arXiv subjects

Mohit Pandey

Publications and source records attributed to Mohit Pandey.

11 recordsLinked to original sources

arXiv2025

Why Pool When You Can Flow? Active Learning with GFlowNets

The scalability of pool-based active learning is limited by the computational cost of evaluating large unlabeled datasets, a challenge that is particularly acute in virtual screening for drug discovery. While active learning strategies such as Bayesian Active Learning by Disagreement (BALD) prioritize informative samples, it remains computationally intensive when scaled to libraries containing billions samples. In this work, we introduce BALD-GFlowNet, a generative active learning framework that circumvents this issue. Our method leverages Generative Flow Networks (GFlowNets) to directly sample objects in proportion to the BALD reward. By replacing traditional pool-based acquisition with generative sampling, BALD-GFlowNet achieves scalability that is independent of the size of the unlabeled pool. In our virtual screening experiment, we show that BALD-GFlowNet achieves a performance comparable to that of standard BALD baseline while generating more structurally diverse molecules, offering a promising direction for efficient and scalable molecular discovery.

cs.LG↗

arXiv2025

Biomarker Integration and Biosensor Technologies Enabling AI-Driven Insights into Biological Aging

As the global population continues to age, there is an increasing demand for ways to accurately quantify the biological processes underlying aging. Biological age, unlike chronological age, reflects an individual's physiological state, offering a more accurate measure of health-span and age-related decline. Aging is a complex, multisystem process involving molecular, cellular, and environmental factors and can be quantified using various biophysical and biochemical markers. This review focuses on four key biochemical markers that have recently been identified by experts as important outcome measures in longevity-promoting interventions: C-Reactive Protein, Insulin like Growth Factor-1, Interleukin-6, and Growth Differentiation Factor-15. With the use of Artificial Intelligence, the analysis and integration of these biomarkers can be significantly enhanced, enabling the identification of complex binding patterns and improving predictive accuracy for biological age estimation and age-related disease risk stratification. Artificial intelligence-driven methods including machine learning, deep learning, and multimodal data integration, facilitate the interpretation of high dimensional datasets and support the development of widely accessible, data-informed tools for health monitoring and disease risk assessment. This paves the way for a future medical system, enabling more personalized and accessible care, offering deeper, data-driven insights into individual health trajectories, risk profiles, and treatment response. The review additionally highlights the key challenges, and future directions for the implementation of artificial intelligence-driven methods in precision aging frameworks.

q-bio.QM↗

arXiv2025

Pretraining Generative Flow Networks with Inexpensive Rewards for Molecular Graph Generation

Generative Flow Networks (GFlowNets) have recently emerged as a suitable framework for generating diverse and high-quality molecular structures by learning from rewards treated as unnormalized distributions. Previous works in this framework often restrict exploration by using predefined molecular fragments as building blocks, limiting the chemical space that can be accessed. In this work, we introduce Atomic GFlowNets (A-GFNs), a foundational generative model leveraging individual atoms as building blocks to explore drug-like chemical space more comprehensively. We propose an unsupervised pre-training approach using drug-like molecule datasets, which teaches A-GFNs about inexpensive yet informative molecular descriptors such as drug-likeliness, topological polar surface area, and synthetic accessibility scores. These properties serve as proxy rewards, guiding A-GFNs towards regions of chemical space that exhibit desirable pharmacological properties. We further implement a goal-conditioned finetuning process, which adapts A-GFNs to optimize for specific target properties. In this work, we pretrain A-GFN on a subset of ZINC dataset, and by employing robust evaluation metrics we show the effectiveness of our approach when compared to other relevant baseline methods for a wide range of drug design tasks. The code is accessible at https://github.com/diamondspark/AGFN.

cs.LG↗

arXiv2024

GFlowNet Pretraining with Inexpensive Rewards

Generative Flow Networks (GFlowNets), a class of generative models have recently emerged as a suitable framework for generating diverse and high-quality molecular structures by learning from unnormalized reward distributions. Previous works in this direction often restrict exploration by using predefined molecular fragments as building blocks, limiting the chemical space that can be accessed. In this work, we introduce Atomic GFlowNets (A-GFNs), a foundational generative model leveraging individual atoms as building blocks to explore drug-like chemical space more comprehensively. We propose an unsupervised pre-training approach using offline drug-like molecule datasets, which conditions A-GFNs on inexpensive yet informative molecular descriptors such as drug-likeliness, topological polar surface area, and synthetic accessibility scores. These properties serve as proxy rewards, guiding A-GFNs towards regions of chemical space that exhibit desirable pharmacological properties. We further our method by implementing a goal-conditioned fine-tuning process, which adapts A-GFNs to optimize for specific target properties. In this work, we pretrain A-GFN on the ZINC15 offline dataset and employ robust evaluation metrics to show the effectiveness of our approach when compared to other relevant baseline methods in drug design.

cs.LG↗

arXiv2023

TacoGFN: Target-conditioned GFlowNet for Structure-based Drug Design

Searching the vast chemical space for drug-like molecules that bind with a protein pocket is a challenging task in drug discovery. Recently, structure-based generative models have been introduced which promise to be more efficient by learning to generate molecules for any given protein structure. However, since they learn the distribution of a limited protein-ligand complex dataset, structure-based methods do not yet outperform optimization-based methods that generate binding molecules for just one pocket. To overcome limitations on data while leveraging learning across protein targets, we choose to model the reward distribution conditioned on pocket structure, instead of the training data distribution. We design TacoGFN, a novel GFlowNet-based approach for structure-based drug design, which can generate molecules conditioned on any protein pocket structure with probabilities proportional to its affinity and property rewards. In the generative setting for CrossDocked2020 benchmark, TacoGFN attains a state-of-the-art success rate of $56.0\%$ and $-8.44$ kcal/mol in median Vina Dock score while improving the generation time by multiple orders of magnitude. Fine-tuning TacoGFN further improves the median Vina Dock score to $-10.93$ kcal/mol and the success rate to $88.8\%$, outperforming all optimization-based methods.

cs.LG↗

arXiv2022

Multibody molecular docking on a quantum annealer

Molecular docking, which aims to find the most stable interacting configuration of a set of molecules, is of critical importance to drug discovery. Although a considerable number of classical algorithms have been developed to carry out molecular docking, most focus on the limiting case of docking two molecules. Since the number of possible configurations of N molecules is exponential in N, those exceptions which permit docking of more than two molecules scale poorly, requiring exponential resources to find high-quality solutions. Here, we introduce a one-hot encoded quadratic unconstrained binary optimization formulation (QUBO) of the multibody molecular docking problem, which is suitable for solution by quantum annealer. Our approach involves a classical pre-computation of pairwise interactions, which scales only quadratically in the number of bodies while permitting well-vetted scoring functions like the Rosetta REF2015 energy function to be used. In a second step, we use the quantum annealer to sample low-energy docked configurations efficiently, considering all possible docked configurations simultaneously through quantum superposition. We show that we are able to minimize the time needed to find diverse low-energy docked configurations by tuning the strength of the penalty used to enforce the one-hot encoding, demonstrating a 3-4 fold improvement in solution quality and diversity over performance achieved with conventional penalty strengths. By mapping the configurational search to a form compatible with current- and future-generation quantum annealers, this work provides an alternative means of solving multibody docking problems that may prove to have performance advantages for large problems, potentially circumventing the exponential scaling of classical approaches and permitting a much more efficient solution to a problem central to drug discovery and validation pipelines.

q-bio.BM↗

arXiv2020

Adiabatic eigenstate deformations as a sensitive probe for quantum chaos

In the past decades, it was recognized that quantum chaos, which is essential for the emergence of statistical mechanics and thermodynamics, manifests itself in the effective description of the eigenstates of chaotic Hamiltonians through random matrix ensembles and the eigenstate thermalization hypothesis. Standard measures of chaos in quantum many-body systems are level statistics and the spectral form factor. In this work, we show that the norm of the adiabatic gauge potential, the generator of adiabatic deformations between eigenstates, serves as a much more sensitive measure of quantum chaos. We are able to detect transitions from non-ergodic to ergodic behavior at perturbation strengths orders of magnitude smaller than those required for standard measures. Using this alternative probe in two generic classes of spin chains, we show that the chaotic threshold decreases exponentially with system size and that one can immediately detect integrability-breaking (chaotic) perturbations by analyzing infinitesimal perturbations even at the integrable point. In some cases, small integrability-breaking is shown to lead to anomalously slow relaxation of the system, exponentially long in system size.

quant-ph↗

arXiv2020

Persistent dark states in anisotropic central spin models

Long-lived dark states, in which an experimentally accessible qubit is not in thermal equilibrium with a surrounding spin bath, are pervasive in solid-state systems. We explain the ubiquity of dark states in a large class of inhomogenous central spin models using the proximity to integrable lines with exact dark eigenstates. At numerically accessible sizes, dark states persist as eigenstates at large deviations from integrability, and the qubit retains memory of its initial polarization at long times. Although the eigenstates of the system are chaotic, exhibiting exponential sensitivity to small perturbations, they do not satisfy the eigenstate thermalization hypothesis. Rather, we predict long relaxation times that increase exponentially with system size. We propose that this intermediate chaotic but non-ergodic regime characterizes mesoscopic quantum dot and diamond defect systems, as we see no numerical tendency towards conventional thermalization with a finite relaxation time.

cond-mat.str-el↗

arXiv2019

Floquet-engineering counterdiabatic protocols in quantum many-body systems

Counterdiabatic (CD) driving presents a way of generating adiabatic dynamics at arbitrary pace, where excitations due to non-adiabaticity are exactly compensated by adding an auxiliary driving term to the Hamiltonian. While this CD term is theoretically known and given by the adiabatic gauge potential, obtaining and implementing this potential in many-body systems is a formidable task, requiring knowledge of the spectral properties of the instantaneous Hamiltonians and control of highly nonlocal multibody interactions. We show how an approximate gauge potential can be systematically built up as a series of nested commutators, remaining well-defined in the thermodynamic limit. Furthermore, the resulting CD driving protocols can be realized up to arbitrary order without leaving the available control space using tools from periodically-driven (Floquet) systems. This is illustrated on few- and many-body quantum systems, where the resulting Floquet protocols significantly suppress dissipation and provide a drastic increase in fidelity.

quant-ph↗

arXiv2018

Floquet-engineered quantum state manipulation in a noisy qubit

Adiabatic evolution is a common strategy for manipulating quantum states and has been employed in diverse fields such as quantum simulation, computation and annealing. However, adiabatic evolution is inherently slow and therefore susceptible to decoherence. Existing methods for speeding up adiabatic evolution require complex many-body operators or are difficult to construct for multi-level systems. Using the tools of Floquet engineering, we design a scheme for high-fidelity quantum state manipulation, utilizing only the interactions available in the original Hamiltonian. We apply this approach to a qubit and experimentally demonstrate its performance with the electronic spin of a Nitrogen-vacancy center in diamond. Our Floquet-engineered protocol achieves state preparation fidelity of $0.994 \pm 0.004$, on the same level as the conventional fast-forward protocol, but is more robust to external noise acting on the qubit. Floquet engineering provides a powerful platform for high-fidelity quantum state manipulation in complex and noisy quantum systems.

quant-ph↗

arXiv2015

Localization of weakly interacting Bose gas in quasiperiodic potential

We study the localization properties of weakly interacting Bose gas in a quasiperiodic potential commonly known as Aubry-André model. Effect of interaction on localization is investigated by computing the `superfluid fraction' and `inverse participation ratio'. For interacting Bosons the inverse participation ratio increases very slowly after the localization transition due to `multisite localization' of the wave function. We also study the localization in Aubry-André model using an alternative approach of classical dynamical map, where the localization is manifested by chaotic classical dynamics. For weakly interacting Bose gas, Bogoliubov quasiparticle spectrum and condensate fraction are calculated in order to study the loss of coherence with increasing disorder strength. Finally we discuss the effect of trapping potential on localization of matter wave.

cond-mat.quant-gas↗