Brief Overview of Foundation Models and LLMs

[Foundation models](https://www.google.com/url?q=https://crfm.stanford.edu/assets/report.pdf&sa=D&source=docs&ust=1725032576800328&usg=AOvVaw3BUJnqgxl1e3mm5WSINKIc) [1] are machine learning models pretrained on large (possibly unlabeled datasets) through the use of self-supervised learning. This pre-training curriculum enables foundation models to be applied to a wide range of downstream tasks. Instead of building AI models from scratch, researchers can use foundation models as a starting point to develop specialized applications more quickly and cost-effectively. In recent years, such models have gained significant interest due to their transformative impact on fields like natural language processing, chemical analysis, and drug discovery. Foundation models that are trained on vast amounts of web data are capable of understanding and generating human-like text. Large Language Models (LLMs), such as BERT and GPT-4, are examples of popular foundation models.

Of course, foundation models don’t only have to be trained on web data. Scientific foundation models, trained on scientific data, are increasingly powerful tools for scientific discovery in general and drug discovery in particular. The [SciFM initiative](https://scifm.ai/) shows how these models can drive research and innovation by supporting collaboration across various disciplines. However such models remain difficult to use for practical scientific applications. Deep Forest Sciences’ Prithvi™ no-code platform enables scientists to fine-tune scientific foundation models for their applications without having to write sophisticated code. In this blog post, we will introduce Prithvi™’s support for no-code foundation model fine-tuning and explain how it can be a powerful tool for drug discovery.

A schematic representation of a foundation model

Deep Forest Sciences in partnership with the DeepChem community, has led the development of the large chemical foundation models [ChemBERTa](https://arxiv.org/abs/2010.09885) and [ChemBERTa-2](https://arxiv.org/abs/2209.01712). ChemBERTa, based on the RoBERTa transformer model, was originally developed for molecular property prediction. This model leverages a large dataset of 77 million SMILES strings from PubChem for self-supervised pre-training. The model uses [HuggingFace](https://huggingface.co/seyonec/ChemBERTa-zinc-base-v1) for streamlined pretraining and fine-tuning and is evaluated on the [MoleculeNet](https://github.com/deepchem/moleculenet) benchmark. [2]

ChemBERTa-2 builds upon the foundations of ChemBERTa by optimizing the pretraining process [3]. The model employs masked-language modeling (MLM) and multi-task regression (MTR) to learn molecular fingerprints, leveraging the simplicity and widespread availability of SMILES strings. The pretraining improvements in ChemBERTa-2 shows that scaling pre-training datasets can significantly improve the performance of the downstream tasks.

Schematic overview of ChemBERTa-based modeling

Despite their potential, fine-tuning LLMs can be complex and requires significant expertise in machine learning. Traditionally, scientists have had to rely on external engineering teams to train or fine-tune these models. To address this, Deep Forest Sciences’ has provided a no-code fine-tuning workflow on our Prithvi platform to enable biologists and chemists to train custom scientific foundation models for their needs.

Prithvi™ is our AI-powered platform to accelerate small molecule drug discovery with scientific foundation models. Prithvi™ has the ability to

To learn more about Prithvi™, check out our earlier [blog post](https://deepforestsci.com/blog/0). In this blog post, we will focus on using Prithvi™ to fine-tune scientific foundation models.

Prithvi offers a no-code solution for fine-tuning scientific foundation models, making advanced machine learning capabilities accessible to biologists and chemists without requiring extensive technical skills. Users can select a suitable pre-trained model from the model library and input their datasets for fine-tuning. In the rest of this tutorial, we will focus on fine-tuning a ChemBERTA-2 model with Prithvi.

We start by uploading the dataset to Prithvi and then perform the following analysis steps:

A schematic representation of no-code dataset splitting

The Train primitive in Prithvi can be used to fine-tune a foundation model. Prithvi uses a default Byte-Pair Encoding (BPE) tokenizer, which breaks down large vocabularies in natural language into smaller subwords. The ChemBERTa-2 pre-training process follows the RoBERTa method, which masks 15% of tokens in each SMILES string and sets a maximum sequence length of 256 characters. This helps the model learn to predict masked tokens consisting of atoms and functional groups. The model learns the relevant molecular context for transferable tasks, such as property prediction.

The model leverages HuggingFace libraries under the hood for streamlined pretraining and fine-tuning and is evaluated on the MoleculeNet benchmark. ChemBERTa scales well with pre-training dataset size, offering competitive downstream performance on MoleculeNet and useful attention-based visualization modalities [2].

Prithvi uses a bidirectional training context with ChemBERTa-2, which has 12 attention heads and 6 layers, resulting in 72 distinct attention mechanisms. This configuration helps the model understand chemical syntax and molecular structures.

A schematic representation of no-code fine-tuning using the train primitive

Prithvi’s Evaluate primitive is used to assess the performance of trained models using metrics such as the Root Mean Square Error (RMSE), which measures the differences between predicted and observed values. This evaluation can be conducted on the train, validation, and test splits as desired.

In the ChemBERTa-2 fine-tuning, we evaluate the models using the RMSE metric on the training, validation, and test splits. Multiple workflow jobs are submitted, one for each model, to ensure thorough evaluation. The evaluated results are then stored in JSON files, providing a comprehensive overview of the model's performance.

A schematic representation of no-code evaluation

Featurization transforms raw data into a format suitable for machine learning. This typically involves converting data into SMILES format. For specific models like ChemBERTa-2, which directly use SMILES strings as input, traditional featurization is bypassed. We use the [Dummy featurizer](https://deepchem.readthedocs.io/en/latest/api_reference/featurizers.html#dummyfeaturizer) to featurize data for the ChemBERTa-2 model, which allows the raw dataset to be used directly without any transformation [4]. This approach preserves the detailed structural information of the molecules and simplifies the workflow to ensure that the ChemBERTa-2 model receives the most informative data for fine-tuning.

A schematic representation of the no-code featurization for the ChemBERTa-2 model

Splitting the Dataset using Scaffold Splits

Prithvi’s Split primitive is used to split datasets into train, valid, and test splits. The primitive supports different splitting algorithms like random splitting and scaffold splitting. In this workflow we use scaffold splitting to split the featurized data into train and test sets such that the molecular scaffolds in each group are as distant from the other as possible. The result is that every single molecule in the test set has a scaffold unique from the train set, and is molecularly distant from the train set. By doing this, we ensure that the test set is maximally distant from the training set, which helps to evaluate the model's generalizability and performance on new molecules proposed in a drug discovery campaign.

The default split ratio is 80:10:10. We get train, valid, and test dataset splits as output from this workflow.

Case Study: Fine-Tuning ChemBERTa-2 on a MoleculeNet Dataset

As a case study, we use Prithvi™ to fine-tune a ChemBERTa-2 model that was pre-trained on a 10M [ZINC dataset](https://paperswithcode.com/dataset/zinc), using the Delaney dataset. The Delaney dataset, also known as ESOL (Estimated SOLubility), is a regression dataset included in the MoleculeNet benchmark. The dataset contains structures and aqueous solubility data for 1128 compounds, making it a useful benchmark for learning to predict aqueous solubility directly from a compound's structure.

After fine-tuning, we compared the results with established experimental data to evaluate the model’s performance. The [MoleculeNet paper](https://arxiv.org/abs/1703.00564) reported an RMSE score of 1.7406 ± 0.0261 (mol/L) for Random Forest (RF) Regression models and 0.8851 ± 0.0292 (mol/L) for Graph Convolutional Networks (GCN) on the Delaney dataset, using an 80:10:10 scaffold split [5]. These benchmarks were obtained through grid search-based hyperparameter tuning, a process that can be easily replicated through Prithvi.

We began by featurizing the dataset on Prithvi™ and then performed scaffold splitting through Prithvi's no-code UI.

The ChemBERTa model supports both masked language modeling (MLM) and multi-task regression for fine-tuning. In this case, we opted for a model pre-trained using MLM on the 10M ZINC dataset. The model employed a tokenizer to convert SMILES strings into input tokens, using a byte-pair encoding tokenizer trained on the PubChem10M dataset, sourced from the HuggingFace model [hub](https://huggingface.co/seyonec/PubChem10M_SMILES_BPE_60k).

We evaluated the fine-tuned model on train, validation and test splits using the RMSE metric. In the figure below, we compare these RMSE scores with those of the Random Forest model and the GCN model reported in the MoleculeNet paper.

Results are comparable to those achieved with graph convolutional networks even with a small foundation model pre-trained on only 10M compound structures. The Deep Forest Sciences team is rolling out new versions of ChemBERTa, pre-trained on much larger datasets, which will enable our partners to achieve state of the art results without having to write any code.

Deep Forest Sciences' Prithvi™  platform offers a no-code solution for fine-tuning scientific foundation models like ChemBERTa for different downstream molecular property prediction tasks. It provides a user-friendly interface and pre-trained models to enable researchers to train their own custom scientific foundation model and simplifies the complex process of model fine-tuning, making these methods accessible to biologists and chemists without extensive machine learning expertise. Prithvi™ will help scientists design and discover powerful new drugs by making cutting edge AI tools accessible and friendly.

[1] Bommasani, R., et al. (2021). "On the opportunities and risks of foundation models." Center for Research on Foundation Models (CRFM), Stanford University. https://crfm.stanford.edu/assets/report.pdf

[2] Chithrananda, S., Grand, G., & Ramsundar, B. (2020). "ChemBERTa:Large-Scale Self-Supervised Pretraining for Molecular Property Prediction." https://arxiv.org/abs/2010.09885

[3] Chithrananda, S., Grand, G., & Ramsundar, B. (2022). "ChemBERTa-2:Towards chemical foundation models." https://arxiv.org/abs/2209.01712

[4] "Featurizers". DeepChem Documentation. https://deepchem.readthedocs.io/en/latest/api_reference/featurizers.html#dummyfeaturizer

[5]  Wu, Z., Ramsundar, B., Feinberg, E. N., et al. (2018). "MoleculeNet:A benchmark for molecular machine learning." arXiv. https://arxiv.org/abs/1703.00564

Deep Forest Sciences’ no-code AI Prithvi™ toolchain accelerates small-molecule drug discovery efforts. Deep Forest Sciences also leads the development of open-source DeepChem framework, and emphasizes supporting open source and open science as fundamental parts of our mission and values. Partner with us to apply our foundational no-code AI technology to hard real-world problems in small molecule drug discovery.

Email us at partnerships@deepforestsci.com to learn more!

No-Code Fine-tuning of Chemical Foundation Models with Prithvi

For many years, accurate in-silico prediction of protein-ligand binding affinities has been a fundamental goal in computational drug design, due to its potential to accelerate the discovery process. However, traditional computational methods have often failed to provide accurate results. Recent advances have improved the state of the art in [Relative Binding Free Energy (RBFE)](https://pubs.acs.org/doi/10.1021/acs.jcim.7b00564) calculations, which use physics-based molecular simulations and statistical mechanics, to the point where accurate protein-ligand binding affinity prediction is now within sight.

Taking a step back, RBFE calculations are computational methods used in the field of computational chemistry and drug discovery to estimate the difference in binding free energy between two or more ligands to a common target (e.g., a protein or enzyme). Binding free energy is a key factor in understanding and predicting the affinity of a ligand for its target, which is crucial in drug design.

The binding free energy is a thermodynamic parameter that represents the balance between the energy gained from the favorable interactions between the ligand and the target (e.g., van der Waals forces, hydrogen bonding) and the energy cost associated with desolvation and conformational changes upon binding. RBFE calculations involve comparing the binding free energies of different ligands and determining the differences between them. The goal is often to rank different ligands based on their binding affinities or to understand the factors contributing to the binding affinity differences.

Examining the thermodynamic cycle at hand, our focus will be on finding the change in binding energy between molecules A and B with respect to a given protein target.

Thermodynamic cycle for relative binding free energy (RBFE) calculations.

Two methodologies exist for computing the difference in binding free energy, between molecules A and B. The direct method, depicted by the horizontal lines above, involves a direct transition from the bound state of A to the unbound state of B. This technique exhibits slow convergence attributed to the substantial conformational difference between initial and final states. Another approach is the “alchemical” path, depicted by vertical lines, which incorporates gradual modifications between the states. This method exhibits faster convergence by implementing smaller, gradual transitions. In this strategy, molecule A undergoes a gradual transformation (or perturbation) into molecule B in both the bound and unbound states (vertical lines). The difference in free energy between these transformations equates to that of the direct method, owing to the closed thermodynamic cycle.

No-Code Relative Binding Free Energy Calculations with Prithvi

Relative binding free energy calculations are typically complex to set up and require significant expertise to prepare and analyze results. To address this challenge, Deep Forest Sciences has built no-code infrastructure on the Prithvi™ platform for researchers to run RBFE calculations with ease. Prithivi streamlines the process of setting up, running, and analyzing RBFE calculations and making physical simulations accessible to a broader range of scientists.

To learn more about Prithvi™, check out our earlier [blog post](https://deepforestsci.com/blog/0). In this blog post, we will focus on using Prithvi™ to perform Relative Binding Free Energy calculations for protein-ligand systems.

As a case study, we leverage Prithvi™ for computing binding free energies of five ligands in conjunction with the TYK2 protein. Subsequently, we compare these outcomes with well-established experimental data to evaluate our results. Under the hood, we use [OpenFE](https://openfree.energy/) to perform RBFE simulations.

The dataset needed to run RBFE simulations is obtained from Open Force Fields’ [openforcefield/protein-ligand-benchmark](https://github.com/openforcefield/protein-ligand-benchmark) repository.

Comparison of experimental ∆G values with calculated ∆G values for pmx/CGenFF and Transformato (TF) for the JNK1 system. The computed ∆∆G values for pmx/CGenFF were taken from Gapsys et al. (2020)

A singular RBFE calculation involves the protein and two ligands: the initial ligand of the transformation and the concluding ligand. These transformations, or perturbations, are represented as edges in a perturbation network— a directed graph where each edge signifies a transformation requiring RBFE computation. Various perturbation networks exist, including minimum spanning, radial, and maximal. Currently, Prithvi supports radial perturbation networks. A radial perturbation network resembles a star graph, where one node has a degree N, while all other nodes have a degree of 1. Prithvi supports maximal and minimal spanning networks as well.

In this simulation, we created a Radial Perturbation Network, positioning ejm_48 as the central node. The network comprised a total of 4 edges.

A radial perturbation network with ejm_48 at the center

We selected TYK2 protein due to its practical advantages for computational studies; notably TYK2 demonstrates easy convergence. Additionally, experimental data is available to validate our results.

For the purpose of this experiment, we focus on 5 Ligands: ejm_48, ejm_31, ejm_46, jmc_28, and jmc_23 (displayed below). To mitigate the cost associated with RBFE calculations, we've constrained the ligand count to 5, consequently limiting the number of “edges” in our calculation to 4 (see the next section for details.)

Prithvi™  aligns ligands and proteins, with ligand coordinates prepared in reference to the protein. Both protein and ligands are protonated to a single tautomeric state, following established practice in the field.

This [datasheet](https://www.google.com/url?q=https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9485484/bin/DataSheet1.pdf&sa=D&source=docs&ust=1721049396624568&usg=AOvVaw1wQ1pYjTUmc_SEOcaEZ1TR) provides experimental data for assessing benchmark results, with Table 1.1 presenting pertinent values specifically for the TYK2 System.

Employing the default settings extracted directly from openFE's [RelativeHybridTopologyProtocol](https://docs.openfree.energy/en/latest/guide/protocols/relativehybridtopology.html), we made a few adjustments. To reduce simulation times, we reduced the number of replicas from 11 to 5. Additionally, we set the lambda_windows to 5 to match the n_replicas.

Each edge in the process took approximately a day for completion, and we executed these computations on an AWS GPU node featuring the following specifications:

We observe that the computed binding free energy values for ejm_31 and jmc_28 closely align with the actual experimental values. In contrast, the binding free energies values for ejm_46 and jmc_23 exhibit some deviation from the experimental results, possibly due to lack of convergence. The cost of running these simulations amounted to $50, which is $12.5 per edge. These calculations were done on a dedicated GPU Node. Costs would be lower (at the cost of additional error handling) when we use SPOT compute instances.

Additionally, we conducted a singular edge simulation, maintaining default values for both n_replicas and lambda_windows, each set to 11. This simulation extended over a period of approximately three days, yielding the subsequent results.

The computed binding free energy value for ejm_47 closely aligns with the reported  experimental value. We believe employing default settings for the mentioned edges could have led to improved results for all other ligands. Nevertheless, considering the substantial computational expenses associated with these simulations, we opted for a reduced number of replicas to mitigate costs.

A schematic of performing no-code RBFE calculations on Prithvi

Prithvi's no-code platform enables scientists and researchers to perform RBFE calculation and analyze results without having to write any custom scripts. Our case study demonstrates how Prithvi™ can accurately compute relative binding free energies (RBFE) for protein-ligand systems that align well with experimental values for certain ligands. The computations, costing only $50 for five ligands, highlight Prithvi's cost-effectiveness. Prithvi™ is highly scalable and can be used to perform RBFE cost-effectively at large scales for your discovery needs.

No-Code Relative Binding Free Energy (RBFE) Calculations with Prithvi

[Molecular Dynamics (MD)](https://en.wikipedia.org/wiki/Molecular_dynamics) is a computational method used to study the physical movements of atoms and molecules, such as proteins and nucleic acids, at the atomic level. The technique simulates the movement of atoms over time based on the interactions and forces between them, following Newton's laws of motion.

Molecular systems are complex and can consist of vast number of atoms and molecules. Analytically predicting the properties and behaviors of these systems under various conditions is extremely challenging. MD simulations avoid analytical solutions by numerically integrating Newton's equations of motion to model the trajectories of individual atoms. At each step, the method calculates the forces exerted on each atom by all other atoms in the system and iteratively updates their positions and velocities over time. Molecular dynamics is considerably more efficient, and can handle larger systems and longer simulation times, than quantum mechanical simulation methods. At the same time, it is far slower than coarse-grained models of pathways, cells or tissues from systems biology.

MD simulations use [force fields](https://en.wikipedia.org/wiki/Force_field_(chemistry)), mathematical models of the potential energy surfaces of molecules. Forces are computed by computing the negative gradient of the potential energy. Force fields include parameters for bond lengths, angles, and non-bonded interactions such as van der Waals forces and electrostatic interactions. The simulation process involves dividing time into small steps (typically at the femtosecond scale), during which the forces on each atom are computed and their positions and velocities are updated accordingly. This iterative approach allows researchers to capture the time-dependent evolution of molecular systems.

A simplified schematic of the molecular dynamics algorithm

Molecular dynamics simulations have recently found greater use in drug discovery. For example, a recent [article](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6209097/) demonstrates how MD simulations provide a powerful complement to experimental methods. MD simulations can help identify the binding modes of potential drugs with their target proteins. Additionally, MD simulations provide information on how proteins change shape and interact with other molecules in their environment. This can be used to better understand the function of proteins and design drugs that modulate their activity. MD simulations can aid in virtual screening by helping to identify alternative binding poses. Experimental techniques like NMR spectroscopy and fluorescence quenching which validate the predictions of MD simulation remain critical, but simulations are slowly becoming useful predictive tools[1].

Advances in computational power and the availability of open source software are powerful trends driving adoption of molecular dynamics. However a number of practical challenges remain to broad adoption. Several research efforts have integrated machine learning with MD simulations to improve accuracy of simulations and enable the prediction of unsampled molecular configurations [2]. In addition, MD simulations remain computationally expensive and generate vast amounts of data, making it challenging to analyze and identify key molecular behaviors.

Effectively using molecular dynamics simulations involves complex and time-consuming calculations and often requires advanced computational skills. Prithvi™ provides a no-code platform for biologists and chemists to run molecular dynamics workflows to simulate and analyze molecular interactions without needing such expertise. Prithvi™ simplifies the entire process of MD simulations and allows the users to upload molecular data, select suitable force fields, control temperature and pressure, and configure other parameters as they require.

The standard workflow for molecular dynamics simulations

To learn more about Prithvi™, check out our earlier [blog post](https://deepforestsci.com/blog/0). In this blog post, we will focus on using Prithvi™ to run protein-based MD simulations.

Prithvi™ offers a no-code platform for running Molecular Dynamics (MD) simulations, making complex molecular analyses accessible without extensive coding. Prithvi’s molecular dynamics workflow allows scientists to specify and run MD simulations directly through its web UI without having to write custom scripts or set up high performance computing (HPC) environments. Prithvi’s MD workflow involves the following steps

Simulation setup allows for several additional parameters, such as step size, simulation duration, and reporting intervals, to be set. For example, a typical simulation might run for 0.5 nanoseconds with a step size of 0.005 picoseconds and log results at regular intervals. This detailed control over simulation parameters allows users to tailor their simulations to specific research needs.

Table 1: Input parameter for running a MD simulation.

Prithvi allows users to use GPU nodes for simulations and provides for easy use of both dedicated and spot instances.

Once the simulation is complete, Prithvi™ generates trajectory files that store the time-dependent coordinates and velocities of the atoms in DCD format. These files provide a comprehensive record of the simulated molecular dynamics, allowing researchers to visualize and analyze the behavior of the molecules over time. Additionally, Prithvi™ stores produced log files detailing the various parameters and conditions throughout the simulation.

A schematic representation of Prithvi’s no-code MD simulation workflow.

MD simulations generate extensive trajectories of atomic positions over time, showing the various conformational states. These resulting trajectories contain large amounts of data. To make sense of this data, clustering techniques are employed to group similar conformations together which allows researchers to identify and analyze significant structural states within the molecular system.

Prithvi’s no-code MD workflow implements several clustering algorithms, like [affinity propagation](https://en.wikipedia.org/wiki/Affinity_propagation), [DBSCAN](https://en.wikipedia.org/wiki/DBSCAN) and [k-means](https://en.wikipedia.org/wiki/K-means_clustering),  to efficiently analyze MD trajectory data generated from MD simulations to group similar molecular structures and form clusters of conformations.

To configure the affinity propagation algorithm, users set parameters such as damping, maximum iterations, convergence iterations, and preference.

Table 2: Input parameters for MD clustering

Prithvi™ runs the selected clustering algorithm and generates several PDB files, each representing a distinct cluster of similar molecular structures identified in the trajectory file. These clusters provide an easy method to identify distinct conformational states, possibly corresponding to different biological function, that the molecule can adopt.

A schematic representation of Prithvi’s no-code MD clustering workflow.

Prithvi provides a molecular docking framework for the users to predict protein-ligand interactions and design molecules that bind to the structures generated by the MD workflow This is done using no-code tools for pocket finding, pose generation and scoring. For detailed information on Prithvi’s no-code molecular docking capabilities, see our earlier [blog post](https://deepforestsci.com/blog/4).

To start an MD simulation on Prithvi™, users upload a Protein Data Bank (PDB) file containing the three-dimensional structure of the molecule(s) to simulate. This file defines the initial positions of all atoms in the biomolecular system. PDB files may contain errors that can cause the MD simulations to become unstable and give unrealistic results. Cleaning involves tasks like removing water molecules, adding missing atoms, fixing structural issues, etc. to prepare the structure for MD simulation.

Under the hood, Prithvi™ makes use of [OpenMM](https://ommprotocol.readthedocs.io/en/latest/forcefields.html). Force fields such as such as, [amber14-all](https://github.com/openmm/openmm/blob/master/wrappers/python/openmm/app/data/amber14-all.xml), [amber14/tip3pfb](https://github.com/openmm/openmm/blob/master/wrappers/python/openmm/app/data/amber14/tip3pfb.xml) and [charmm36/tip5pew](https://github.com/openmm/openmm/blob/master/wrappers/python/openmm/app/data/charmm36/tip5pew.xml), provide mathematical models that describe potential energy surfaces and the forces acting on atoms within molecular systems. Prithvi™ supports several force fields, which provide different methods of modeling interactions between atoms.

Prithvi™ uses the [LangevinMiddleIntegrator](http://docs.openmm.org/latest/api-python/generated/openmm.openmm.LangevinMiddleIntegrator.html#openmm.openmm.LangevinMiddleIntegrator.setConstraintTolerance) by default but also supports other integrators. Prithvi™ enables users to control the temperature and pressure of the system to simulate different environmental conditions, such as a temperature of 300 K (27℃) to mimic biological environments. This control allows researchers to study how molecular interactions change under various conditions.

Molecular Dynamics has been used widely to study the structure and dynamics of macromolecules like proteins. MD simulations can be particularly valuable in lead optimization where they can help refine ligand interactions, predict binding pocket rearrangements, and improve ligand binding poses. Unlike traditional docking methods, MD simulations offer the potential for more accurate predictions of ligand binding affinities [1]. However, this process is computationally intensive and requires sophisticated knowledge. Prithvi™ provides a no-code molecular dynamics workflow for the researchers to easily run MD simulations and clustering for understanding and analyzing molecular behavior and interactions, without requiring advanced computational expertise.

MD simulations facilitate virtual screening by enabling systematic study of the flexibility of binding pockets, and can help identify a broader range of potential binders compared to docking with a single protein structure. MD simulations also help understand conformational changes that influence specific signaling patterns, needed for designing drugs with precise therapeutic effects [1]. Using no-code primitives, Prithvi™ streamlines the entire process of running molecular dynamics simulation by allowing users to upload molecular data, select suitable force fields, and control environmental conditions as they require. In future blog posts, we will discuss more advanced applications of this technology, such as showing how Prithvi™ can be used to facilitate the design of allosteric inhibitors.

[1] Hollingsworth, S. A., & Dror, R. O. (2018). Molecular dynamics simulation for all. Neuron, 99(6), 1129–1143. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6209097/

[2] Wang, Y., Ribeiro, J. M. L., & Tiwary, P. (2019). Machine learning approaches for analyzing and enhancing molecular dynamics simulations. arXiv preprint arXiv:1909.11748. https://arxiv.org/abs/1909.11748.

Email us at [partnerships@deepforestsci.com](partnerships@deepforestsci.com) to learn more!

No-Code Molecular Dynamics with Prithvi

In drug discovery, exploring vast chemical spaces to find potential drug candidates is a major challenge. While traditional techniques like high throughput screening (HTS) have successfully generated numerous leads for development, such methods have major limitations. HTS setups remain costly and are limited to screening about a million compounds.

[DNA-encoded libraries (DEL)](https://en.wikipedia.org/wiki/DNA-encoded_chemical_library) are a recent development that allow for synthesis and screening of chemical libraries containing billions of small molecule compounds. DELs use water-based combinatorial chemistry and a split-and-pool method to generate diverse compounds. These compounds are screened against target proteins via affinity selection in a single-tube approach, unlike high throughput screening, which requires each compound to be screened individually.

A schematic of the DNA-Encoded Library (DEL) procedure, showcasing the creation, reaction, sequencing, selection, and enrichment of DNA templates

In DELs, oligonucleotides serve as a unique identifier for each compound in the library. These oligonucleotides are short DNA sequences that are chemically linked to small molecule compounds. During the synthesis process, each compound is tagged with a specific oligonucleotide sequence which encodes information about the chemical building blocks from which the compounds were synthesized. When compounds bind to a target protein, the attached oligonucleotides can be rapidly sequenced. This allows researchers to identify successful interactions without the need for separate testing of each compound.

DNA encoded libraries have considerable synergy with machine learning methods, since they naturally generate very large scale datasets. A seminal [article](https://arxiv.org/abs/2002.02530) from Google in 2020 created a lot of interest in the field of DEL-ML by demonstrating how DEL data could be used to train large machine learning models for hit finding. Trained DEL-ML models could be used to rapidly screen large compound libraries to predict activity. Subsequent work has shown that machine learning models like Random Forests (RF) and Graph Convolutional Neural Networks (GCNN) trained on DEL data are effective for predicting the activity of new compounds selected from vendor libraries. The systematic use of DEL-ML holds the potential to significantly reduce the time and cost needed to bring new molecules to market, and the approach has consequently generated significant interest in the industry.

However, computational processing of DEL data remains challenging. There is considerable assay noise in DEL data, which has led to the use of denoising strategies such as “disynthon aggregation,” which looks at subfragments of DEL compounds to lower noise. More recent [research](https://pubs.acs.org/doi/10.1021/acsomega.3c01775) has started to explore the use of directly learning on full DEL compounds as well.

Conceptual framework for training ML models using DEL data

Although DEL-ML has found considerable adoption, deploying machine learning solutions for DEL data has remained a difficult technical endeavor, requiring considerable software engineering expertise. To address this need, Deep Forest Sciences has released infrastructure on our Prithvi™  platform to enable biologists and chemists to build no-code DEL-ML models. We anticipate that our technology will enable biologists and chemists to use DEL data systematically in their discovery pipelines.

An expanded Prithvi DEL pipeline showing the conceptual steps in running a no-code DEL-ML workflow.

This workflow consists of a series of no-code primitives which can be invoked by users directly from Prithvi’s web user interface (UI). Advanced users can save workflows as reproducible scripts which can be invoked from the UI or alternatively from a Python API. The workflows have been tested on live DEL projects at large scale with existing partners.

Let’s discuss each workflow in the Prithvi DEL-ML pipeline.

The Visualize primitive in the Prithvi DEL pipeline renders DEL data in a 3D “feature cube.”. This tool provides users a comprehensive view of DEL data, and helps identify “line” and “plane” features (one or two conjoined DEL building block fragments which are commonly represented in hits). Each axis represents a building block and color-coded counts help identify hits.

A schematic representation of no-code visualization

An example of a DEL-cube (on synthetic data) visualized by Prithvi™ DelCubeViewer

Input DEL data is noisy due to noise in sequencing. A common denoising strategy is to use aggregated sequencing counts based on shared “disynthon” representations. For instance, in a three-cycle library of the form A−B−C, sums aggregating counts for the A−B, A−C, and B−C disynthons can be generated to lower noise.

Suitable cutoffs can be set for aggregated reads (sequence counts) to ensure that small sequence counts don't influence the balance of the training data. The following [formula](https://pubs.acs.org/doi/10.1021/acsomega.3c01775) has been proposed in the literature as a strategy for converting sequence counts into enrichment scores,

where p~0~ is the observed probability (p~0~ = c~0~/n), p~i~ is the expected probability (p~i~ = 1/ library size), c~0~ is the observed count, and n is the total sequencing reads.

This normalized z-score calculation approximates the DNA sequencing process as random sampling with replacement using a binomial distribution. It quantifies the enrichment level of an observed count compared to the expected count, factoring in the sequencing read depth and library sizes. The average of these scores sets a threshold that categorizes each disynthon pair as a "hit" or "not hit."

Example of sample DEL data: R1_REAGENT, R2_REAGENT, and R3_REAGENT are the columns from which disynthons are generated, and the values from the SequenceCount column are used to calculate the enrichment scores.

Prithvi’s Deldenoise primitive implements denoising strategies following these and other best practices from the literature to enable users to easily convert raw assay data into a format suitable for machine learning.

Prithvi is our AI-powered platform to accelerate small molecule drug discovery with scientific foundation models. Prithvi has the ability to

A schematic representation of the no-code DELdenoise primitive

The Featurize primitive in Prithvi is used to featurize the denoised DEL data. (Here “featurization” denotes the process of converting inputs into a form suitable for machine learning.) This step involves transforming denoised outputs, such as disynthons, which can be represented as pairs of SMILES, into a disjoint union of graph structures (for GCNNs) or fingerprint vectors (for RFs).

A schematic representation of the no-code featurization

The featurized data is split into training, validation, and test splits using the Split primitive. In [scaffold splitting](https://www.oloren.ai/blog/scaff-split), the data is split into training and test sets such that the [molecular scaffolds](https://pubs.acs.org/doi/10.1021/jm9602928) in each group are as distant from the other as possible. The result is that every single molecule in the test set has a scaffold unique from the training set, and is structurally dissimilar from the training set. Ensuring the test set is maximally distant from the training set helps test models for generalizability and provides a better measure of their performance on new molecules.

A schematic representation of the no-code dataset splitting

The Hyperparameter Tuning primitive in Prithvi uses the training and validation splits of the featurized DEL data to train several representative machine learning models, like GCNs and RFs, using different combinations of hyperparameters from the provided hyperparameter space. Prithvi employs a hyperparameter search technique, which systematically searches through a specified hyperparameter space and evaluates the performance of the model for each combination of hyperparameters.

Hyperparameter tuning is crucial in the context of DEL data due to the large chemical diversity and inherent noise in the library. The output of this workflow provides the best model, the best hyperparameters, and a collection of scores for all evaluated models for hit predictions.

A schematic representation of no-code hyperparameter tuning

We use the Train primitive to train an ensemble of GCN and RF models to enhance the accuracy of hit predictions from DEL data. Prithvi supports no-code construction of ensembles, which are collections of multiple copies of a base model each trained on the same training data but with different random seeds. The predictions from these individual models are then averaged to generate a final robust result.

Because ensemble models combine predictions from multiple models, they improve accuracy by reducing the impact of individual model biases. This approach leads to better hit identification by minimizing false positives and offering consistent predictions across diverse DEL datasets.

A schematic representation of no-code training

To learn more about Prithvi, check out our earlier [blog post](https://deepforest.substack.com/p/accelerating-small-molecule-discovery). In this blog post, we will discuss the DEL-ML capabilities of Prithvi.

A schematic representation of no-code ensemble model prediction by Prithvi™

Prithvi’s Evaluate primitive is used to evaluate the performance of trained models using metrics such as [ROC-AUC](https://en.wikipedia.org/wiki/Receiver_operating_characteristic), which measures the area under the Receiver Operating Characteristic curve. This evaluation can be run on the train, valid, and test splits as desired. Prithvi computes these statistics for each model and saves the results in a CSV file.

The Infer primitive is used to generate hit predictions for SMILES on specified vendor catalogs using trained ensembles of models.

A schematic representation of no-code inference

Threshold Hits and Cluster for Chemical Diversity

Users can use the Threshold primitives to select the top 1% (or any other desired percentage) of predictions as hits. Once thresholds are identified from the inference results, users can run the Cluster primitive to cluster these hits and select representative top hits from cluster centers.

A schematic representation for running no-code threshold analysis and clustering.

Prithvi’s DEL pipeline offers no-code machine learning tools designed to model, analyze, and virtually screen small molecule compounds. Prithvi streamlines the entire DEL-ML workflow, from visualizing and denoising DEL data to training models and identifying hits. Prithvi’s no-code infrastructure allows scientists to focus more on discovering new hits and less on the technical complexities of machine learning pipelines.

To perform Prithvi DEL-ML analysis, users can run the following no-code workflow:

DNA-encoded library (DEL) technology can significantly accelerate the discovery process by enabling systematic hit discovery. Prithvi™ provides a powerful no-code tool for DEL-ML analysis, enabling biologists and chemists to process DEL data without needing sophisticated Python pipelines. Prithvi streamlines the entire process of modeling, analysis, and screening of small molecule compounds into a series of no-code primitives, which can be modularly composed and modified as users require. Each stage of Prithvi’s DEL pipeline, from data visualization to advanced hit prediction, is designed to optimize the discovery and evaluation of potential drug candidates, allowing scientists to focus more on discovery and less on technical complexities.

No-Code DEL-ML (DNA-Encoded Library Machine Learning) Analysis with Prithvi 

A Brief Introduction to Molecular Docking

Molecular docking is a computational method that predicts how small molecules, with atomic weight less than ~500 dalton, bind to larger biomolecular targets like proteins. Docking roughly simulates interactions between molecules to determine favorable binding positions for [ligands](https://en.wikipedia.org/wiki/Ligand_(biochemistry)) to proteins and is utilized in drug discovery for screening and understanding molecular interactions.

Prithvi provides a powerful no-code framework to execute complex drug discovery workflows. This blog post showcases Prithvi’s no-code molecular docking workflows and capabilities. Biologists and chemists can leverage Prithvi’s powerful UI to run complex analyses without having to write custom scripts. Prithvi also supports very-large scale docking workflows. In a future blog post, we will discuss how to run very-large scale docking using Prithvi.

Deep Forest Sciences’ AI-powered Prithvi toolchain accelerates small-molecule drug discovery efforts. Deep Forest Sciences also leads the development of open-source DeepChem framework, and emphasizes supporting open source and open science as fundamental parts of our mission and values. Partner with us to apply our foundational AI technologies to hard real-world problems in small molecule drug discovery.

Email us at partnerships@deepforestsci.com

To learn more about Prithvi, check out our earlier [blog post](https://deepforest.substack.com/p/accelerating-small-molecule-discovery). In this post, we will focus on using Prithvi for molecular docking analyses, but in future posts we will highlight Prithvi’s capabilities in free energy calculations, molecular dynamics, and more.

Proteins or enzymes related to biological processes or disease pathways are the central targets of modern drug discovery. Drug discoverers seek to design compounds, or ligands, that interact with these targets in various ways. A binding site (or pocket) is a region of the target where a drug-like molecule can effectively interact with it. The drug-target pair will tend to occupy low energy states since those are more physically favorable. A particular choice of drug-target interaction shape is called a binding pose. The shape of the molecule in this pose is called a conformation. Molecular docking aims to predict low-energy ligand conformations associated with high-affinity binding sites of targets. Additional biomolecules, like DNA, can also be involved when drugging systems like transcription factors.

Several families of classical algorithms predict low-energy binding affinity using score functions based on physical or chemical equations [1]. Ligand chirality, Van der Waals forces, electrostatics, and hydrophobicity are important in calculating binding affinity and are often explicitly included in physical scoring functions. However, such methods often have high runtime and associated computational costs. Recent advances in deep learning have enabled algorithms to learn from molecular data and enhance pose search space while reducing the parameters mentioned earlier, but possibly at the cost of failing to generalize to proteins not present in the training dataset.

This blog post highlights Prithvi’s molecular docking features to explore protein-ligand interactions. Prithvi provides a protein-handling infrastructure suite to run docking jobs. This tool broadens a company’s molecular search space, and helps identify potential drug-like molecules. Below is a general workflow to perform a docking simulation on Prithvi:

Prithvi uses open-source tools like DeepChem and AutoDock Vina to perform docking. Open-source tools enable us to apply state-of-the-art scientific tooling maintained by a broader community. Prithvi’s no-code infrastructure makes advanced computational methods easily accessible to biologists and chemists. In this blog post, we perform a case study to showcase Prithvi’s no-code molecular docking capabilities by docking five organic compounds (Naringin, Capsaicin, Gallic Acid, Psychotrine, and Quercetin) against the SARS-CoV-2 main protease (PDB ID:5R80) [2].

Example of Docking with Prithvi. From right to left, ligand conformers are sampled, and possible binding pockets are shown in red. Then, protein-ligand candidates are generated in PDB format. At last, binding affinity scores for each pose are saved in a CSV file.

Each role outlined in Prithvi’s docking workflow is described below.

Molecular docking workflow with Prithvi. Each block in the figure represents a primitive, which is a setup to specify parameters to run scientific computational tasks.

Conformational Sampling: Sample Low-Energy 3D Molecular Shapes

Docking aims to select a low-energy binding pose among different binding poses. A molecular conformation is one of the plausible spatial arrangements of atoms in a molecule. Conformer sampling is based on axis rotations and atomic bond restrictions [6]. Therefore, the ligand’s structure limits the number of conformers. Larger molecules generate more conformers due to increased degrees of freedom (e.g., rotation modes). Ligands are normally input into Prithvi as SDF, PDB, or SMILE files. If desired molecules aren’t known upfront, scientists can use Prithvi’s built-in editor tools to draw the molecular structure and download it as the previous file types. Pairing docking algorithms with good conformational sampling explores a broader ligand space, and increases chances of identifying native-like binding poses increase.

Prithvi’s conformational sampling primitive asks for an input file, containing a ligand, the number of conformers to generate and the output filename to store the sampled conformation. Prithvi currently computes the results by using RDKit to generate conformers, but will soon support more powerful sampling methods.

A schematic representation of the no-code conformational sampling.

Two different conformers of Capsaicin after applying the conformational sampling primitive.

Finding Binding Pockets: Mapping the Topography of Molecular Interaction Sites

Protein-ligand interactions take place at preferred sites on the target protein commonly referred to as binding pockets. Docking can be performed with or without knowledge of the binding pocket. Docking performed without knowledge of a binding pocket is referred to as blind docking, in contrast to docking performed with a specified binding pocket which is non-blind. Non-blind docking is usually more efficient but in some cases a good binding pocket is not known.

In blind docking, the whole protein surface must be searched for possible binding pockets. In contrast, with non-blind docking users can specify bounding boxes as parameters, strategically defining spaces where a ligand is most likely to bind. Users can visualize these bounding boxes directly through Prithvi.

Several methods have been proposed to detect binding pockets including geometry-based, energy-based, sequence-based, hybrid, and deep learning algorithms [3]. Different algorithms can be more or less effective for different systems.

Prithvi supports a geometric-based approach based on DeepChem’s docking utilities. This method computes the convex hull of the protein and uses it to identify faces that may be suitable for docking. More advanced strategies that leverage Voronoi tessellations and deep learning-based surface scanning methods are under development and will be added to Prithvi shortly.

A schematic representation of no-code pocket finding.

Three bounding boxes for SARS-CoV-2 main protease (PDB ID: 5R80) identified using the pocket finding primitive.

Pose Generation: The Molecular Arrangement

After exploring the conformer space and binding sites, the next step is coupling the ligand to the receptor. Molecular docking aims to predict protein-ligand binding mode candidates, known as poses [1]. Based on user-defined parameters, a ligand is assessed at different potential binding pockets on the protein surface, and favorable poses are sampled (or generated in the case of deep methods). The poses with lowest binding energy have greater energetic stability.

The pose generation primitive in Prithvi currently uses AutoDock Vina and Gnina as its docking engines [4]. Parameters must be specified for the ligand, the protein to dock against, exhaustiveness, and the number of poses to generate. Centroid and box dimension parameters can optionally be provided to perform non-blind docking. We are also working on incorporating new deep-learning and equivariant pose generation methods into our pipeline.

A schematic representation of no-code pose generation.

Three generated poses of capsaicin, gallic acid, and naringin binding with SARS-CoV-2 main protease (PDB ID: 5R80).

Pose Scoring: Quantitative Assessment of Molecular Affinity

Pose scoring quantifies energetically favorable binding poses by estimating the strength of the protein-ligand binding interaction. As mentioned previously, this interaction depends on chemical and physical factors such as Van der Waals forces, electrostatic interactions, and hydrogen bonding which contribute to the numerical score, typically calculated in kcal/mol [1]. Docking methods aim to predict pose orientation and stability by comparing binding energies. For Autodock Vina, this optimization task involves genetic algorithms iterating over search space to identify favorable conformers with low binding energy.

Prithvi’s pose scoring primitive needs a docked ligand, a target as inputs, and a filename to store results as a CSV file. Audock Vina, based on simple physical approximation, or Gnina’s score function, based on a convolutional neural network, can be used to compute a binding score [5].

A schematic representation of the no-code pose scoring.

The score for docked naringin with 5R80 using Autodock Vina scoring functions.

A Case Study with Prithvi’s No-Code Docking

We start the case study with a control experiment to evaluate Prithvi’s performance on a known benchmark system. Autodock Vina provides an example that docks imatinib with the kinase domain of c-Abl (PDB:1IEP) [7] ([link](https://autodock-vina.readthedocs.io/en/latest/docking_python.html)). We run the same experiment with Prithvi and display below the scores from Prithvi’s no-code Vina docking and the reference Autodock Vina scores. Files [1iep](https://github.com/ccsb-scripps/AutoDock-Vina/blob/develop/example/basic_docking/data/1iep_receptorH.pdb) and [imatinib](https://github.com/ccsb-scripps/AutoDock-Vina/blob/develop/example/basic_docking/data/1iep_ligand.sdf) were retrieved from Autodock Vina’s repository.

A SARS-CoV-2 main protease receptor co-crystal structure (PDB ID: 5R80).

First, 5 conformers were sampled for each ligand. Then, the pocket-finding primitive was used to identify possible binding sites on the SARS-CoV-2 main protease. Visualizing binding pockets can help support or discard hypotheses about where drug candidates may bind. Below, a binding pocket is presented for the main protease co-crystal structure where molecular fragments are enclosed [2].

Pocket for SARS-CoV-2 co-crystal main protease (5R80).

The pose generation primitive parameters were set to output 2 poses for each ligand conformer. Below are presented protein-ligand complexes of each ligand and SARS-CoV-2 main protease. Poses are likely to bind inside a bounding box colored yellow retrieved from Pocket Finding primitive. This structure is the same as above and serves as a guide to proceed with docking.

Protein-ligand poses using Prithvi’s Pose Generation primitive with a bounding box in yellow. From right to left ligands presented are Gallic Acid, Naringin, Quercetin, Psychotrine, and Capsaicin.

Results from Prithvi’s score primitive are presented below:

Top pose of naringin and 5R80 with -8.46 kcal/mol

Naringin presented the lowest score compared with other candidates. Naringin’s lowest energy binding pose occupies a binding site formed by LEU-141, GLU-166, and GLN-189 residues, matching results from the reference study [2].

3D comparison of Naringenin top scoring pose with SARS-CoV 2 Main Protease (PDB ID: 5R80). First, we present Prithvi’s pose and the second pose of the case study [2].

Generated pose for 1iep-imatinib using Prithvi primitives.

Next, we study some potential inhibitors for the SARS-CoV-2 main protease [2]. Active components from medicinal plants have shown beneficial effects such as antioxidant, antidiabetic, or anti-inflammatory activity with good toxicity profiles. Consequently, following [2], we choose the following compounds for docking against the main protease:

Prithvi offers an easy-to-use UI to visualize conformations and binding poses. The images below were rendered using Prtihvi’s visualization features. Below are five organic compounds and a SARS-CoV-2 main protease co-crystal structure.

The five ligands we investigate in this case study.

No-Code Molecular Docking with Prithvi

Building quantitative structure activity relationship (QSAR) models that predict molecular properties is an important task for a small-molecule biotech/pharma company. Robust models that predict properties such as solubility, lipophilicity, and other ADME/Tox measurements can be valuable tools in the design process for new small molecules. In this blog post, we discuss how the Prithvi platform facilitates building and benchmarking chemoinformatic models. In particular, we build predictive models on some datasets from the [MoleculeNet](https://github.com/deepchem/moleculenet) benchmark collection.

Prithvi is our AI-powered platform to accelerate small molecule drug discovery with scientific foundation models. It has the ability to

To learn more about Prithvi, check out our earlier [blog post](https://deepforest.substack.com/p/accelerating-small-molecule-discovery).  In this post, we will focus on using Prithvi for ligand-based chemical machine learning, but in future posts we will highlight Prithvi’s capabilities in structure-based drug discovery and design as well.

MoleculeNet is a widely used benchmark for molecular machine learning tasks (co-created by our CEO [Dr. Ramsundar](https://rbharath.github.io/about/) during his Ph.D. at Stanford), which focuses on molecular machine learning and is integrated with DeepChem. Predicting the aqueous solubility of a compound or ESOL (Estimated SOLubility) directly from its structure is one task benchmarked in MoleculeNet through the Delaney dataset. Delaney (also referred to as ESOL) is a regression dataset containing structures and aqueous solubility data for 1128 compounds.

The MoleculeNet [paper](https://arxiv.org/abs/1703.00564) reported an RMSE score of 1.7406 ± 0.0261 (mol/L) on the Delaney (ESOL) dataset for Random Forest (RF) Regression models and RMSE score of 0.8851 +- 0.0292 (mol/L) on the Delaney (ESOL) dataset for Graph Convolutional Network (GCN), each on a 80:10:10 scaffold split. This performance is based on the average results from three models trained with hyperparameters obtained through grid search-based hyperparameter tuning. This same analysis workflow can be easily replicated through Prithvi.

We start by uploading the Delaney dataset to Prithvi and then perform the following analysis steps

A representation of paracetamol being converted into a molecular fingerprint.