A Brief Introduction to Molecular Docking

Molecular docking is a computational method that predicts how small molecules, with atomic weight beneath roughly 500 dalton, bind to larger biomolecular targets like proteins. Docking simulates the interaction between molecules to determine favorable binding orientations and is utilized in drug discovery for screening, lead optimization, and understanding molecular interactions.

Prithvi is our AI-powered platform to accelerate small molecule drug discovery with scientific foundation models. Prithvi facilitates large-scale use of chemical foundation models, molecular dynamics, docking, FEP, and more. Prithvi’s protein suite can:-

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 and work through a study case to identify potential drug candidates. In future posts, we will highlight Prithvi’s capabilities in structure-based drug discovery and design.

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

Molecular docking is a core tool for structural bioinformatics that simulates protein-ligand interactions. Proteins or enzymes are macromolecular targets related to biological processes or disease pathways. Additional molecular structures, like DNA, can also be involved in systems like transcription factors. Molecular docking aims to predict low-energy ligand conformations and orientations associated with high-affinity binding sites of targets.

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 fundamental in calculating binding affinity and are often explicitly included in physical scoring functions.  However, such methods require higher runtime and 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.

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 molecular search space, leading to identifying potential drug-like candidates. 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. In this blog post, we perform a case study to showcase Prithvi’s molecular docking capabilities. Five chemical compounds (Naringin, Capsaicin, Gallic Acid, Psychotrine, and Quercetin) are docked against a SARS-CoV-2 main protease (PDB ID:5R80) to find potential COVID-19 treatments [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 applications for drug discovery rely on selecting a low-energy binding pose among different molecular candidates. 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 nature limits the number of conformers. Larger molecules generate more conformers due to degrees of freedom (e.g., rotation modes). When referring to a ligand, it is normally input as a SDF, PDB, or SMILE file. If the desired molecules aren’t known upfront, scientists can use Prithvi’s editor tools to illustrate the molecular structure and download it as the previous file types.

Molecular docking’s effectiveness is evaluated by predicting a ligand's relevant binding poses within the protein's binding site. Docking algorithms with conformational sampling explore a broader ligand space, and chances of identifying native-like binding poses increase.

The conformational sampling primitive asks for an input file, a ligand, the number of conformers to generate and output to store the molecular arrangements as PDB files. Prithvi computes the results by using RDKit to generate conformers.

Two conformers after applying conformational sampling primitive.

Finding Binding Pockets- Mapping the Topography of Molecular Interaction Sites

Protein-ligand interactions take place at preferred target sites known as binding pockets. Storage of these virtual spaces is commonly included in molecular docking workflows. The docking process encompasses two distinct types:-

In blind docking, whole protein serves as the target, and possible binding pockets are calculated. In contrast, with non-blind docking users can specify bounding boxes as parameters, strategically defining a virtual space where a ligand is most likely to bind. Scientists can visualize these 3D arrangements and analyze a binding pose.

Computer-based methods in recent years have enabled several approaches to detect binding pockets, geometric-based, energetic-based, sequence-based, hybrid, and DL algorithms [3]. Computational criteria, including efficiency and precision, are decisive when selecting a suitable pocket-finding method.

Prithvi supports a geometric-based approach under DeepChem docking utilities by using Scipy’s ConvexHull and several strategies including Voronoi Tessellation and deep learning-based methods are under development.

Three binding pockets for SARS-CoV-2 main protease (PDB ID: 5R80) using 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. Common approaches for determining plausible complexes are systematic, stochastic, deterministic, or generative models [1].

Based on user-defined parameters, a ligand is assessed at different sites within the protein, and favorable poses are generated. Significant poses have the lowest binding energy, indicating greater energetic stability. The protein and docked ligand are combined to create a complex in PDB format.

The pose generation primitive in Prithvi uses AutoDock Vina as its primary docking engine [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 are optional whether to perform non-blind docking. We are working on incorporating new deep-learning and equivariant pose generation methods into our pipeline.

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 evaluating the strength of a ligand binding to a target. This depends on chemical and physical factors such as Van der Waals forces, electrostatic interactions, and hydrogen bonding which contributes to the numerical score calculated in kcal/mol [1]. Docking aims to predict pose orientation and stability by using and comparing binding energy. 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. Scoring functions are computed using AutoDock Vina and machine learning methods such as Atomic Convolutions are being trained to include other options [5].

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

Case Study to Showcase Prithvi’s Docking Capabilities

First, we did a control run to evaluate Prithvi’s performance on a known system. Autodock Vina provides an example of molecule inhibitor imatinib with the kinase domain of c-Abl (PDB:1IEP) which transforms cells and causes myelogenous leukemia [7]. For more details on Vina docking scripts please click [here](https://autodock-vina.readthedocs.io/en/latest/docking_python.html). Scores from Prithvi and Vina and the generated pose figure are presented:-

Ligands used for Molecular Docking: Capsaicin, Quercetin, Gallic Acid, Psychotrine, and Naringin—SARS-CoV-2 main protease receptor co-crystal structure (PDB ID: 5R80).

First, 5 conformers were sampled for each ligand. This operation performs rotation depending on molecular chirality. Then, the pocket-finding primitive was used to identify possible binding sites where molecules are likely to dock. This feature gives a clearer idea of generated poses aligned with binding pockets. Visualizing the binding pockets helps to 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).

Protein dimerization is an important biological phenomenon where proteins interact with similar structures (homo and hetero-dimers) to form a functional assembly. This interaction is fundamental and common in biological activities such as enzymatic activation, pathogenic pathways, etc. Also, cysteine protease regions are crucial for viral replication. This characteristic benefits covalent inhibitors, which are molecular structures that bind to a target such as an enzyme or other protein. Covalent inhibitors are attracted to these regions and most of the bounding boxes surround CYS (cysteine) areas. Below is the 5R80 Dimer with molecular fragments.

SARS-CoV-2 Main Protease Dimer co-crystal structure downloaded from RCSB PDB ([5R80](https://www.rcsb.org/structure/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.

A quantitative analysis provides a method to rank generated poses. This was done using Prithvi’s pose scoring primitive. Results are presented in a table from lowest to highest binding energy (kcal/mol).

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

Naringin presented the lowest score compared with other candidates. In terms of stability, we can say that 5R80 prefers to interact with naringin. Also, Naringin was surrounded by a binding site formed by LEU-141, GLU-166, and GLN-189 residues.

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.

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.

Next, we present a study case to assess molecular docking capabilities with Prithvi. Following inspiration from a recent study, we focus on potential inhibitors' interaction with SARS-CoV-2 main protease [2]. Active components from medicinal plants have shown beneficial effects such as antioxidants, antidiabetic, or anti-inflammatory with safe to use in terms of toxicity.

Five active compounds were used as ligands:-

Prithvi’s primitives offer a friendly UI to simulate protein-ligand interactions and visualize resulting poses. Images and results presented were taken using Prtihvi’s visualization features. Below are five drug candidates and SARS-CoV-2 main protease co-crystal structure.

Prithvi provides an intuitive setup to execute powerful scientific workflows. Our web platform extends different primitives for material and structure-based drug discovery. Prithvi facilitates scientists to use structural results to identify potential therapeutic agents related to specific diseases. Along this blog post, we showcase Prithvi’s molecular docking workflow and capabilities.

By incorporating protein ensemble tools that rely on open-source and AI algorithms, Prithvi enables users to conduct robust protein-ligand simulations. Prithvi’s docking suite extends its capabilities to assess and score several ligands and conformers as indicated in the case study, facilitating larger-scale molecular analysis. This broad search-space exploration extends to the identification of potential novel drug candidates.

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.

Since graph models like GCN take molecular graphs as input,  we use a custom featurizer which converts SMILES into graph structures where nodes and edges contain atom and bond features respectively.

A representation of the molecular graph featurization for paracetamol.

In this workflow, we can use our no-code UI to run the featurization process. We first specify the dataset to be featurized, in our case Delaney. Then we select the ECFP (circular/morgan fingerprint) featurizer and add custom parameters like 1024-bit size and 2 radius. We name the output featurized dataset as “feat_delaney_ecfp”. Similarly for GCN, we select our graph featurizer ‘MolGraphConv’, and store the output as “feat_delaney_fgcn”. Prithvi supports working with compute nodes of various specifications. We will choose a small node (vCPU - 4, RAM - 8 GB) to run this workflow since the dataset is small.

Symbolic representation of the no-code featurization procedure.

Splitting the Dataset using Scaffold Splits

Prithvi provides a Split workflow which splits datasets into for train, valid and test splits. Different splitting algorithms like random splitting, scaffold splitting, and others are supported.

We specify the featurized dataset “feat_delaney_ecfp” or “feat_delaney_fgcn” as our input and select the Scaffold split option on the no-code UI. In [scaffold splitting](https://www.oloren.ai/blog/scaff-split), the data is split into train 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 train set, and is molecularly distant from the train set. Ensuring the test set is maximally distant from the training set helps test models for generalizability and provides a better measure of its 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.

Symbolic representation of the no-code splitting procedure.

Hyperparameters Tuning for Parameters for RF and GCNs

Now we use the Hyperparameter Tuning workflow of Prithvi to test parameters across a specified hyperparameter search space. For simplicity, we will use a parameter grid, and run all the possible combinations to find the best model parameters to use based on a specified metric.

We select the random forest regressor model and define the hyperparameter space as follows:

Similarly, we select the graph convolutional network model and the train split. We define the hyperparameter space as follows:

We select the RMSE metric for tuning and choose a GPU node (vCPU - 4, RAM - 8 GB, GPU - 1) for tuning. Then we run hyperparameter search jobs for both the models.

Symbolic representation of no-code hyperparameter tuning.

Training Models using Selected Hyperparameters

We use the Train workflow to train models on specific parameters. Prithvi allows users to start multiple jobs through a single workflow. For each job, we choose the random forest regressor model or graph convolutional network model and the train split along with the best parameters from the Hyperparameter Tuning workflow. We run each training workflow on a GPU node (vCPU - 4, RAM - 8 GB, GPU - 1).

Symbolic representation of training 3 models using a no-code workflow.

Evaluating Models on Train/Valid/Test splits for RFs and GCNs

Next we use the Evaluate workflow of Prithvi to evaluate the model performance on specified datasets and metrics. After our models for each model type are trained, we evaluate models using the RMSE metric on the training, validation, and test splits. Here, again we submit multiple workflow jobs, one for each model. The evaluated results are stored in json files. These files can be downloaded locally for later use.

Symbolic representation of evaluating models on train/valid/test splits using a no-code workflow.

Prithvi provides a Featurize workflow which extracts atom-, bond- and molecule-level features from the [SMILES](https://pubs.acs.org/doi/10.1021/ci00057a005) strings in the dataset. SMILES (Simplified molecular-input line-entry system) is a line notation method to represent molecules as well as reactions. For example, we can represent ethane as follows:

A simple example of a molecule being converted to SMILES format.

We use extended connectivity fingerprints (ECFP) for the random forest models. During the [ECFP featurization](https://medium.com/@hacertilbec/introduction-to-cheminformatics-circular-fingerprints-f557cdc77492) process, local structural properties of the molecule are represented as a fixed-length binary array whose each element is either 1 or 0.

We observe that results for Random Forest models are similar to MoleculeNet numbers while the Graph Convolutional Network models do slightly worse. More systematic hyperparameter tuning should be able to close the gap for the graph convolution models.

Building good models that predict molecular properties can be valuable tools for designing new small molecules. Our Prithvi platform provides systematic infrastructure to build and test chemoinformatic machine learning models all through our easy no-code UI. Enabling chemists and biologists to build strong predictive models without having to write extensive python scripts unlocks the door to systematically leveraging ML and AI in the drug discovery process. In future blog posts in this series, we will explore how Prithvi can be used to build state-of-art models that can help scientists design new molecules that they would not have discovered otherwise.

Building and Benchmarking No-Code Chemoinformatic Machine Learning Models with Prithvi

We are adopting a new name Prithvi for our core platform (previously named Chiron). As Gaia denotes the Earth in Greek, Prithvi holds the same meaning in Sanskrit. Prithvi will serve as a foundation for accelerating projects in drug discovery and beyond and aligns with our core mission to accelerate AI-driven scientific discovery.

Stay tuned for more updates from Prithvi and Deep Forest Sciences!

Deep Forest Sciences is a deep tech R&D company leading the development of a no-code AI platform for small molecule discovery and design. Deep Forest Sciences emphasizes supporting open source and open science and considers them fundamental to our mission and values. Partner with us to apply our foundational AI technologies to hard real-world problems in drug discovery.

Introducing Prithvi

Density-Functional Theory (DFT) is a cornerstone of quantum chemistry that provides a more efficient method to understand the electronic structure of atoms, molecules, and materials. DFT simplifies the complex Schrödinger equation for many-electron systems by replacing the intricate many-body wave function with an electron density function which depends only on three spatial coordinates. This substitution allows us to describe the electronic ground state of complex molecular systems through single-particle equations and an effective potential comprised of three key components,

In this blog post, we discuss our recent [article](https://arxiv.org/abs/2309.15985) accepted into ICML'2023 Synergy and Scientific and Machine Learning Modeling (SynS & ML) workshop, and discuss our open-source model for Differentiable DFT which leverages machine learning to learn a predictive exchange correlation functionals. This work aligns with the broader efforts of Deep Forest Sciences, built around the open-source [DeepChem](https://deepchem.io/)ecosystem, to leverage powerful machine learning tools (including large chemical foundation models like [ChemBERTa](https://arxiv.org/abs/2010.09885) and [ChemBERTa-2](https://arxiv.org/abs/2209.01712)) to accelerate small molecule drug discovery. The integration of DeepChem with DFT could have significant implications for the field of drug discovery by helping us improve covalent docking and retrosynthesis.

To make differentiable DFT more efficient and accessible, we've developed open-source infrastructure aimed at training neural exchange correlation functionals. Our aim is to standardize and streamline the process of learning these functionals. We have open-sourced the model in [DeepChem](https://deepchem.io/), an open-source Python library designed to facilitate scientific machine learning calculations by breaking them down into structured workflows. DeepChem is tightly integrated into the Deep Forest Sciences ecosystem.

Table 1. MAE scores in kcal/mol for two test datasets.

The DFT community has started moving towards differentiable DFT due to challenges with defining XC functionals for complex systems. This 2021 [paper](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.127.126403) from Oxford presents a model, XCNN, that leverages deep neural networks to learn XC functionals and combine them with traditional XC functionals. This innovative approach is not system-specific, allowing it to be applied to a broad spectrum of physical systems. Another recent advancement is the development of the [DM21](https://www.science.org/doi/abs/10.1126/science.abj6511) functional by DeepMind, computed using fictional systems with fractional charge and spin constraints in order to overcome errors encountered by traditional XC functionals. DM21's method is particularly valuable when dealing with charge densities involving mobile charges and spins.

We aim to standardize differentiable DFT workflows, by making the process of training neural network XC functionals more accessible through DeepChem’s infrastructure. This standardization retains model accuracy, and is validated through experiments and comparisons with the original XCNN model. The [XCModel](https://deepchem.readthedocs.io/en/latest/api_reference/models.html#density-functional-theory-model-xcmodel), based on the XCNN implementation in DeepChem (adapted and modified from Oxford’s open source [release](https://github.com/mfkasim1/xcnn)), provides flexibility by including different loss functions, metrics, and machine learning models present in DeepChem. This infrastructure will facilitate the incorporation of more comprehensive density information and fractional constraints akin to the [DM21 functional](https://deepmind.google/discover/blog/simulating-matter-on-the-quantum-scale-with-ai/) from DeepMind and will help foster the development of new hybrid functionals trained using both neural networks and a fully differentiable quantum chemistry library. The diagram below shows an architectural schematic of DeepChem’s differentiable DFT infrastructure.

Schematic of XCModel architecture—loading yaml files with DFTYamlLoader, transforming them into mol objects with DFT data classes, initializing the Neural Network LDA layer (NNXC), and combining the functional with a traditional LDA functional. The hybrid xc solves Kohn-Sham equations, computes the total energy after self-consistent iteration convergence, and facilitates utilization with DeepChem functions for evaluation and prediction.

We tested our implementation by computing a hydrogen dissociation curve as shown in the figure below. In this experiment, we compare Mean Absolute Error (MAE) values for ionization potential and atomization energy using different functionals. Although our DeepChem implementation slightly trails the original, we expect to close this gap with further refinement. The experiment was executed on a 16GB RAM CPU, and illustrates how the XCModel accurately tracks dissociation energy, indicating its potential for precision in future larger-scale simulations.

Hydrogen dissociation curve. The Orange graph represents the exact values of energy required for H2 to dissociate at different bond lengths. The Blue graph represents the predicted energies using the trained model.

Open-source infrastructure for differentiable DFT in DeepChem can help simplify complex calculations, foster collaboration, and improve accuracy . Our research aligns with Deep Forest Sciences’ broader aim to accelerate AI-driven drug discovery. Future developments, including fractional constraints on fictional systems and Meta-PBE layers, can make the model even more versatile and accurate, especially for future applications in drug discovery such as covalent inhibitor design.

Deep Forest Sciences Prithvi suite provides a rich toolkit of computational approaches for drug discovery. Get in touch if you would like to use our cutting-edge AI approaches to accelerate your small molecule discovery efforts today!

Email us at partnerships@deepforestsci.com!

Open Source Infrastructure for Differentiable Density Functional Theory

Drug discovery is a long and complex process that requires significant time and resources. The pharmaceutical industry has historically depended on an expensive trial-and-error process that tests large numbers of compounds in the lab to identify potential drug candidates. The process is expensive, costing $2.6 billion per approved drug, and time consuming, taking 10-15 years to bring a new drug to the market. The full pipeline has very low yield, with tens of thousands (or even millions or billions) of compounds tested to produce a single clinical candidate, and a 90% failure rate between entering the clinic and reaching patients.

Modern high throughput assays enable very large numbers of compounds to be tested at early stages of discovery, but the number of viable compounds drops rapidly as a project proceeds through the stages of the drug discovery pipeline.

Each stage of the drug discovery pipeline brings both significant expenses and risks. Very few compounds reach the end of the pipeline and help patients. NME here means “novel molecular entity.” Adapted from source

AI techniques hold out the promise of leveraging data to transform trial-and-error into guided data-driven design. By leveraging powerful machine learning models in combination with rich biological, chemical and physical priors, it ought to be possible to lower failure rates and accelerate timelines for new therapeutics. However, AI techniques remain difficult to develop and deploy. Building large machine learning models requires sophisticated algorithmic understanding, robust cloud infrastructure, and extensive engineering know-how. While the first wave of AI-driven biotech companies has built internal AI infrastructure for discovery campaigns (at high cost), their tools remain internal and specialized to particular biological or chemical systems. As a result, most biotech teams today still lack the ability to deploy and utilize effective AI infrastructure for drug discovery.

Prithvi, our new AI powered drug discovery platform, aims to close this gap and enable world-class scientific teams to use state-of-art AI systems. Built upon years of our academic and open-source research, Prithvi allows scientists to deploy powerful AI tools without having to write a single line of code.

Prithvi uses AI and scientific machine learning algorithms to accelerate the drug discovery process. More specifically, Prithvi analyzes biological and chemical datasets in order to build machine learned models that can identify small molecule hits and optimize small molecule leads. Prithvi leverages pre-trained chemical foundation models in order to lower data requirements for practical drug discovery. Prithvi can help predict the efficacy of potential compounds, identify potential side effects, and prioritize compounds to test in secondary assays.

Prithvi can be used today to perform a broad range of scientific machine learning workflows to accelerate the drug discovery process. In particular, Prithvi can be used to

Early uses of AI in drug discovery have often been limited by low data availability. Prithvi leverages pre-trained foundation models to help systematically lower data requirements so biotech teams can start to leverage AI earlier in their design process.

Our research team is working to expand Prithvi’s applicability to every step of the drug discovery and design process. We believe that integrating Prithvi into your drug discovery pipeline can reduce your time to IND by months or more.

Prithvi is built around a powerful scientific machine learning engine that enables our users to run sophisticated scientific calculations without having to write any new code. Prithvi features proprietary datasets and models that users can license to bootstrap their AI efforts before building in-house datasets. Prithvi has out-of-box workflows to

Prithvi’s scientific machine learning engine is built around a powerful cloud infrastructure that allows users to run sophisticated workflows without having to write any code. Prithvi’s no-code capabilities empower scientists to leverage their own biological and chemical expertise without having to go through an intermediary data science team. We believe that providing scientists easy access to powerful AI workflows will unlock creativity and catalyze breakthrough discoveries.

Prithvi is built around the DeepChem ecosystem, an open source framework for molecular drug discovery and scientific machine learning. DeepChem was created by our CEO, Dr. Bharath Ramsundar, during the course of his PhD at Stanford. DeepChem offers a wide range of tools and models for scientific machine learning and deep learning, enabling researchers to predict properties, perform virtual screening, and conduct molecular featurization. DeepChem has been cited thousands of times and is used broadly in the academic and industrial drug discovery ecosystems (see this non-exhaustive listing of DeepChem-powered science). The accompanying book, Deep Learning for the Life Sciences, authored by Dr. Ramsundar and colleagues, has become a standard computational reference for the drug discovery community.

Like DataBricks, which is built around the Spark open source ecosystem created by academic research at UC Berkeley, Deep Forest Sciences is built around the DeepChem ecosystem. Prithvi productionizes and extends the DeepChem stack to handle real-world drug discovery. Prithvi leverages its cloud infrastructure to train models and gather datasets at an industrial scale not feasible with vanilla DeepChem

In particular, Deep Forest Sciences has taken a bet on the power of large scientific foundation models. The DeepChem team, led by Dr. Ramsundar, has developed the ChemBERTa and ChemBERTa-2 models, which helped demonstrate the power of chemical self-supervision to learn meaningful representations of drug-like molecules. By harnessing chemical foundation models such as ChemBERTa, Prithvi can leverage vast unlabeled datasets of chemical structures to unlock AI methods in real-world drug discovery even with limited data. Deep Forest Sciences has built both in-house cloud infrastructure and a network of collaborations with leading scientific institutions in order to train and deploy large foundation models for drug discovery and other applications.

Deep Forest Sciences has built a track record of successful partnerships with pharmaceutical companies and academic institutions to apply AI tools to real-world problems. For example, we recently partnered with Third Rock Ventures to accelerate AI-driven drug discovery for their portfolio companies using Prithvi, and have several additional partnerships in the works. These projects have validated Prithvi on real-world drug discovery campaigns.

Prithvi, our AI-powered scientific discovery platform, accelerates drug discovery in order to bring new treatments to patients faster. By equipping world-class biologists and chemists with world-class AI tools, Prithvi can unlock breakthrough scientific insights by enabling scientists to marry their intuition to cutting-edge AI workflows. Partnered with Deep Forest Sciences’ team of experts, Prithvi will provide your scientists the tools they need to drug challenging targets and design breakthrough therapeutics that will help save lives.

AI methods have already unlocked major scientific advances but have yet to revolutionize drug discovery. Prithvi will cross this chasm and transform how scientists discover new medicine by enabling AI-driven drug design.

Note- We have adopted a new name Prithvi for our core platform (previously named Chiron).

Blog Posts