Academic Databases
Explore Our Semantic Databases
STEM & Tech
Math, Physics, CS & Engineering...
Economics
Macro, Micro, Econometrics...
Business
Management, Marketing, Finance...
Health & Life Sciences
Biology, Medicine, Neuroscience...
Humanities & Social Sciences
Psychology, Law, Philosophy, History...
General Studies
General & interdisciplinary studies...
Explore Topics
Features

AI-Powered Molecule Design for Faster Drug Discovery

Added: 2026-08-07

Deep Dive

Introduction to Molecular Chemistry and Representations: Familiarity with how molecules are structurally represented computationally, such as SMILES strings and molecular graphs.

Stereochemistry by Pradeep Sir | Representations and their Interconversions | Lec-9 | Chem Academy

Stereochemistry by Pradeep Sir | Representations and their Interconversions | Lec-9 | Chem Academy

12.9K

This segment introduces the fundamental concept of molecular representations in chemistry. The instructor explains that molecular representations are visual tools used to depict the three-dimensional structure of molecules. The main types of representations discussed include Newman projections, Sawhorse projections, and Fischer projections. The instructor emphasizes that understanding these representations is essential for visualizing molecular geometry and stereochemistry. The segment sets the foundation for learning how to convert between different representations and interpret molecular structures.

Click to watch →

Molecular Representations pt 1

Molecular Representations pt 1

24.1K

Organic chemistry uses multiple molecular representation systems: molecular formulas show only atom counts (C3H8O for isopropyl alcohol), condensed structures show connectivity, Lewis structures display all bonds and lone pairs, and zig-zag/bond-line structures provide the most efficient visual representation where each corner represents a carbon atom and hydrogens attached to carbons are implied; key valency rules include carbon needing 4 bonds, hydrogen needing 1 bond, oxygen needing 2-3 bonds, and halogens needing 1 bond, with heteroatoms (non-carbon/hydrogen elements) always requiring explicit drawing.

Click to watch →

Chem-1A-Ch-2-Molecular-Compounds-2015-08-25

Chem-1A-Ch-2-Molecular-Compounds-2015-08-25

165

Molecular compounds consist of atoms held together by shared electrons (covalent bonds), unlike ionic compounds held by electrostatic attraction. A molecule is the basic unit of a molecular compound. Molecular formulas show atom counts (e.g., N2H4 for hydrazine). Different representations include structural formulas, ball-and-stick models, and space-filling models. In organic chemistry, carbon atoms are often simplified in line structures since carbon always makes four bonds and hydrogen always makes one.

Click to watch →

COMPRENDRE la représentation d'une MOLÉCULE | Physique-Chimie

COMPRENDRE la représentation d'une MOLÉCULE | Physique-Chimie

1.2K

A molecule is the smallest part of a pure substance (solid, liquid, or gas) that retains all its chemical properties, consisting of atoms linked by chemical bonds. Molecules can be represented through chemical formulas (indicating atom count and bond types), molecular models (using colored spheres for different atoms), or three types of formulas: the simple formula brute showing atom symbols and counts, the semi-developed formula showing non-hydrogen atom bonds, and the developed formula showing all atoms and bonds.

Click to watch →

Représentation des molécules

Représentation des molécules

34K

This section covers the foundational concepts of representing molecules in chemistry. A molecule is defined as an assembly of atoms connected by chemical bonds. Chemists use multiple representation methods based on their needs: molecular formula shows only atom counts (e.g., C5H11NO2 for valine) but cannot distinguish constitutional isomers (different molecules with same formula like ethanol vs. dimethyl ether). Expanded structural formula shows all bonds but is too cumbersome. Condensed structural formula groups CH bonds for efficiency but misses double bonds. Lewis structure shows all electrons including lone pairs but is very time-consuming. These methods reflect the trade-off between information completeness and practical usability in chemical communication.

Click to watch →

Deep Dive

Fundamentals of Machine Learning and Deep Learning: Understanding core neural network architectures, training processes, and optimization techniques.

Machine Learning And Deep Learning - Fundamentals And Applications [Introduction Video]

Machine Learning And Deep Learning - Fundamentals And Applications [Introduction Video]

61.3K

Machine Learning is a subset of Artificial Intelligence that enables machines to learn patterns from data without being explicitly programmed, while Deep Learning represents an advanced form of Machine Learning using artificial neural networks that adapt and learn from large amounts of training data. The field encompasses various learning types: Supervised Learning (with known desired outputs), Unsupervised Learning (without known outputs), Semi-Supervised Learning (with limited outputs), and Reinforcement Learning (learning through rewards from action sequences). Key mathematical foundations include linear algebra (vectors, dot products, eigenvectors, eigenvalues) and probability theory (mean, variance, normal distributions). Applications span diverse domains such as species recognition, speech processing, biometrics, optical character recognition, DNA sequence analysis, biomedical imaging, finance, e-commerce, robotics, healthcare, and autonomous systems.

Click to watch →

UP LEKHPAL COMPUTER MARATHON CLASS | UP LEKHPAL COMPUTER CLASSES | UPSSSC COMPUTER CLASSES

UP LEKHPAL COMPUTER MARATHON CLASS | UP LEKHPAL COMPUTER CLASSES | UPSSSC COMPUTER CLASSES

5.1K

Machine Learning is a subset of Artificial Intelligence where computers learn from data to perform tasks without explicit programming. It involves learning from past data, identifying patterns, and making predictions. Key examples include email spam detection, recommendation systems, face recognition, and voice recognition. The process involves data collection, preparation, model training, testing, and deployment. Machine Learning has four main types: Supervised Learning (uses labeled data), Unsupervised Learning (discovers hidden patterns in unlabeled data), Reinforcement Learning (learns through rewards and penalties), and Semi-Supervised Learning (combines both approaches). Supervised Learning includes Classification (predicting discrete categories) and Regression (predicting continuous values) with algorithms like SVM, Decision Trees, Naive Bayes, KNN, Linear Regression, and Polynomial Regression. Unsupervised Learning includes Clustering, Association, and Dimensionality Reduction with algorithms like K-Means, DBSCAN, and PCA. Deep Learning is a subset of Machine Learning that uses artificial neural networks to learn from large amounts of data, mimicking the human brain's structure. Neural Networks consist of Input Layer, Hidden Layer(s), and Output Layer with components including Neurons, Weights, Biases, and Activation Functions (ReLU, Tanh, Sigmoid, Softmax). Types include Artificial Neural Networks, Convolutional Neural Networks (CNN) for image processing, Recurrent Neural Networks (RNN) for sequential data, and LSTM for long-term memory.

Click to watch →

Machine Learning y Deep Learning

Machine Learning y Deep Learning

120

Machine learning enables computers to learn from data without explicit programming, identifying patterns and improving over time. Algorithms serve as the core recipes guiding this process, with three main types: supervised learning (learning from labeled data), unsupervised learning (finding patterns in unlabeled data), and reinforcement learning (learning through rewards and penalties). Deep learning, a specialized subfield, uses neural networks with multiple layers to learn hierarchical features automatically, enabling complex tasks like image recognition and natural language processing. Unlike traditional machine learning that requires manual feature engineering, deep learning can automatically extract features from raw data, though it demands significantly more computational resources and specialized hardware like GPUs.

Click to watch →

Machine Learning and Deep Learning (Business Applications) | From A Business Professor

Machine Learning and Deep Learning (Business Applications) | From A Business Professor

1.2K

Machine learning (ML) is a subset of AI that enables systems to learn from experience without explicit programming, using data to identify patterns and improve over time. ML differs from traditional programming by teaching computers to recognize patterns through exposure to examples rather than direct instruction. There are three main types: supervised learning (training on labeled datasets with input-output pairs), unsupervised learning (inferring structure from unlabeled data), and reinforcement learning (learning through rewards/punishments). Deep learning (DL) is a specialized ML subset using neural networks with multiple layers inspired by the human brain, capable of handling unstructured data like images and text. DL adds depth and sophistication to the learning process, enabling complex pattern recognition.

Click to watch →

Ai & ML सीखो फ्री में | FREE AI/ML Courses in 2025 | Complete Roadmap

Ai & ML सीखो फ्री में | FREE AI/ML Courses in 2025 | Complete Roadmap

20.1K

Three essential Python libraries form the foundation: NumPy for numerical computing and arrays, Pandas for data manipulation and analysis, and Matplotlib for visualization. Machine Learning is divided into three categories: Supervised Learning (regression and classification with labeled data), Unsupervised Learning (clustering and dimensionality reduction with unlabeled data), and Reinforcement Learning (reward-based learning for applications like self-driving cars). Recommended resources include Andrew Ng's Coursera course, Google's ML Crash Course, Cornell CS 4780 lectures, and Abhishek Thakur's practical ML book. Deep Learning involves neural networks including ANN, CNN, RNN, and GAN. Key resources include AI For Coders' practical approach, D2L AI's interactive textbook, Andrej Karpathy's Neural Networks Zero to Hero series, and UvA Deep Learning Notebooks covering computer vision and NLP.

Click to watch →

Deep Dive

Basic Graph Theory: Conceptual knowledge of graphs, including nodes (atoms), edges (bonds), and adjacency matrices, which form the basis of Graph Neural Networks (GNNs).

Basic Graph Theory I - vertices, edges, loops, and equivalent graphs

Basic Graph Theory I - vertices, edges, loops, and equivalent graphs

44.1K

In graph theory, a graph consists of vertices (dots) connected by edges (lines), where a loop is an edge connecting a vertex to itself; two graphs are equivalent if they have the same number of vertices and edges connecting the same pairs of vertices, regardless of how they are visually drawn.

Click to watch →

Graph Theory | Basic Concept Of Graph Theory | Discrete Mathematics | Bsc 3 year| L-1

Graph Theory | Basic Concept Of Graph Theory | Discrete Mathematics | Bsc 3 year| L-1

26K

A graph is a mathematical structure consisting of two sets: vertices (V) and edges (E), where edges connect vertices. Key concepts include: (1) Types of graphs - trivial graph (one vertex, no edges), empty graph (multiple vertices, no edges), directed graph (edges with direction), undirected graph (edges without direction), self-loop (edge connecting a vertex to itself), parallel edges (multiple edges between same vertices), simple graph (no self-loops or parallel edges), finite graph (fixed number of vertices/edges), and infinite graph (unlimited vertices/edges); (2) Vertex degree - the number of edges incident to a vertex, with self-loops counting twice; (3) Handshaking Theorem - the sum of all vertex degrees equals twice the number of edges; (4) Even and odd degree vertices - a vertex is even if its degree is even, odd if its degree is odd; (5) Theorem - the number of odd-degree vertices in any graph is always even; (6) In-degree and out-degree in directed graphs - in-degree counts incoming edges, out-degree counts outgoing edges.

Click to watch →

ΠΛΗ20 ΜΑΘΗΜΑ 4.1: ΒΑΣΙΚΟΙ ΟΡΙΣΜΟΙ ΘΕΩΡΙΑΣ ΓΡΑΦΩΝ - ΘΕΩΡΙΑ 1 από 2

ΠΛΗ20 ΜΑΘΗΜΑ 4.1: ΒΑΣΙΚΟΙ ΟΡΙΣΜΟΙ ΘΕΩΡΙΑΣ ΓΡΑΦΩΝ - ΘΕΩΡΙΑ 1 από 2

43.7K

This lesson introduces fundamental graph theory concepts including undirected graphs (vertices and edges without direction), directed graphs (edges with direction), paths (sequences of edges connecting vertices), cycles (paths starting and ending at the same vertex), simple graphs (no loops or parallel edges), complete graphs (cliques with all possible edges between vertices), connected graphs (any two vertices connected by a path), connected components (maximal connected subgraphs), cut vertices and bridges (removing which disconnects the graph), and graph complements (containing all edges not in the original graph). Key formulas include: complete graph with n vertices has n(n-1)/2 edges, and the sum of edges in a graph and its complement equals the edges of the corresponding complete graph.

Click to watch →

Introduction to Graph Theory|Discrete Mathematics|BBA|BCA|B.COM|Dream Maths

Introduction to Graph Theory|Discrete Mathematics|BBA|BCA|B.COM|Dream Maths

271.9K

Graph theory studies diagrams composed of vertices (points) connected by edges (lines or curves). A graph is denoted as G = (V, E), where V represents vertices and E represents edges. Two fundamental graph types exist: directed graphs with arrowed edges specifying direction (having initial and terminal vertices), and undirected graphs without arrows. Key vertex properties include adjacency (vertices sharing a direct connection), loops (edges connecting a vertex to itself), and isolated vertices (unconnected vertices). Parallel edges share identical initial and terminal vertices. Adjacent edges share exactly one common vertex but are not parallel. These foundational concepts establish the vocabulary and framework for all subsequent graph theory study.

Click to watch →

Basic Concepts

Basic Concepts

86.1K

A graph is defined as a pair (V, E) where V is the set of vertices and E is the set of edges (subsets of V × V). Key concepts include: (1) Degree of a vertex is the number of edges incident to it; (2) A graph is k-regular if all vertices have degree k; (3) The Handshaking Lemma states that the sum of all vertex degrees equals twice the number of edges; (4) In any graph, the number of vertices with odd degree is always even; (5) A complete graph Kn has every pair of vertices adjacent, with n(n-1)/2 edges.

Click to watch →

Deep Dive

Principles of Natural Language Processing (NLP): Basic understanding of sequence-to-sequence models and transformer architectures, which are often adapted for processing chemical sequences.

Natural Language Processing (NLP), Deep Learning (DL), Machine Learning (ML) & Transformers

Natural Language Processing (NLP), Deep Learning (DL), Machine Learning (ML) & Transformers

1.4K

Natural Language Processing (NLP) is a subfield of artificial intelligence focused on enabling machines to understand, interpret, generate, and respond to human language meaningfully. The four core functions are: understand, interpret, generate, and respond. NLP's primary goals include understanding syntax (grammar) and semantics (meaning), enabling machine translation, speech recognition, text classification, and sentiment analysis. Common NLP tasks encompass text classification (spam detection), named entity recognition (identifying names, places, dates), machine translation (language conversion), part-of-speech tagging (identifying grammatical categories), question answering (responding to factual queries), text summarization (generating concise document summaries), and conversational AI (powering chatbots).

Click to watch →

NLP Demystified 1: Introduction

NLP Demystified 1: Introduction

42.2K

Natural Language Processing (NLP) is a multidisciplinary field combining math, linguistics, and computer science that enables computers to perform useful tasks with natural language data such as translation, summarization, question answering, and speech recognition. Despite dramatic advances, NLP remains challenging because natural language is inherently ambiguous, contextual, and full of metaphors and expressions that require sophisticated computational approaches to resolve. The field has evolved from early rules-based systems in the mid-20th century through statistical models in the 1980s to today's deep learning approaches, with modern production systems often combining all three approaches. This course covers both classical methods and deep learning techniques, including text preprocessing, vectorization, machine learning workflows, text classification, neural networks, word embeddings, sequence modeling, attention mechanisms, and transformer architectures.

Click to watch →

NLP Explained: Natural Language Processing Basics for Beginners

NLP Explained: Natural Language Processing Basics for Beginners

Natural Language Processing (NLP) is a branch of artificial intelligence that enables computers to understand, interpret, and generate human language by converting unstructured text (like spoken words or written sentences) into structured data that machines can process; this involves techniques such as tokenization (breaking text into individual words), stemming (reducing words to their root form), lemmatization (finding dictionary-based word roots), part-of-speech tagging (identifying grammatical roles), and named entity recognition (identifying specific entities like people or places), with applications including machine translation, virtual assistants, sentiment analysis, and spam detection.

Click to watch →

Introduction to NLP | NLP Lecture 1 | End to End NLP Course

Introduction to NLP | NLP Lecture 1 | End to End NLP Course

403K

Natural Language Processing (NLP) is an interdisciplinary field combining linguistics, computer science, and artificial intelligence that enables computers to understand, interpret, and generate human language. NLP has evolved through three main approaches: rule-based methods (1900s-1990s), machine learning approaches (since 1980s), and deep learning approaches (since 2010). Common NLP tasks include text classification, sentiment analysis, information retrieval, parts of speech tagging, machine translation, speech-to-text conversion, and text summarization. Despite significant advances, NLP faces challenges such as ambiguity, polysemy, slang, irony, sarcasm, and the vast diversity of human languages worldwide.

Click to watch →

Natural Language Processing In 5 Minutes | What Is NLP And How Does It Work? | Simplilearn

Natural Language Processing In 5 Minutes | What Is NLP And How Does It Work? | Simplilearn

804.6K

Natural Language Processing (NLP) is a branch of artificial intelligence that enables machines to read, understand, and derive meaning from human languages by combining linguistics and computer science; the process involves several key steps including segmentation (breaking text into sentences), tokenization (splitting sentences into individual words), removing stop words (non-essential words like 'the' and 'are'), stemming (reducing words to their base form), lemmatization (identifying root words), part-of-speech tagging (labeling words as nouns, verbs, etc.), and named entity recognition (identifying proper nouns like people and places), ultimately allowing machines to comprehend and respond to human language intelligently.

Click to watch →

Deep Dive

The Traditional Drug Discovery Pipeline: A general overview of the steps involved in bringing a drug to market, including target identification, lead generation, and clinical trials.

AI & Drug Discovery | Nirnith Devireddy | InSilicoMinds | IEFR Lecture Series | SciKonnect Academy

AI & Drug Discovery | Nirnith Devireddy | InSilicoMinds | IEFR Lecture Series | SciKonnect Academy

129

The traditional drug development pipeline progresses through distinct phases: Discovery (target identification, hit generation, lead identification, lead optimization), Preclinical/Non-clinical (animal models), and Clinical stages (Phase 1, Phase 2, Phase 3). This process is time-consuming, costly, and carries very low success rates. Computational approaches can intervene at multiple points: Discovery (drug repurposing, target identification, bioactivity predictions, molecular docking), Non-clinical (toxicity endpoints, cardiotoxicity, mutagenicity, ADMET properties), and Clinical stages (simulating safety, efficacy, dose response in various populations and disease states).

Click to watch →

Bioinformatics and Drug Discovery: Will We Reach 1000 Novel Molecules per Year?

Bioinformatics and Drug Discovery: Will We Reach 1000 Novel Molecules per Year?

478

The traditional drug discovery pipeline involves sequential stages: plant authentication, extraction based on polarity, structural elucidation, biological screening, instrumental analysis, molecular docking, retro-synthesis, and ADMET profiling. Clinical trials progress through four phases evaluating safety, efficacy, and special populations. Starting with 10,000 molecules, the process filters down to 250 pre-clinical candidates, then to 5 clinical candidates, with only one potentially approved. This represents a success rate of approximately 0.01%, highlighting the inefficiency of classical approaches.

Click to watch →

Drug Discovery and Design Class 3: Introduction to Computer-Aided Drug Design

Drug Discovery and Design Class 3: Introduction to Computer-Aided Drug Design

188

The traditional drug discovery pipeline consists of five main stages: (1) Target identification - identifying proteins, enzymes, or genes that play key roles in disease mechanisms; (2) Hit identification - screening thousands to millions of compounds using high-throughput screening to find molecules that bind to the target; (3) Lead optimization - modifying hit compounds to improve potency, selectivity, safety, and stability; (4) Preclinical and clinical testing; and (5) Regulatory approval. This process typically takes 10-15 years and costs billions of dollars, with a clinical success rate below 10%.

Click to watch →

Sam Altman Says AI Will Cure Cancer. I Looked Into It.

Sam Altman Says AI Will Cure Cancer. I Looked Into It.

522

The traditional drug discovery process follows a funnel-like pipeline that is time-consuming and expensive. It begins with preclinical research where researchers identify disease-related proteins by comparing healthy tissue with unhealthy tissue to find implicated genes. This target identification stage alone often takes 1-3 years because establishing causal links between proteins and diseases requires extensive academic research.

Click to watch →

AlphaFold 3 | El invento más importante del siglo

AlphaFold 3 | El invento más importante del siglo

128.5K

Traditional drug discovery follows an experimental approach: identifying a disease-causing protein, then finding a chemical compound (the 'key') that fits perfectly into the protein's binding site (the 'lock'). Pharmaceutical companies maintain massive physical libraries of millions of compounds. Robots test compounds one by one in Petri dishes. This process costs over $2 billion and takes 10-15 years on average, making it a game of trial and error.

Click to watch →

AI for Medicine

Drug Discovery Costs

Accelerating Search

Molecule Languages

Modeling Molecules

Inspiration Sources

Key Challenges

Target Binding

Human AI Team

Data Sharing

AI for Medicine

0:15

Playing Section

1
Lei Li introduces AI and machine learning for medicine.
2
Focuses on generating molecules for drug purposes.
3
Motivated to help humanity after early Covid.

Deep Dive

Generative Chemistry Models: Deep dive into advanced generative frameworks such as Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and Diffusion Models applied to molecular design.

Machine Learning in Chemistry

Machine Learning in Chemistry

4.5K

Generative models for molecules face interconnected challenges: validity (decoding to chemically valid structures), stability (acid/base/photo/thermal stability), and synthesizability (ability to be made in lab). Common representations like SMILES decode to invalid structures (e.g., carbon with five bonds). Solutions include validity-enforcing objectives or representations. Even valid molecules may lack stability or synthesizability. A novel approach teaches elementary chemistry concepts alongside target property prediction—training models to predict multiple chemical properties (even uncorrelated ones) enriches the latent space with chemical realism. Internal energy, spatial extent, zero-point vibrational energy, and heat of formation serve as transfer learning variables. This dramatically improves extrapolation—validity increases from ~4-5% to ~40% in high-energy regimes and from ~12% to ~25% in low-band-gap regimes. The community needs better surrogates for stability and synthesizability. Alternative approaches constrain chemical space to synthesizable molecules using reaction predictors, though additional constraints may limit exploration. The fundamental question remains: should we prioritize mimicking existing chemistry or discovering fundamentally new molecules?

Click to watch →

Generative Chemistry in real world application | AI for Drug Discovery | Jessica Lanini

Generative Chemistry in real world application | AI for Drug Discovery | Jessica Lanini

510

Generative chemistry employs three core components: molecule generators using generative models, multi-objective optimization algorithms navigating chemical space, and post-processing for ranking solutions. Molecular representation methods include string-based approaches using LSTM autoencoders and graph-based methods modeling atoms as vertices and bonds as edges. MolER, developed through Novartis-Microsoft collaboration, addresses scaffold-fixing challenges by incorporating motive-level generation alongside atom-level generation. Motives represent common substructures derived from public data, enabling probability distribution computation and imposed scaffold constraints. Multi-objective optimization navigates the vast chemical space (estimated at 10^60 molecules) toward desired properties, generating thousands of candidates requiring systematic evaluation and filtering.

Click to watch →

Generative AI for Molecular Design & Synthesis

Generative AI for Molecular Design & Synthesis

17.4K

Generative AI explores vast chemical space by aligning computational thinking with chemist capabilities and robotic synthesis platforms. Constrained approaches using commercial building blocks guarantee synthesizability but limit exploration, while unconstrained approaches maximize creativity at practical cost. Current models compete with top medicinal chemists in predicting reaction compatibility, significantly improving average performance. The future direction involves demonstrating utility through experimental validation and developing methods to steer models toward better structures through scoring and evaluation.

Click to watch →

Кафедра ММП | 25 | Введение в глубокое обучение | ВМК МГУ

Кафедра ММП | 25 | Введение в глубокое обучение | ВМК МГУ

117

This section introduces the fundamental concepts of generative models and their contrast with discriminative models. Discriminative models learn relationships between conditions and parameters to distinguish classes, while generative models aim to model the complete joint probability distribution of data and labels. The section covers the historical development of GANs, proposed by Ian Goodfellow in 2014, and explains the adversarial training paradigm where a generator creates synthetic data while a discriminator learns to distinguish real from generated samples. Practical applications in chemistry for molecule generation are mentioned as examples of useful generative model applications.

Click to watch →

Expanding AI and next generation simulation tools to accelerate scientific discovery

Expanding AI and next generation simulation tools to accelerate scientific discovery

434

Foundation models in chemistry are created by training on massive datasets of simulation data, such as millions of density functional theory calculations, to develop foundational understanding of molecular assembly principles. These models can then be fine-tuned on experimental data for specific applications, enabling rapid property prediction. A collaboration with Pacific Northwest National Laboratory demonstrated this approach by screening 32 million candidate solid-state electrolyte structures, narrowing to 500,000 stable materials through AI inference, then to a handful of candidates through HPC validation. One resulting material achieved 70% less lithium than best-in-class electrolytes while functioning in a working battery. Generative chemistry extends this capability by enabling researchers to escape existing molecular databases and discover truly novel compounds in unexplored regions of chemical space. The workflow begins with expert-defined seed molecules, generates diverse candidate pools through AI fine-tuning, and applies multi-stage screening to identify synthetizable, useful molecules. This represents a paradigm shift toward computational creativity in drug discovery and materials development.

Click to watch →

Deep Dive

Reinforcement Learning for Molecular Optimization: Learning how to steer generative models to design molecules with specific target properties using reward-based feedback loops.

webinar recording: AI Driven De Novo Design with REINVENT

webinar recording: AI Driven De Novo Design with REINVENT

5.9K

Reinforcement learning applies machine learning principles to molecular design by training generative models as agents to maximize rewards from scoring functions. The iterative process involves generating molecule batches, scoring them against objectives (activity, similarity, diversity), and learning to improve generation strategies. Key considerations include balancing exploration (discovering new scaffolds) versus exploitation (refining known compounds). Diversity filters help manage this trade-off by maintaining compound variety in generated populations. This framework enables automated optimization of complex molecular properties without explicit programming of chemical rules.

Click to watch →

Deep Generative models for molecular de novo design by Iktos: A real life case study

Deep Generative models for molecular de novo design by Iktos: A real life case study

1.1K

Actos implements a hybrid system combining generative models with reinforcement learning. The framework uses QC models trained on client data to score molecules across multiple objectives (activity, selectivity, solubility, stability). Scores feed back into the generative model, directing exploration toward optimal regions of chemical space. This enables rapid convergence to multi-parametric optimal solutions that would require extensive human-led experimentation.

Click to watch →

Esben Bjerrum: Faster and more diverse de novo molecular optimization using augmented SMILES

Esben Bjerrum: Faster and more diverse de novo molecular optimization using augmented SMILES

193

The double loop architecture enhances standard reinforcement learning by introducing an inner loop where multiple SMILES augmentations are generated and scored before updating the agent. In each inner iteration, different SMILES representations of the same molecule are created, and their scores (from expensive computations like docking) are reused since they represent the same compound. Only after accumulating updates from multiple augmentations does the outer loop generate a new batch. This approach dramatically reduces expensive scoring evaluations while maintaining or improving optimization performance. Research demonstrates that 5-10 augmentations per loop provide optimal results for most tasks, balancing exploitation (finding good solutions quickly) against exploration (maintaining diversity). Too few augmentations result in slow learning, while too many cause overfitting and reduced diversity.

Click to watch →

9ADD Thursday lectures

9ADD Thursday lectures

136

Molecular generation architectures include: (1) RNNs treating SMILES as sentences, good at reproducing similar molecules but struggle with scaffold hopping; (2) Variational Autoencoders encoding molecules into latent spaces for multi-objective optimization but risking invalid molecules; (3) GANs with generative and discriminative networks, combined with reinforcement learning for multi-objective optimization; (4) Reinforcement learning where agents perform actions (adding atoms) in environments (QSAR models, docking scores) with reward schemes. RL can overcome RNN limitations but improper reward schemes can lead to nonsensical molecules. DrugX is a reinforcement learning framework using two networks: an agent network and a prior/exploration network. The prior network grounds molecules in chemical space (like ChEMBL data), while the agent optimizes objectives. The epsilon parameter controls the balance between exploration and exploitation. Objectives typically include QSAR predictions and synthetic accessibility scores. Synthetic accessibility prediction faces applicability domain challenges—models trained on ChEMBL compounds perform poorly on generated molecules from different chemical space.

Click to watch →

Webinar 13: “Are we there yet?” Emerging drug discovery challenges for SARS-CoV-2 with AI driven MD

Webinar 13: “Are we there yet?” Emerging drug discovery challenges for SARS-CoV-2 with AI driven MD

249

This section explains how reinforcement learning enables automated molecular optimization by combining physics-based simulations with AI-driven exploration. The framework defines action spaces including fragment growth and chemical transformations, with rewards based on binding free energy, docking scores, and RMSD values. Through continuous interaction, learners explore chemical space toward tighter binders. Case studies demonstrate successful optimization of kinase inhibitors and main protease ligands, showing progressive improvement in binding metrics over multiple iterations. This approach automates lead optimization without requiring explicit expert guidance at each step.

Click to watch →

Deep Dive

AI-driven Retrosynthesis Planning: Exploring how machine learning algorithms predict the chemical reactions and synthetic pathways required to physically manufacture a designed molecule in the lab.

Cheminfo Stories 2021 APAC | Chemical.AI retrosynthesis platform with ChemAxon modules

Cheminfo Stories 2021 APAC | Chemical.AI retrosynthesis platform with ChemAxon modules

257

AI-driven retrosynthesis planning platforms like Cheminfo Stories 2021 APAC's ChemAI can predict multiple synthetic routes for complex molecules by combining machine learning algorithms with chemistry expertise, enabling chemists to explore diverse reaction pathways more efficiently than traditional methods. The platform uses techniques such as molecular fingerprint generation for rapid fragment comparison, reaction template extraction for building structural databases, and substructure matching for single-step reaction prediction, ultimately helping researchers identify feasible synthesis routes with supporting references in significantly reduced time.

Click to watch →

Dr. Samuel Genheden (AstraZeneca) --- Exploiting artificial intelligence in synthesis planning

Dr. Samuel Genheden (AstraZeneca) --- Exploiting artificial intelligence in synthesis planning

180

AI-assisted retrosynthesis planning uses machine learning models to break down target molecules into synthetic precursors, enabling chemists to design efficient synthesis routes for drug development. Modern approaches include template-based methods that apply learned chemistry rules to decompose molecules, and generative transformer models that predict retrosynthetic disconnections without relying on predefined templates. These computational tools help medicinal chemists quickly generate synthesis plans for new drug candidates, while process chemists optimize large-scale manufacturing processes. The field continues evolving with constrained retrosynthesis techniques that preserve specific bond-breaking patterns for structure-activity relationship studies, and ongoing challenges include improving model reliability, expanding chemical space coverage, and bridging the gap between computational predictions and experimental validation.

Click to watch →

逆合成解析ツール CAS Retrosynthesis

逆合成解析ツール CAS Retrosynthesis

155

CAS Retrosynthesis is an AI-driven tool that automatically analyzes synthetic pathways by learning from CAS SciFinder's extensive reaction database to propose optimal routes, with key features including configurable reaction rules (common vs. rare reactions), break and protect bond settings, cost estimation, yield prediction, and real-time route modification capabilities that enable chemists to efficiently explore diverse synthetic strategies for target molecules.

Click to watch →

Chemistry42 PHARMA.AI WEBINAR WEEK Nov 2024

Chemistry42 PHARMA.AI WEBINAR WEEK Nov 2024

880

AI-powered molecular generation platforms like Chemistry 42 combine generative models with reinforcement learning to create molecules with desired properties, while retrosynthesis tools use disconnection-based approaches and expert-coded reaction templates to generate synthetic routes that consider functional group compatibility, physiochemical properties, and stereochemistry, enabling researchers to efficiently design and synthesize complex drug candidates.

Click to watch →

Application of artificial intel. (AI) to chemical reaction engineering and processes intensification

Application of artificial intel. (AI) to chemical reaction engineering and processes intensification

1.3K

AI transforms chemical synthesis planning through data-driven approaches: (1) Reaction templates encode expert-curated transformations, while machine learning extracts symbolic rules from millions of published reactions; (2) Deep learning models convert product representations to reactant predictions, enabling recursive tree searches for multi-step pathways; (3) Supervised learning predicts reaction conditions (reagents, catalysts, solvents, temperatures) from reaction examples; (4) Pathway scoring evaluates feasibility, atom economy, green chemistry metrics, and potential for telescoping; (5) Tools like ASOS provide accessible interfaces for recommending synthetic routes. Despite advances, predicting full procedural details (flask preparation, addition order, purification) remains challenging due to data limitations and the complexity of bridging AI predictions with empirical laboratory execution.

Click to watch →

Deep Dive

ADMET Prediction Models: Utilizing deep learning to predict the Absorption, Distribution, Metabolism, Excretion, and Toxicity profiles of newly generated drug candidates.

SwissADME a web tool to support pharmacokinetic optimization for drug discovery

SwissADME a web tool to support pharmacokinetic optimization for drug discovery

21.6K

Computational ADMET prediction balances robustness, speed, and interpretability for drug discovery. SwissADME implements three lipophilicity approaches: fragmental methods (fast but limited chemical space), topological methods (QSPR with statistical models, open-source but prone to overfitting), and physics-based methods (physically interpretable but computationally intensive). Binary classification models predict physiological barriers using optimized decision boundaries—the BOILED-Egg model predicts gastrointestinal absorption with cLogP vs TPSA descriptors, while similar approaches predict BBB permeation with narrower acceptance criteria. SVM classification models predict P-glycoprotein substrate status achieving ~72% cross-validation accuracy. These models capture passive transport mechanisms but cannot account for active transporter effects, motivating consensus approaches averaging predictions from multiple methods to improve reliability across diverse chemical spaces.

Click to watch →

AI 기반 약물 예측 플랫폼 AI-Exp, 무엇이 다를까? (장지돈 한국화학연구원 선임연구원)

AI 기반 약물 예측 플랫폼 AI-Exp, 무엇이 다를까? (장지돈 한국화학연구원 선임연구원)

192

This section details the technical architecture of ADMET prediction models. Data sources include public databases and in-house Korean Compound Bank data. Feature engineering converts molecular structures into numerical vectors using fingerprints, graph convolution, and physicochemical descriptors. Models are categorized as classification (binary) or regression (quantitative). The heart toxicity model uses ensemble methods with three models to minimize false negatives, significantly improving negative predictive value. Quantitative models like XGBoost predict specific IC50 values, providing more intuitive results for researchers.

Click to watch →

약물의 ADMET 특성 예측 모델 개발 | 아이펠톤

약물의 ADMET 특성 예측 모델 개발 | 아이펠톤

292

Drug development costs $2.6B and takes over 5 years, with 90% of candidates failing in clinical trials. ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) properties are critical for predicting drug safety and efficacy. Molecular data is stored as SMILES codes and processed using RDKit to convert to 2D graphs or 3D coordinates. Three representation modalities exist: 1D (sequence-based NLP approaches like BERT), 2D (graph neural networks capturing atomic interactions), and 3D (spatial arrangements requiring energy minimization). Advanced architectures include Mixture of Experts (combining specialized sub-models with routing gates), InfoGraph and InfoMax (information-theoretic GNN improvements), and TopoPool (topological data analysis using persistent homology). Systematic hyperparameter tuning was performed across all approaches to optimize performance for ADMET prediction tasks.

Click to watch →

Introducción a Propiedades ADMET por Quimioinformática para Selección de Fármacos

Introducción a Propiedades ADMET por Quimioinformática para Selección de Fármacos

451

ADMET prediction software includes free tools like ADMETLAB, SwissADMET, and PKCSM, and commercial tools like ACD/Labs. ADMETLAB provides: physicochemical properties (MW, LogP, LogS, TPSA, rotatable bonds, volume), Lipinski and Veber rule compliance, ADMET predictions (intestinal absorption, P-gp status, blood-brain barrier permeability, protein binding, metabolic stability, clearance, half-life), and toxicity predictions. SwissADMET evaluates LogP, LogD, TPSA, intestinal absorption (Caco-2 cells), blood-brain barrier permeability, protein binding, metabolic stability, clearance, half-life, and toxicity. Databases like PubChem, DrugBank, and ChemSpider provide compound information. The Globally Harmonized System (GHS) classifies toxicity based on LD50 values.

Click to watch →

국내 유일의 은행💵, 국내 유일의 플랫폼! | 🐤곽곽 ep.17

국내 유일의 은행💵, 국내 유일의 플랫폼! | 🐤곽곽 ep.17

137.2K

The ADMET prediction platform uses artificial intelligence to predict drug properties including Absorption, Distribution, Metabolism, Excretion, and Toxicity. Researchers draw chemical structures, and AI models predict properties like Blood-Brain Barrier permeability. The platform has 23,000 monthly active users globally and has been maintained for over 10 years. It has been cited in over 500 papers annually since 2023. The platform addresses the 50-70% failure rate in drug development by helping researchers identify promising compounds earlier. An incentive model allows data contributors to earn rewards when others use their models, creating a sustainable ecosystem for scientific collaboration.

Click to watch →

Deep Dive

Real-World AI Drug Discovery Case Studies: Examining clinical-stage drug candidates designed by AI companies (e.g., Insilico Medicine, Exscientia) to understand the practical challenges of translation from silicon to clinic.

9 20202155 한채림 AI기반 신약설계

9 20202155 한채림 AI기반 신약설계

604

A practical case study demonstrates AI's effectiveness in drug discovery. A pharmaceutical company partnered with an AI firm to develop a 300mg therapeutic drug targeting DDR1 kinase. Using AI, they identified a new target compound, completed synthesis and testing in just 46 days (21 days for synthesis and 25 days for testing). This represents a dramatic reduction from the traditional 5-year timeline, demonstrating AI's potential to significantly accelerate drug development.

Click to watch →

How Artificial Intelligence is changing drug discovery

How Artificial Intelligence is changing drug discovery

16.2K

In 2021, German biotechnology company Evotec discovered a new anti-cancer molecule in just eight months using Extensia's AI design platform. Recently, the world's first AI-designed drug entered clinical trials, demonstrating practical applications of AI in pharmaceutical research.

Click to watch →

QUIBIM WEBINAR | AI and Radiomics in Drug Development Challenges and Approaches

QUIBIM WEBINAR | AI and Radiomics in Drug Development Challenges and Approaches

229

Two notable real-world examples demonstrate AI's impact in drug discovery: (1) BenevolentAI used machine learning to mine structured datasets and medical literature during the COVID-19 pandemic, discovering that baricitinib (an approved JAK inhibitor for rheumatoid arthritis) had antiviral properties by mediating SARS-CoV-2 entry into cells, achieving this in days rather than the typical lengthy drug development timeline; (2) X-Chem fully designed a molecule for a known target (serotonin 5-HT1A receptor agonist), representing the first fully AI-designed compound to enter clinical trials, demonstrating AI's capability to design new molecules for existing targets.

Click to watch →

Week 5: Abraham Heifets, Deep Learning for Drug Discovery

Week 5: Abraham Heifets, Deep Learning for Drug Discovery

1.7K

The speaker presents real-world applications of AI in drug discovery, including an Ebola project where a neural network predicted a molecule that blocked viral entry in cell culture by about 80%. The compound worked for the right reason (blocking the specific entry mechanism) rather than coincidentally, demonstrating that AI can identify molecules with genuine biological activity.

Click to watch →

"Artificial Intelligence in Therapeutics" by reputed Speakers: Dr Nayan Choudhary & Dr Amit Dey

"Artificial Intelligence in Therapeutics" by reputed Speakers: Dr Nayan Choudhary & Dr Amit Dey

154

A real-life example demonstrates AI's impact: a company scanned research to shortlist 20 targets for hypopharyngeal carcinoma, then used AI to narrow it down to the CDK20 receptor. AlphaFold AI predicted CDK20's 3D structure, and Chemistry42 AI platform designed chemicals targeting CDK20 receptors, narrowing down to seven drug candidates. This entire procedure took only 30 days, whereas traditional clinical trials would have taken 6-12 months. This represents a dramatic time reduction in drug discovery.

Click to watch →

AI for Medicine

Drug Discovery Costs

Accelerating Search

Molecule Languages

Modeling Molecules

Inspiration Sources

Key Challenges

Target Binding

Human AI Team

Data Sharing

AI for Medicine

0:15

Playing Section

1
Lei Li introduces AI and machine learning for medicine.
2
Focuses on generating molecules for drug purposes.
3
Motivated to help humanity after early Covid.

26K views3626:32@GeekWireOriginal Release: 2024-09-30

Artificial intelligence and machine learning algorithms, particularly generative models like transformers and graph neural networks, can accelerate the traditional drug discovery process by generating new molecule candidates from a vast chemical space (estimated at 10^30 molecules), using the same computational approaches that power natural language processing to design molecules that are effective, safe, and manufacturable for treating diseases.

[00:00:15]We are talking about computer science and the search for new medicine.

[00:00:18]I'd like to introduce you right now to Lei Li, Language Technologies Institute, Carnegie Mellon University School of Computer Science. Lei.

[00:00:27]Thank you so much for making the time to speak with me today.

[00:00:30]Thank you for hosting me.

[00:00:32]Tell me a little bit about your job and what you're doing at CMU.

[00:00:37]I'm a computer scientist working on artificial intelligence and machine learning.

[00:00:43]The type of job I do is mostly research on developing novel algorithms using data to develop autonomous software that can develop intelligence to smart tasks like our humans.

[00:01:05]The particular problems I'm interested in, generative algorithm, generative models that can generate text, generate image, generate new content including generating new molecules for drug purpose.

[00:01:26]It must be very rewarding to be working on solving problems that can benefit humanity.

[00:01:32]Yes, exactly.

[00:01:33]The reason I started working on molecule generation and molecule design for drug purpose was I started around 2020, so I was attending a workshop, a conference in MIT that was right before Covid.

[00:01:55]So in February where at that time there was Covid in China, but in US there wasn't too many cases yet.

[00:02:03]So I was able to attend the conference in person and the conference was talking about molecule design and drug discovery and manufacturing using AI.

[00:02:14]I was writing there talking about all this and I wasn't working on anything yet, but I was listening to all those advanced research at that time and right after the conference I received the email saying, oh, someone attended the conference, got Covid, so are you affected?

[00:02:38]Fortunately enough I didn't got affected by that, but then that was a moment I realized, oh, it's very important.

[00:02:48]It's a very important topic and also AI can play a bigger role in this area that can be really beneficial for all humans.

[00:03:02]Well, I mean look, that's when mRNA leapt into the basic public space and as something that we realized was recently programmable.

[00:03:12]This is effectively what you're working on in a different capacity.

[00:03:16]So take me through this. What is it that you're doing in the space right now.

[00:03:20]Talking about biomedicine. This is a huge space.

[00:03:24]If we talk about pharmaceutical research and the path or the effort to develop effective drug takes a lot of effort from basic research, trying to identify the target for disease and then trying to identify some molecule that can cure this particular target, bite to the target and be effective in our human body.

[00:03:57]We're not being toxic to our human body and then after that we need to do a clinical trial like phase one, phase two, phase three and then manufacture it to be cost effective and deliver it to the market.

[00:04:14]So the whole process can take over 10 years with 1 billion to even 10 billion US dollars to develop one effective drug.

[00:04:29]So that's a very long duration and our goal is to accelerate certain parts, certain steps in the procedure. In particular, we want to accelerate the design and search of this molecules both before clinical trial and after clinical trial. We have to do the clinical, we have to have the real patient to try these molecules, but before that it also takes several years to find an effective drug candidate.

[00:05:10]So our goal is to use AI to accelerate this process so that instead of taking several years, we can just use our high performance computer GPUs, use the same AI algorithms that we are using for natural regulatory processing computer vision to find those molecules that can be potentially effective for certain disease target while satisfy certain chemical property we want. It should be saludable in our human body.

[00:05:48]It's not toxic to our human body.

[00:05:51]It should bide well to a particular target for the disease.

[00:05:57]It should be easy to manufacture and synthesize this molecule in the lab and in factory.

[00:06:04]This is a very interesting point that you bring up.

[00:06:08]I was a health reporter in the year 2000 and they were talking about the concerns over R and D for drug development.

[00:06:16]This is what 24 years ago potentially bankrupting our system for all of the reasons you just outlined and it's still a factor and now we have this interesting work going on in AI and ML that is able to very much more quickly, this is in its infancy, start pulling apart and finding these new molecules.

[00:06:39]But what I don't understand is how NLP works.

[00:06:42]Is it because of how it looks in sort of space and time?

[00:06:48]Why is natural language processing a part of something like this to help discover these new molecules? Can you help me understand that?

[00:06:56]That's great.

[00:06:59]So there are different kinds of drug molecules.

[00:07:04]Usually when we are talking about target-based drug design, we are referring to small molecule.

[00:07:11]So are the molecules with below let's say 50 or 100 heavy atoms other than hydrogen?

[00:07:22]So like carbon for example.

[00:07:25]So if we think about the potential space of all possible molecule with just 50 heavy atoms, so it can be, it's a very large number.

[00:07:41]It's maybe more than the number of say sand on earth.

[00:07:50]So the rough estimate is 10 to 30, that's the one. After that we have like 30 zeros.

[00:07:59]So this is a huge space and these are all possible molecules, but it doesn't mean we can all synthesize all this molecule and it doesn't mean all these molecules can be effective.

[00:08:15]We want to find from this very large space, very tiny portion, those are the promising candidates that can be potentially effective for certain disease.

[00:08:32]And the way we represent these molecules, those are small molecules. So the way we represent them, we can use two ways to represent them.

[00:08:42]First we can write them as a chemical formula, it's just like a string. So we can write it just like a English sentence, it's like letter.

[00:08:53]Instead we have a different letter system in this chemical language, chemical string.

[00:09:01]So we can use the same way we are modeling natural language modeling our English sentences.

[00:09:10]So we are modeling this English character sequences.

[00:09:14]We can use the same algorithm to model these chemical atom sequences and we can develop the same generative AI algorithm such as this variational auto encoder and the transformer which is behind ChatGPT.

[00:09:35]So we can use exactly the same large language model to model the molecules.

[00:09:44]The slight difference is for molecule, it's not only sequence.

[00:09:49]We also have graph because these atoms that they are abiding to each other, they have some bond, they form a graph, they have 3D shapes as well. They have some geometry as well.

[00:10:02]So we want to model their geometry.

[00:10:06]So for that purpose we need to use a second type of model called graph neural network.

[00:10:14]So right now many models including our own models are based on a combination of transformer which is modeling the sequence and graph neural network, which is a modeling the geometry combination of the two.

[00:10:33]We can better model this molecule sequence and their geometry then we can find from this very large space and to generate those effective molecule candidates.

[00:10:49]My mind is blown. Let me start with that.

[00:10:52]When you first realized that this was successful, that you were able to effectively discover new molecules by combining these two particular types of machine learning and AI and potentially reduce the amount of time.

[00:11:10]Was this an aha moment or was this a slow burn?

[00:11:14]You realized it was working and had potential?

[00:11:16]Was there a moment when you knew we're really onto something here?

[00:11:20]We got a lot of insight from earlier research, both from our own group and also from other teams like deep mite.

[00:11:31]They have work on the AlphaGo.

[00:11:36]It's also very similar problem.

[00:11:39]They want to search best gaming strategy in the game of Go.

[00:11:44]If you consider the possible play strategies in the game, it's 360.

[00:11:56]That's also very large space.

[00:11:59]So we are searching in the similar space in the chemical space and in their problem they can use reinforcement learning, generative AI to find effectively find the winning solution even better than the best human champion.

[00:12:17]So that's actually the moment we got some inspiration.

[00:12:23]We think it's possible or we do have a very large search space and a lot of the space even containing those molecules that never experiment, never testing the lab environment, we don't know yet.

[00:12:40]We never synthesize those, but we can possibly use this algorithm, AI algorithm from this large space to effective fender path to search for the most effective molecule.

[00:12:56]I love that you mentioned scientific collaboration effectively collaboration has been a theme across every one of these interviews that I've done and I can see how attacking the problem both from a structure of these molecules as well as a function of these molecules is what will ultimately, hopefully deliver better therapies in the end for people.

[00:13:21]How difficult is it to build these models?

[00:13:26]I feel like this particular field of study is relatively new in the big scheme of things. How challenging is it to build these?

[00:13:35]It is very challenging, the most challenging part.

[00:13:44]First we need to find the right model architecture to represent this molecule. So I mentioned a small molecule, but there are also a larger molecule we need to model as well like peptide, protein, DNA as well.

[00:14:05]Usually for drug for disease target, target is usually a protein, so we need to model this protein, the target as well.

[00:14:19]So the problem for small molecule drug design is given target, let's say for example Alzheimer's disease.

[00:14:31]So Alzheimer's is a particularly difficult disease because we don't know too much about its target.

[00:14:38]We have some suspect candidate for the targets but not confirmed yet. For those target, usually those target are protein.

[00:14:51]We want to find some small molecule.

[00:14:55]This molecule can be searched found from a molecule database.

[00:15:01]This is what this pharmaceutical company are doing.

[00:15:05]Usually they have a very large database with say 100 million molecules, but this database is only a small portion of the much larger hope possible molecule space we can search for.

[00:15:24]Then we want to find from those large spaces, small molecule that can fit into the target.

[00:15:30]This is called binding. For binding, usually there should be a pocket on the protein.

[00:15:39]This pocket has certain shape, which means our small molecule should have some certain shape as well so that it can fit into that pocket so that it can bind together to behave with the ideal desired function in our human body. And if we want to develop effective molecule using AI, we need to have data to train our AI model for new disease target like Alzheimer disease or many other cancers.

[00:16:23]We don't have much collected data from the lab environment.

[00:16:28]So possibly we have some measured molecule and protein binding affinity like how effective they are and the strength of the binding, but we only have a few with very small data or even no data. We need to train a AI model that can effectively generate those molecule. So that's a key challenge.

[00:16:55]That means we need to have the right model architecture to represent the molecule, represent the protein, represent their binding as well. We need to have sample efficient training algorithm so that we can work well with small data.

[00:17:19]We want to develop models that can effectively tell the function between the small molecule and disease target tell their chemical properties and we need to effectively synthesize those molecule to tell whether a particular molecule can be synthesized or not. If it's from the data already no database, we know how to synthesize them, but if we generate a new kind of molecule that has never been synthesized before, we also need to develop and to generate the synthesize steps so that we can synthesize it in the chemical lab.

[00:18:09]Do you think, given everything you've described across the process and that process you just described, is the new drug discovery that can take 10 years and a billion dollars from discovery to execution, do you think in your crystal ball, which is not how scientists tend to, to think that this could be a process that could almost to completion be run by AI and algorithms?

[00:18:38]Ultimately?

[00:18:39]I would say it's this process is it's always human, and AI interaction process like collaborative process.

[00:18:54]Of course we need an advanced AI exam, we need computer scientists in this procedure, but we also need a domain expert, which has those chemistry pharmaceutical expert, and also lab scientists who really have the domain knowledge about the particular disease and the target and then they can use these AI tools in their daily drug development life cycle.

[00:19:28]You mentioned the key challenge obviously in getting good data.

[00:19:32]Again a common theme across all of the AI ML discussion that we have had here. How do you get ahead of that challenge?

[00:19:41]How are you hoping to tackle that and grow these data sets to train these models?

[00:19:46]Yeah, this is a great challenge.

[00:19:49]So first we need to collaborate.

[00:19:53]So we are calling for open science like collaborative research so that we can team up with other universities like MIT or other places and also industry, these pharmaceutical companies so that we can work collaboratively to the extent that maybe they can share some data, not really for commercial purpose, but for research purpose.

[00:20:28]We can push forward the research and we can potentially collaborate with, for example, CDC and NIH to build a common consortium to, so each of this party will contribute a little bit of data to the data repositories so that collaboratively we can use the data to train a bigger model.

[00:20:55]I can see that's already happening now for some domain. For example, there's a very large protein database that's kept being collected for over several decades, protein data bank, PDB data.

[00:21:13]So we are using that to develop our protein models as well.

[00:21:18]So it's highly effective. Yes, they are, right.

[00:21:21]So this data is very important and I'm looking forward to collaborate with more, especially from industry partners as well.

[00:21:33]Lei, what's the best part of your job?

[00:21:36]Oh, the best part of my job. I think the excitement is from the moment.

[00:21:47]I think that there are two I feel really excited about.

[00:21:51]One is I can turn my algorithm into some real findings.

[00:22:00]So recently we just published two papers on protein design and we are pushing this into synthesize the proteins that found using our algorithm and test that in the lab.

[00:22:18]We're working with our partners in chemistry department of the universities, California, in Santa Barbara, and we're very close to synthesize this new protein and we are very excited about to see this algorithm actually work and can be effective in real biomedical purpose. So that's one moment in time.

[00:22:48]I'm very excited. The other moment is I'm always enthusiastic to work with students.

[00:23:00]I have two students graduating this year, and at the moment they defend their thesis and they become real PhD.

[00:23:15]I feel really proud of them.

[00:23:18]I think that's our contribution as an educator and as a researcher at university.

[00:23:25]Is there anything I'm not asking you, there's a million things, but is there something big that I should be asking you that I didn't.

[00:23:31]Under this small molecule, we are also working on protein design, which I also find very interesting and exciting.

[00:23:45]In particular, recently Google, I just published their Alpha 43 paper predicting the structures for protein and molecule complex. We are working on slightly different problem.

[00:24:04]We are working on protein design, which is given a particular function.

[00:24:10]We want to design this protein so that it'll have this desired function. For example, it can be a protein that meets this green for recent light when it's attached to a certain biomolecule target.

[00:24:31]Another example is designing enzymes that can be used in biochemical reactions so that it can accelerate the reaction and manufacture the biochemical materials. So again, we are using AI algorithm, generating AI algorithm to design these protein.

[00:24:58]Our key idea here is we are designing not just their sequence, we're designing their sequence and their structure protein structures together.

[00:25:09]So we're using this core design approach that we can design better proteins with to function.

[00:25:19]I want to talk a little bit about the, I guess, luminescence because this would be a situation where you could in real time determine if a therapy is working, correct.

[00:25:31]I mean that is when you're talking about cells lighting up, I've seen this in mouse models where they're trying to determine if course A or course B is effective in treatment.

[00:25:45]Tell me a little bit more about that.

[00:25:46]Yes, exactly.

[00:25:48]So this can be used to tell whether a certain biological pathway is activated or deactivated so that we can identify what goes wrong and where it goes wrong.

[00:26:01]We can identify the biological target and then we can design certain drug molecule for the purpose.

[00:26:10]Lei. I really appreciate your time, Lei.

[00:26:12]Thank you very much for your time and perspective.

[00:26:15]Thank you.

[00:26:16]You've been listening to Does Compute the podcast that examines how computer science is building useful stuff. That works. Thanks so much for joining us.

[00:26:24]I'm Steph Stricklen.

Up Next

Reverse Engineering a Patient Monitor's Binary Protocol for Vulnerabilities

@mattbrwn

46.5K views•2025-02-10

Building Real-Time ML Pipelines with Feature Stores and MLOps Frameworks

@ODSCAI

5.1K views•2022-02-20

Bypassing Tor Censorship: Bridges and Pluggable Transport Guide

@Coding_ForEveryone

397 views•2024-06-11

Neural Networks Explained: Math, Layers, and Learning Fundamentals

@3blue1brown

21.9M views•2017-10-05

Related Study Plans & Knowledge Roadmaps

Structured learning paths in Artificial Intelligence

View All Artificial Intelligence Topics

artificial-intelligence

Building Recommenders: Two-Tower & Vector DBs

01Foundations of Python & Machine Learning Basics

02Introduction to Recommender Systems

03Deep Learning & Representation Learning

+ 3 more learning modules

Explore Roadmap

artificial-intelligence

Latent Diffusion: PyTorch, UNet & CLIP

01Deep Learning & PyTorch Foundations

02Convolutional Networks & Autoencoders

03The Math and Mechanics of Diffusion Models

+ 3 more learning modules

Explore Roadmap

artificial-intelligence

NeRFs: 3D Reconstruction & Instant-NGP

01Math & Python Foundations for 3D Vision

02Photogrammetry & Camera Pose Estimation

03Neural Radiance Fields (NeRF) Core Concepts

+ 2 more learning modules

Explore Roadmap

artificial-intelligence

TFT Forecasting: Energy Demand & Deep Learning

01Programming & Data Science Foundations

02Introduction to Machine Learning & Time Series

03Deep Learning & Sequential Models

+ 3 more learning modules

Explore Roadmap