaigns utilize HPC or cloud computing resources. While they have become more readily available over the past few years, massive computing resources often remain costly [5, 22, 35]. Consequently, when planning and executing an ultra-large screening campaign, the computing environment choice is entangled with the screening timeframe and budget. You may need to benchmark your tool(s), optimize
11 Ultra-Large-Scale Virtual Screening 307
the performance for the chosen computing environment, select appropriate data splits, and monitor progress and potential errors. Below, we outline some key questions we and others have found helpful in planning VS on the ultra-large scale and navigating the specic challenges of data-intensive screening campaigns.
Should one use CPU or GPU? This question is often predetermined by your choice of tools. However, workows can sometimes run on either CPU or GPU, and dedicated GPU implementations of otherwise CPU-bound tools exist (e.g., the docking tool AutoDock4 and its corresponding GPU-reimplementation AutoDockGPU) [36, 37]. GPUs are best suited when tasks can be split into many independent subtasks to execute in parallel. With the independence criterion met, GPUs excel at processing particularly high volumes of data. On the other hand, tasks that involve sequential steps and require communication between different subtasks are more suitable for processing on CPUs.
When faced with the choice, throughput may dictate your decision, but cost and availability should also be considered: GPUs tend to be more expensive and, in an age of deep learning (DL) popularity, often more sought after. For example, as a rule, the prediction step in a DL workow is faster on GPU. However, if CPU resources are signicantly cheaper and awarded more swiftly, it may be more economical to run predictions on a larger number of CPUs to offset the speed difference [23]. Tech­nically, it would also be possible to run the training step on CPUs. However, the speed difference is much more pronounced during training, making training on CPUs impractical. When you choose the processing hardware for your ultra-large­scale screening workows, we recommend factoring in the different costs and queuing times for different types of resources in planning the various workow stages.
Get to know your tool(s)Proling and parallelization When planning an ultra­large VS campaign, a prerequisite for optimizing workows and budgeting time and resources is a clear picture of the expected resource consumption and behavior of your tool(s). In most HPC and cloud computing environments, you are required to specify and reserve computing resour ces, for example, the number of CPUs and GPUs and the memory required by your job. Resource reservations will often directly determine your billing but also affect, for example, your queuing time. In a nutshell, you want to ensure sufcient resources for your workow to run without problems but avoid a major resource overhead youd pay for without being able to utilize it.
Most computing environments offer special queues or instances for testing to help you determine how many CPUs/GPUs and how much memory to reserve. Test environments often have limited resources and a short wall time, but are awarded swiftly and enable you to prole tools in the relevant production environment. Some environments will offer graphical tools for monitoring the resources consumed by your test job(s). Alternatively, HPCs commonly rely on batch job management tools such as SLURM, which can provide an overview of resource consumption with dedicated commands such as seff, sacct, or scontrol. It is worth familiarizing yourself with the relevant proling commands of the tool used in your environment to review,
308 I. Pöhner et al.
e.g., CPU and GPU utilization and memory consumption of your job. Finally, as almost all HPC environments are UNIX-based, handy command-line tools such as time, (h)top, and free can help you in monitoring runtime and resource consumption in the UNIX command line .
In practice, to decide on the number of computing units to reserve, start by determining the scaling behavior of your tool(s). Is your tool single-threaded or internally parallelized, that is, can it utilize more than a single CPU (or GPU)? If so, ideally, you would nd a linear scaling behavior: Whenever you double the number of computing units, the computing time halves [7]. However, many tools do not (fully) scale linearly or have a hard limit on the number of computing cores they can utilize. In the latter case, when exceeding, for example, 8 CPUs, runtime will no longer improve. Efcient usage of multiple GPUs is, as a rule, more challenging to implement. Thus, many GPU-based tools rely on using only a single GPU.
Often, you can nd hints on the scaling behavior from method publications, published use cases, repositories, or online documentation. Alternatively or addi­tionally, you can also prole the behavior of your tool(s) yourself in a test environ­ment. Time a small test run with a diverse random set of compounds with incrementally increasing xed numbers of computing units (set in your test reser­vation and the toolmost tools will have settings/command line switches when internally parallelized). It may also be advisable to use a couple of different dataset sizes and monitor how the memory footprint is affected by input data size.
Once you established a CPU/GPU count where your tool scales favorably and assessed its memory footprint in dependence on the input data, you can put it all together and plan the degree of external parallelization, that is, decide how many instances of your tool to run in parallel in the production system. Keep license limitations in mind at this stage, where they apply. Having determined the number of instances and required CPU, GPU, and memory resources, you should now be able to estimate the overall runtime depending on your chosen compound librarywhich leads to the question of how to distribute an ultra-large dataset between the different parallel instances.
Splitting your input data As with any rule, there are exceptions, but most of the time, you want to split your ultra-large input data into at least as many chunks as processing instances launched. Keep in mind that different compounds have differ­ent processing speeds and most similar speeds can be expected for closely related compounds. Thus, in particular for combinatorial enumerated libraries, you should randomize the compound order in the workow input. As a rule, compounds with high molecular weight and/or many rotatable bonds require higher processing time. Randomization will help to avoid idling resources due to compound batches with many compounds having a signicantly slower processing speed.
For similar reasons, it may also be benecial to serially run several smaller batches rather than one larger batch, as individual runtime differences would decrease and be less likely to result in idling resources. However, note that many HPCs use distributed le systems such as Lustre. One caveat is their unfavorable behavior with heavy le I/O stemming from large numbers of small les. Thus, if
11 Ultra-Large-Scale Virtual Screening 309
you split your data into too many chunks, your runtime may suffer a negative impact from slow le system response (be also mindful of temporary les in this context). In addition, HPCs often offer fast local storage solutions based on NVMe/SSD, that can help to circumvent le I/O problems.
Optimal le sizes should ensure that memory does not become an issue and individual jobs can nish processing well within the system wall time. If the le system negatively affects the performance of your workow, you can often coun­teract: Many computing environments offer fast local storage resources, and solu­tions like RAMDisk can move load off the le system and into memory.
An exception applies when you fetch compound data from remote databases rather than working with les. In this case, network speed will inuence your throughput. You may want to run separate queries for each instance of your tool. If you massively parallelize and run thousands of instances and queries simulta­neously, this can in itself create problems when operating on a single ultra-large database. For databases, a counteracting measure is a concept called sharding, reminiscent of batch creation for les, where you dist ribute data into individual smaller databases.
Finally, if you have different tools in your workow, you may want to separately optimize the size of compound batches/shards for each tool, as different tools can have rather different throughput and requirements [4].
3.3 Ultra-Large VS in a Nutshell: Key Questions
and Trouble-Shooting at a Glance
For all of the pitfalls we discussed above, a lot more could be said, which would be beyond the scope of this work. We hope, however, to provide a starting point and relevant keywords for prospective users. To summarize, if you plan on embarking on your own ultra-large VS campaign and can answer the following questions, you are off to a good start:
Questions to answer when planning ultra-large VS
How experienced are you in running workows in an HPC/cloud environment, and is there user
support available for optimization or troubleshooting? Are your tools single- or multi-threaded? How do the tools scale with multiple CPUs/GPUs? What is the memory footprint, and how is it affected by input data sizes? Are there any license limitations? How many tools can you run in parallel in available queues/instances of your computing
environment? What is the available wall time? Is your input data read from les or a remote database? Do you require specialized storage solutions to circumvent heavy I/O load on distributed le
systems?
310 I. Pöhner et al.
Table 11.2 Typical VS performance bottlenecks, how to recognize them, and potential countermeasures
Type Symptoms Solutions CPU/
GPU bound
File I/O bound
Memory bound
Network bound
100% utilization on one or more CPU cores/GPUs, potentially while other reserved ones idle
Low CPU utilization, high le system utilization
Low CPU utilization with high swap memory usage, program crash
Low CPU usage, low le I/O Prestage or preload data prior to
Increase parallelization by running more instances or increase method internal parallelization (if possible). Reserve/ spread over more computing resources
Use fast local storage or fast SSD stor­age. Move (temporary) les to RAMDisk
Reserve more memory or reduce mem­ory usage by processing smaller chunks of data
processing
If, despite your best efforts, you nd your performance unsatisfactory, Table 11.2 summarizes typical performance bottlenecks and how to recognize them and pro­poses potential solutions.
Arguably, different methods have their own specic challenges and not all of them can be covered in this work. Thus, consult other campaigns using the tools of your choosing. The following sections will provide a vast set of examples to help you inform your approach. Additionally, for further reading, several published examples discuss lessons learned from their ultra-large screening or key challenges faced in method development for large-scale screening (e.g., references [4, 7, 23]), and published protocols provide step-by-step instructions (e.g., references [24, 38]).

4 Ligands on the Ultra-Large Scale

At the heart of any ultra-large VS are the ultra-large ligand libraries. As we discussed ultra-large chemical libraries and spaces, we established their massive sizes and that specialized approaches for liga nd-based VS are required when handling nonenumerated chemical spaces. This section will discuss examples of ligand­based 2D searches in enumerated and nonenumerated compound collections and dive deeper into 3D conformer generation strategies and 3D ligand-based VS.
4.1 Ultra-Large 2D Similarity Searches
In an era of ultra-large make-on-demand chemical libraries and vast commercially available nonenumerated chemical spaces, the search for analogs and substructure
11 Ultra-Large-Scale Virtual Screening 311
Fig. 11.1 Schematic illustration of a classical ngerprint-based search. The query molecule is transformed into a bit vector describing its structural features, such as secondary amine(blue) and pyridine heterocycle(green). In the example, different structural features (bits) are represented by different colors, where white indicates that the bit is off (i.e., the feature is not present). Once generated, the query bit vector is compared one by one with similarly constructed bit vectors for every compound in a large compound database
matches of known ligands represents a straightforward way to study structure– activity relationships (SARs) and extend the hit pool.
2D similarity searches achieve particularly high throughput and require limited to no preparation of the ultra-large libraries. They are rarely applied as the sole mode of VS, but their throughput makes them particularly valuable as prelters to more computationally demanding techniques. However, they require prior knowledge of compounds/scaffolds with the desired properties and have limited potential in proposing entirely novel scaffolds. Additionally, as also illustrated schematically in Fig. 11.1 on the example of classical ngerprint-based searches, many approaches to 2D similarity searches rely on an exhaustive pairwise comparison and become slow and impractical when proces sing ultra-large libraries [ 9 ].
To enable efcient searches in ultra-large libraries, recent works have proposed various novel approaches and strategies that address the limited throughput. These involve, for example, novel types of ngerprints, the use of specialized formats and database operations, graph-based search methods, or searches in the reagent space [ 9 ].
In classical ngerprint-based searches (see Fig. 11.1), the bulk of the computation time is spent on ngerprint generation. Thus, precomputing ngerprint bit popula­tion statistics can speed up the process (e.g., in chemf p [39] and MolSofts Giga­Search [40]), and some software vendors provide precom puted ngerprints for ultra­large databases to remove the ngerprint computing step altogether. Examples include Schrödingers FP-Sim GPU and OpenEyes Molecules as a Service (MaaS) module in their cloud-based Orion software [41, 42]. Schrödingers GPU Similarity search is available for billion-scale library screening via LiveDesign and has also been released as an open-source incentive version (see Table 11.10 in the Appendix) [9, 41, 43]. With four GPUs, Schrödingers GPU Similarity search was able to screen 1 billion molecules in around 0.2 s [43]. OpenEyes MaaS achieved a 2D similarity search of 800 million Enamine REAL molecules in 3 s or less. However, to achieve this processing speed, data in MaaS is held in memory. MaaS therefore required an Amazon Web Services (AWS) instance with 768 GB of
312 I. Pöhner et al.
memory, which underlines how memory limitations can become major bottlenecks even for these relatively fast approac hes [9].
The recent release of ZINC20 strives to support direct similarity searches by relying on NextMove Software Ltd.s tools, SmallWorld and Arthor [20, 44]. SmallWorld utilizes maximum common subgra phs and graph edit distance in their graph-based similarity search. Importantly, using graph-based instead of ngerprint-based approaches is more suitable for and adaptable to large databases [9, 44]. Arthor is focused on high-performance SMARTS pattern and substructure matching and scales favorably on chemical libraries with hundreds of millions of compounds while handling relevant tautomers and resonances [9, 45]. ZINCs choice of tools underl ines that, while not every similarity search tool ts the bill, efcient algorithms to sift through the massive state-of-the-art ultra-large libraries exist.
Another strategy tackles huge make-on-demand combinatorial chemical spaces by relying on their building blocks. Over 20 years ago, a graph-based strategy, FTrees-FS, was developed [46]. This method relies on so-called Feature Trees, and its searches conceptually resemble pharmacophoric-feature-based screens. Only building blocks and corresponding reactions that match the query get enumerated into virtual compounds, which allows for the efcient screening of ultra-large libraries [9, 46].
A more recent topological ngerprint similarity search called SpaceLight can, likewise, take advantage of combinatorial chemical spaces like the Enamine REAL Space [9, 47, 48]. It can process more than 20 billion compounds within seconds on a desktop computer and scales with the number of fragments/building blocks, rather than the number of virtual products [9]. Finally, another tool, termed SpaceMACS, was developed to enable a search for specic structural features within ultra-large chemical spaces [9, 49]. Since SpaceMACS rst loads spaces into memory, it requires signicant memory resources, but once loaded, ultra-large libraries and spaces can be screened for desired structural features within seconds [9].
FTrees-FS has, for example, been one of the various screening strategies employed by Glaab et al. [50] to discover novel inhibitors of SARS-CoV-2 main protease (Mpro, or 3CLpro in their work). They performed a ligand-based similarity search of close to 900 million compounds, largely sourced from ZINC15, which were compared to previously reported inhibitors of SARS-CoV-2 and SARS-CoV Mpro. Br üschweiler et al. [51] aimed to target the DNA -binding domain of the nuclear factor erythroid 2-related factor 2 (NRF2). The binding pocket of NRF2 has limited druggability and consequently generally displays modest hit rates at best. Thus, to nd additional analogs of their initial fragment-based screening hits, the authors used the inniSee software, similar in concept to FTrees-FS, to search 15.5 billion virtual compounds from the Enamine REAL Space [9, 26, 51].
Given their dependency on previous knowledge and limited generalizability, one should take particular care when selecting the template (sub)structure(s) for 2D similarity searches. At the same time, the presented application examples demon­strate how 2D similarity searches can act as powerful screening tools in conjunction with other methods or support the extension of the hit pool in later stages of VS.
11 Ultra-Large-Scale Virtual Screening 313
4.2 Recent Examples of Ultra-Large 3D Ligand-Based
Screening
Recent examples of 3D ligand-based screening cover, in particular, ultra-large 3D shape similarity searches, as schematically illustrated in Fig. 11.2, which commonly leverage GPUs for increased throughput.
For example, Grebner et al. [4] utilized the GPU-based FastROCS method in OpenEyes cloud-based platform Orion to screen subsets of the Astra Zeneca virtual chemistry space with up to 12.7 billion compounds. The shape similarity searches were initiated from known crystallographic ligands of ten diverse target receptors, including G Protein-Coupled Receptor Kinase 2, estrogen receptor, dipeptidyl peptidase 4, and MEK1 kinase. In their work, the authors present a comprehensive benchmark of FastROCS on AWS. They report detailed database generation and shape search timings and resource requirements for screens of varying sizes: Screen­ing 51 million, 172 million, 1.33 billion, and 12.7 billion molecules, they demon­strate the scalability of 3D shape similarity searches [4].
To address the growing need for novel anti-mycobacterial agents, Michino et al. [21] targeted the challenging mycobacterial lipoamide dehydrogenase (Lpd). An extensive experimental high-throughput assay of around 2.5 million compounds had identied only two lead compounds. To leverage recent advances in ultra-large VS and extend the hit pool, the authors utilized the GPU Shape screen method by Schrödinger to screen 1.12 billion compounds from Enamine REAL. Around 800,000 candidates entered a post-ltering process including molecular docking, and a nal set of around 100 candidates was purchased, yielding six novel hit compounds with IC narrow chemical space, the new hit compounds displayed some variation compared to previously known inhibitors. The authors did not report the time invested in
s between 5 and 100 μM. While Lpd requires a rather specic
50
Fig. 11.2 Schematic illustration of a shape-based search. First, multiple 3D conformers are generated for the query molecule (only one is shown exemplarily in the illustration). For each conformer, molecular shape and, often, 3D-pharmacophore-type features are determined. Next, the overlap between the query shape/feature ensemble and predetermined shape and feature descriptors of all compounds in a large database can be computed
314 I. Pöhner et al.
Table 11.3 Summary of recent ultra-large 2D and 3D ligand-based VS campaigns
Compounds
15.5 billion NRF2 inniSee NA Brüschweiler
Up to 12.7 billion
1.12 billion Lpd GPU
900 million SARS-CoV-2
Reported throughput covers only the screening process and does not account for database preparation NRF2 nuclear factor erythroid 2-related factor 2, Lpd mycobacterial lipoamide dehydrogenase, NA not available
Target (s) screened Tool Approx. throughput Reference(s)
et al. [51]
10 diverse targets
Mpro
FastROCS 480,000 - 1.5 million cpds/
min/GPU 120,000 cpds/min/GPU Michino et al.
shape FTrees-FS NA Glaab et al. [50]
Grebner et al. [4]
[21]
library preparation, but the shape screen achieved an impressive throughput of around 120,000 compounds per minute per employed GPU [21].
Both 2D and 3D ligand-based VS examples are summarized in Table 11.3.
4.3 3D Conformer Generation on the Ultra-Large Scale:
Shortcuts and Acceleration Strategies
Whenever approaches like ligand shape similarity searches or docking utilize the ligand 3D structural infor mation, ligand conformer generation becomes the rst step of the VS workow. As already discussed in Sect. 3.1, conformer generation will noticeably increase the amount and size of the ligand data. Additionally, when working with ultra-large libraries, a signicant portion of computing time and resources will be spent on the generation of the 3D conformers.
For example, to use the GPU-based Rapid Isotere Discovery Engine (RIDE) by MolSoft [52] for a 3D similarity search of SAVI, the necessary generation of a conformationally expanded version of the database consumes around eight million CPU hours [9]. On the other hand, once the 3D ligand database is created, RIDE can achieve a screening performance of around 500,000 conformations per second per GPU. Within hours, the full SAVI can thus be screened against a search template [9]. 3D conformer generation is a single-time effort as the conformationally expanded database can be used for any number of searches.
Thus, beyond following the general guidelines outlined in the previous section to optimize ligand conformer generation parameters and tool performance, obtaining pregenerated 3D conformers can save time and resources. In the spirit of reusable digital assets, the results of such one-time efforts are increasingly made available to remove the conformer generation step from othersworkows altogether.