publications

2026

1. Can AI Conduct Autonomous Scientific Research? Case Studies on Two Real-World Tasks

   Shreyansh Agrawal, Harsh Anadkat, Kiran Athimoolam, Harsh Bhardwaj, Trishul Chowdhury, Shengtao Gao, Purva Kamat, and 5 more authors

   bioRxiv 2026

   Abs Citations

   Recent advances in artificial intelligence (AI) have prompted claims about autonomous "AI scientists," yet systematic evaluations of these capabilities remain scarce. This exploratory study investigates whether current AI frameworks can execute scientific research tasks beyond isolated demonstrations. We tested eight open-source AI frameworks (Agent Laboratory, AutoGen, BabyAGI, GPT Researcher, MOOSE-Chem2, SciAgents, SciMON, and Virtual Lab) on two tasks that aimed to reproduce research on algorithm development from recent papers in uncertainty quantification and protein interaction discovery. In our evaluation, no framework completed a full research cycle from literature understanding through computational execution to validated results and scientific paper writing. While all systems showed competence in conceptual tasks such as planning and summarization, they consistently failed at robust implementation. Every framework produced sophisticated hallucinations. Deployment proved demanding, requiring substantial debugging and technical expertise, which undermines common claims about the democratization of science with AI. Despite these limitations, the frameworks showed promise as research assistants for methodological planning and ideation under careful human supervision. Our findings suggest that the explored AI systems cannot yet autonomously conduct scientific research, but may provide real value for specific subtasks within the research workflow. We offer preliminary observations to help researchers and developers better understand the gap between advertised and actual capabilities of AI in science.Competing Interest StatementThe authors have declared no competing interest.

2025

1. 

   Multimodal out-of-distribution individual uncertainty quantification enhances binding affinity prediction for polypharmacology

   Amitesh Badkul, Li Xie, Shuo Zhang, and Lei Xie

   Nature Machine Intelligence 2025

   Abs PDF Code Poster Citations

   Polypharmacology, a single drug that targets multiple proteins, holds promise for addressing unmet medical needs. Achieving accurate, reliable and scalable predictions of protein–ligand binding affinity across multiple proteins is crucial to realizing the potential of polypharmacology. Machine learning offers a powerful tool for multitarget binding affinity prediction. However, three major challenges remain: generalizing predictions to out-of-distribution compounds that are structurally different from those in the training data; quantifying the uncertainty of predictions in out-of-distribution scenarios where the assumption underlying existing methods does not hold; and scaling to billions of compounds, which remains unattainable for current structure-based methods. Here, to overcome these challenges, we propose a model-agnostic anomaly detection-based individual uncertainty quantification method: embedding Mahalanobis Outlier Scoring and Anomaly Identification via Clustering (eMOSAIC). eMOSAIC features the divergence between the multimodal representations of known cases and unseen instances and quantifies individual prediction uncertainty on a compound-by-compound basis. We integrate eMOSAIC with a multimodal deep neural network for multitarget ligand binding affinity predictions, leveraging a structure-informed large protein language model. Comprehensive validation in out-of-distribution settings demonstrates that eMOSAIC significantly outperforms state-of-the-art sequence-based and structure-based methods as well as existing uncertainty quantification approaches. These findings underscore eMOSAIC’s potential to advance real-world polypharmacology and other applications that require robust predictions and scalable solutions.

2. 

   Adaptive Individual Uncertainty under Out-Of-Distribution Shift with Expert-Routed Conformal Prediction

   Amitesh Badkul, and Lei Xie

   2025

   Abs PDF Citations

   Reliable, informative, and individual uncertainty quantification (UQ) remains missing in current ML community. This hinders the effective application of AI/ML to risk-sensitive domains. Most methods either fail to provide coverage on new data, inflate intervals so broadly that they are not actionable, or assign uncertainties that do not track actual error, especially under a distribution shift. In high-stakes drug discovery, protein-ligand affinity (PLI) prediction is especially challenging as assay noise is heterogeneous, chemical space is imbalanced and large, and practical evaluations routinely involve distribution shift. In this work, we introduce a novel uncertainty quantification method, Trustworthy Expert Split-conformal with Scaled Estimation for Efficient Reliable Adaptive intervals (TESSERA), that provides per-sample uncertainty with reliable coverage guarantee, informative and adaptive prediction interval widths that track the absolute error. We evaluate on protein-ligand binding affinity prediction under both independent and identically distributed (i.i.d.) and scaffold-based out-of-distribution (OOD) splits, comparing against strong UQ baselines. TESSERA attains near-nominal coverage and the best coverage-width trade-off as measured by the Coverage-Width Criterion (CWC), while maintaining competitive adaptivity (lowest Area Under the Sparsification Error (AUSE)). Size-Stratified Coverage (SSC) further confirms that intervals are right-sized, indicating width increases when data are scarce or noisy, and remain tight when predictions are reliable. By unifying Mixture of Expert (MoE) diversity with conformal calibration, TESSERA delivers trustworthy, tight, and adaptive uncertainties that are well-suited to selective prediction and downstream decision-making in the drug-discovery pipeline and other applications.

3. 

   Wavelet Transform and Machine Learning-Driven Multi-Class Classification of Chest X-Ray Images for COVID-19 Diagnosis

   Amitesh Badkul, Inturi Vamsi, and Radhika Sudha

   International Conference on Integrating Cognitive Science and Computational Intelligence 2025

   Abs PDF Citations

   Chest radiography is a fast, low cost, and widely available first line test for suspected pneumonia and other respiratory infections, which makes automation clinically valuable where CT and subspecialty radiologists are limited. Yet adoption of automated CXR analysis using machine learning has been slowed by the memory and compute demands of large deep convolutional networks, their training time, and the need for GPU infrastructure, which limit deployment at the edge and in resource constrained settings. In acute care, separating COVID-19, bacterial pneumonia, and viral pneumonia is necessary for isolation, preventive care, so the classification is valuable and directly actionable by hospitals and clinics. We present a lightweight pipeline for multiclass CXR classification that applies a single level discrete wavelet transform and computes 11 statistical and texture descriptors per subband, producing a 44 dimensional feature vector for tabular classifiers. We evaluate CXRs stratified across four classes (healthy, COVID-19, bacterial pneumonia, viral pneumonia). Random Forest and XGBoost reach 96.83% and 97.76% test accuracy, respectively, outperforming transfer learned DCNN baselines on the same data. Ablations show that features pooled across subbands outperform any single subband and that texture descriptors including entropy, contrast, homogeneity, dissimilarity, and correlation carry most of the signal. Among wavelet bases, single level Bior6.8 performs best, while deeper decompositions reduce accuracy at this data scale. Because computation is limited to a grayscale wavelet transform and tabular learning, the method runs on commodity CPUs, has a small memory footprint, and is simple to reproduce, with strong per class performance for COVID-19, bacterial pneumonia, and viral pneu- monia that supports further treatment.

2024

1. JME

   Comparative study of DCNN and image processing based classification of chest X-rays for identification of COVID-19 patients using fine-tuning

   Amitesh Badkul, Inturi Vamsi, and Radhika Sudha

   Journal of Medical Engineering & Technology 2024

   Abs Citations

   The conventional detection of COVID-19 by evaluating the CT scan images is tiresome, often experiences high inter-observer variability and uncertainty issues. This work proposes the automatic detection and classification of COVID-19 by analysing the chest X-ray images (CXR) with the deep convolutional neural network (DCNN) models through a fine-tuning and pre-training approach. CXR images pertaining to four health scenarios, namely, healthy, COVID-19, bacterial pneumonia and viral pneumonia, are considered and subjected to data augmentation. Two types of input datasets are prepared; in which dataset I contains the original image dataset categorised under four classes, whereas the original CXR images are subjected to image pre-processing via Contrast Limited Adaptive Histogram Equalisation (CLAHE) algorithm and Blackhat Morphological Operation (BMO) for devising the input dataset II. Both datasets are supplied as input to various DCNN models such as DenseNet, MobileNet, ResNet, VGG16, and Xception for achieving multi-class classification. It is observed that the classification accuracies are improved, and the classification errors are reduced with the image pre-processing. Overall, the VGG16 model resulted in better classification accuracies and reduced classification errors while accomplishing multi-class classification. Thus, the proposed work would assist the clinical diagnosis, and reduce the workload of the front-line healthcare workforce and medical professionals.

2. MGV

   A comparative study of DeepLabCut and other open-source pupillometry data analysis algorithms – Which to choose?

   Amitesh Badkul, Sonakshi Mishra, and Srinivasa Prasad Kommajosyula

   Machine Graphics and Vision Dec 2024

   Abs PDF Citations

   <p>Pupillometry measures pupil size, and several open-source algorithms are available to analyse pupillometry data. However, only a few studies compared these algorithms’ accuracy and computational resources. This study aims to compare the accuracy of computer vision-based algorithms (Swirski, Starburst, PuRe, ElSe, ExCuSe algorithms) and the machine learning algorithm, DeepLabCut, to the double-blinded human examiners (gold-standard). Training of DeepLabCut with different architectures and a variable number of markers (2-9 markers) was done on an open-source dataset. The duration of training was statistically longer for the ResNet152 model compared to the MobileNet model. The pupil diameters in computer vision-based software such as PuRe, Starburst, and Swirski were statistically different from human measurements. MobileNet 2 and 3 marker models were the closest to the human measurements. In conclusion, this work highlights the efficiency of lower marker models based on MobileNet architecture in DeepLabCut, which consumes fewer computational resources and is more accurate.</p>

2023

1. 

   End-to-end sequence-structure-function meta-learning predicts genome-wide chemical-protein interactions for dark proteins

   Tian Cai, Li Xie, Shuo Zhang, Muge Chen, Di He, Amitesh Badkul, Yang Liu, and 4 more authors

   PLOS Computational Biology Dec 2023

   Abs PDF Code Citations

   Discovering chemical-protein interactions for millions of chemicals across the entire human and pathogen genomes is instrumental for chemical genomics, protein function prediction, drug discovery, and other applications. However, more than 90% of gene families remain dark, i.e., their small molecular ligands are undiscovered due to experimental limitations and human biases. Existing computational approaches typically fail when the unlabeled dark protein of interest differs from those with known ligands or structures. To address this challenge, we developed a deep learning framework PortalCG. PortalCG consists of four novel components: (i) a 3-dimensional ligand binding site enhanced sequence pre-training strategy to represent the whole universe of protein sequences in recognition of evolutionary linkage of ligand binding sites across gene families, (ii) an end-to-end pretraining-fine-tuning strategy to simulate the folding process of protein-ligand interactions and reduce the impact of inaccuracy of predicted structures on function predictions under a sequence-structure-function paradigm, (iii) a new out-of-cluster meta-learning algorithm that extracts and accumulates information learned from predicting ligands of distinct gene families (meta-data) and applies the meta-data to a dark gene family, and (iv) stress model selection that uses different gene families in the test data from those in the training and development data sets to facilitate model deployment in a real-world scenario. In extensive and rigorous benchmark experiments, PortalCG considerably outperformed state-of-the-art techniques of machine learning and protein-ligand docking when applied to dark gene families, and demonstrated its generalization power for off-target predictions and compound screenings under out-of-distribution (OOD) scenarios. Furthermore, in an external validation for the multi-target compound screening, the performance of PortalCG surpassed the human design. Our results also suggested that a differentiable sequence-structure-function deep learning framework where protein structure information serve as an intermediate layer could be superior to conventional methodology where the use of predicted protein structures for predicting protein functions from sequences. We applied PortalCG to two case studies to exemplify its potential in drug discovery: designing selective dual-antagonists of Dopamine receptors for the treatment of Opioid Use Disorder, and illuminating the undruggable human genome for targeting diseases that do not have effective and safe therapeutics. Our results suggested that PortalCG is a viable solution to the OOD problem in exploring the understudied protein functional space.

2. NeurIPS 2023

   TrustAffinity: accurate, reliable and scalable out-of-distribution protein-ligand binding affinity prediction using trustworthy deep learning

   Amitesh Badkul, Li Xie, Shuo Zhang, and Lei Xie

   NeurIPS 2023 Workshop on New Frontiers of AI for Drug Discovery and Development & AAAI 2024 Workshop on LLMs4Bio Dec 2023

   Abs PDF Poster Citations

   Accurate, reliable and scalable predictions of protein-ligand binding affinity have a great potential to accelerate drug discovery. Despite considerable efforts, three challenges remain: out-of-distribution (OOD) generalizations for understudied proteins or compounds from unlabeled protein families or chemical scaffolds, uncertainty quantification of individual predictions, and scalability to billions of compounds. We propose a sequence-based deep learning framework, TrustAffinity, to address aforementioned challenges. TrustAffinity synthesizes a structure-informed protein language model, efficient uncertainty quantification based on residue-estimation and novel uncertainty regularized optimization. We extensively validate TrustAffinity in multiple OOD settings. TrustAffinity significantly outperforms state-of-the-art computational methods by a large margin. It achieves a Pearson’s correlation between predicted and actual binding affinities above 0.9 with a high confidence and at least three orders of magnitude of faster than protein-ligand docking, highlighting its potential in real-world drug discovery. We further demonstrate TrustAffinity’s practicality through an Opioid Use Disorder lead discovery case study.

2022

1. SURI

   RNN-driven Approaches to Self-healing Compound Synthesis

   Amitesh Badkul, and Ashif Iquebal

   Dec 2022

   Abs PDF Citations

   3D printing technology has revolutionized manufacturing processes, offering en- hanced precision andversatilityin product design. However, the materials commonly used in this domain often exhibit brittleness, leading to concerns about their durabil- ity. The frequent and irreversible damage to these materials necessitates a solution to enhance their longevity and reduce maintenance. Self-healing materials, characterized by their ability to recover from damage au- tonomously, present a promising avenue to address this challenge. Hydrogen bond- ing, a fundamental atomic interaction, plays a pivotal role in facilitating the self- healing properties of materials. Yet, systematically exploring the chemical space to identify compounds with optimal hydrogen bonding for self-healing remains a com- plex task. This research aims to employ Recurrent Neural Networks-based (RNNs) algorithms to navigate this vast chemical space, striving to design compounds that harness the potential of hydrogen bonding for enhanced self-healing properties.