Transparent medical image AI via an image-text foundation model grounded in medical literature.

Building trustworthy and transparent image-based medical artificial intelligence (AI) systems requires the ability to interrogate data and models at all stages of the development pipeline, from training models to post-deployment monitoring. Ideally, the data and associated AI systems could be described using terms already familiar to physicians, but this requires medical datasets densely annotated with semantically meaningful concepts. In the present study, we present a foundation model approach, named MONET (medical concept retriever), which learns how to connect medical images with text and densely scores images on concept presence to enable important tasks in medical AI development and deployment such as data auditing, model auditing and model interpretation. Dermatology provides a demanding use case for the versatility of MONET, due to the heterogeneity in diseases, skin tones and imaging modalities. We trained MONET based on 105,550 dermatological images paired with natural language descriptions from a large collection of medical literature. MONET can accurately annotate concepts across dermatology images as verified by board-certified dermatologists, competitively with supervised models built on previously concept-annotated dermatology datasets of clinical images. We demonstrate how MONET enables AI transparency across the entire AI system development pipeline, from building inherently interpretable models to dataset and model auditing, including a case study dissecting the results of an AI clinical trial.

Auditing the inference processes of medical-image classifiers by leveraging generative AI and the expertise of physicians.

The inferences of most machine-learning models powering medical artificial intelligence are difficult to interpret. Here we report a general framework for model auditing that combines insights from medical experts with a highly expressive form of explainable artificial intelligence. Specifically, we leveraged the expertise of dermatologists for the clinical task of differentiating melanomas from melanoma 'lookalikes' on the basis of dermoscopic and clinical images of the skin, and the power of generative models to render 'counterfactual' images to understand the 'reasoning' processes of five medical-image classifiers. By altering image attributes to produce analogous images that elicit a different prediction by the classifiers, and by asking physicians to identify medically meaningful features in the images, the counterfactual images revealed that the classifiers rely both on features used by human dermatologists, such as lesional pigmentation patterns, and on undesirable features, such as background skin texture and colour balance. The framework can be applied to any specialized medical domain to make the powerful inference processes of machine-learning models medically understandable.

Fostering transparent medical image AI via an image-text foundation model grounded in medical literature.

Building trustworthy and transparent image-based medical AI systems requires the ability to interrogate data and models at all stages of the development pipeline: from training models to post-deployment monitoring. Ideally, the data and associated AI systems could be described using terms already familiar to physicians, but this requires medical datasets densely annotated with semantically meaningful concepts. Here, we present a foundation model approach, named MONET (Medical cONcept rETriever), which learns how to connect medical images with text and generates dense concept annotations to enable tasks in AI transparency from model auditing to model interpretation. Dermatology provides a demanding use case for the versatility of MONET, due to the heterogeneity in diseases, skin tones, and imaging modalities. We trained MONET on the basis of 105,550 dermatological images paired with natural language descriptions from a large collection of medical literature. MONET can accurately annotate concepts across dermatology images as verified by board-certified dermatologists, outperforming supervised models built on previously concept-annotated dermatology datasets. We demonstrate how MONET enables AI transparency across the entire AI development pipeline from dataset auditing to model auditing to building inherently interpretable models.

Dissection of medical AI reasoning processes via physician and generative-AI collaboration.

Despite the proliferation and clinical deployment of artificial intelligence (AI)-based medical software devices, most remain black boxes that are uninterpretable to key stakeholders including patients, physicians, and even the developers of the devices. Here, we present a general model auditing framework that combines insights from medical experts with a highly expressive form of explainable AI that leverages generative models, to understand the reasoning processes of AI devices. We then apply this framework to generate the first thorough, medically interpretable picture of the reasoning processes of machine-learning-based medical image AI. In our synergistic framework, a generative model first renders "counterfactual" medical images, which in essence visually represent the reasoning process of a medical AI device, and then physicians translate these counterfactual images to medically meaningful features. As our use case, we audit five high-profile AI devices in dermatology, an area of particular interest since dermatology AI devices are beginning to achieve deployment globally. We reveal how dermatology AI devices rely both on features used by human dermatologists, such as lesional pigmentation patterns, as well as multiple, previously unreported, potentially undesirable features, such as background skin texture and image color balance. Our study also sets a precedent for the rigorous application of explainable AI to understand AI in any specialized domain and provides a means for practitioners, clinicians, and regulators to uncloak AI's powerful but previously enigmatic reasoning processes in a medically understandable way.

A Suite of Tutorials for the WESTPA 2.0 Rare-Events Sampling Software [Article v2.0].

The weighted ensemble (WE) strategy has been demonstrated to be highly efficient in generating pathways and rate constants for rare events such as protein folding and protein binding using atomistic molecular dynamics simulations. Here we present two sets of tutorials instructing users in the best practices for preparing, carrying out, and analyzing WE simulations for various applications using the WESTPA software. The first set of more basic tutorials describes a range of simulation types, from a molecular association process in explicit solvent to more complex processes such as host-guest association, peptide conformational sampling, and protein folding. The second set ecompasses six advanced tutorials instructing users in the best practices of using key new features and plugins/extensions of the WESTPA 2.0 software package, which consists of major upgrades for larger systems and/or slower processes. The advanced tutorials demonstrate the use of the following key features: (i) a generalized resampler module for the creation of "binless" schemes, (ii) a minimal adaptive binning scheme for more efficient surmounting of free energy barriers, (iii) streamlined handling of large simulation datasets using an HDF5 framework, (iv) two different schemes for more efficient rate-constant estimation, (v) a Python API for simplified analysis of WE simulations, and (vi) plugins/extensions for Markovian Weighted Ensemble Milestoning and WE rule-based modeling for systems biology models. Applications of the advanced tutorials include atomistic and non-spatial models, and consist of complex processes such as protein folding and the membrane permeability of a drug-like molecule. Users are expected to already have significant experience with running conventional molecular dynamics or systems biology simulations.

WESTPA 2.0: High-Performance Upgrades for Weighted Ensemble Simulations and Analysis of Longer-Timescale Applications.

The weighted ensemble (WE) family of methods is one of several statistical mechanics-based path sampling strategies that can provide estimates of key observables (rate constants and pathways) using a fraction of the time required by direct simulation methods such as molecular dynamics or discrete-state stochastic algorithms. WE methods oversee numerous parallel trajectories using intermittent overhead operations at fixed time intervals, enabling facile interoperability with any dynamics engine. Here, we report on the major upgrades to the WESTPA software package, an open-source, high-performance framework that implements both basic and recently developed WE methods. These upgrades offer substantial improvements over traditional WE methods. The key features of the new WESTPA 2.0 software enhance the efficiency and ease of use: an adaptive binning scheme for more efficient surmounting of large free energy barriers, streamlined handling of large simulation data sets, exponentially improved analysis of kinetics, and developer-friendly tools for creating new WE methods, including a Python API and resampler module for implementing both binned and "binless" WE strategies.

The RED scheme: Rate-constant estimation from pre-steady state weighted ensemble simulations.

We present the Rate from Event Durations (RED) scheme, a new scheme that more efficiently calculates rate constants using the weighted ensemble path sampling strategy. This scheme enables rate-constant estimation from shorter trajectories by incorporating the probability distribution of event durations, or barrier-crossing times, from a simulation. We have applied the RED scheme to weighted ensemble simulations of a variety of rare-event processes that range in complexity: residue-level simulations of protein conformational switching, atomistic simulations of Na+/Cl- association in explicit solvent, and atomistic simulations of protein-protein association in explicit solvent. Rate constants were estimated with up to 50% greater efficiency than the original weighted ensemble scheme. Importantly, our scheme accounts for the systematic error that results from statistical bias toward the observation of events with short durations and reweights the event duration distribution accordingly. The RED scheme is relevant to any simulation strategy that involves unbiased trajectories of similar length to the most probable event duration, including weighted ensemble, milestoning, and standard simulations as well as the construction of Markov state models.

AI for radiographic COVID-19 detection selects shortcuts over signal.

Artificial intelligence (AI) researchers and radiologists have recently reported AI systems that accurately detect COVID-19 in chest radiographs. However, the robustness of these systems remains unclear. Using state-of-the-art techniques in explainable AI, we demonstrate that recent deep learning systems to detect COVID-19 from chest radiographs rely on confounding factors rather than medical pathology, creating an alarming situation in which the systems appear accurate, but fail when tested in new hospitals. We observe that the approach to obtain training data for these AI systems introduces a nearly ideal scenario for AI to learn these spurious "shortcuts." Because this approach to data collection has also been used to obtain training data for detection of COVID-19 in computed tomography scans and for medical imaging tasks related to other diseases, our study reveals a far-reaching problem in medical imaging AI. In addition, we show that evaluation of a model on external data is insufficient to ensure AI systems rely on medically relevant pathology, since the undesired "shortcuts" learned by AI systems may not impair performance in new hospitals. These findings demonstrate that explainable AI should be seen as a prerequisite to clinical deployment of ML healthcare models.

A twist in the road less traveled: The AMBER ff15ipq-m force field for protein mimetics.

We present a new force field, AMBER ff15ipq-m, for simulations of protein mimetics in applications from therapeutics to biomaterials. This force field is an expansion of the AMBER ff15ipq force field that was developed for canonical proteins and enables the modeling of four classes of artificial backbone units that are commonly used alongside natural α residues in blended or "heterogeneous" backbones: chirality-reversed D-α-residues, the Cα-methylated α-residue Aib, homologated ß-residues (ß3) bearing proteinogenic side chains, and two cyclic ß residues (ßcyc; APC and ACPC). The ff15ipq-m force field includes 472 unique atomic charges and 148 unique torsion terms. Consistent with the AMBER IPolQ lineage of force fields, the charges were derived using the Implicitly Polarized Charge (IPolQ) scheme in the presence of explicit solvent. To our knowledge, no general force field reported to date models the combination of artificial building blocks examined here. In addition, we have derived Karplus coefficients for the calculation of backbone amide J-coupling constants for ß3Ala and ACPC ß residues. The AMBER ff15ipq-m force field reproduces experimentally observed J-coupling constants in simple tetrapeptides and maintains the expected conformational propensities in reported structures of proteins/peptides containing the artificial building blocks of interest-all on the µs timescale. These encouraging results demonstrate the power and robustness of the IPolQ lineage of force fields in modeling the structure and dynamics of natural proteins as well as mimetics with protein-inspired artificial backbones in atomic detail.

A Suite of Tutorials for the WESTPA Rare-Events Sampling Software [Article v1.0].

The weighted ensemble (WE) strategy has been demonstrated to be highly efficient in generating pathways and rate constants for rare events such as protein folding and protein binding using atomistic molecular dynamics simulations. Here we present five tutorials instructing users in the best practices for preparing, carrying out, and analyzing WE simulations for various applications using the WESTPA software. Users are expected to already have significant experience with running standard molecular dynamics simulations using the underlying dynamics engine of interest (e.g. Amber, Gromacs, OpenMM). The tutorials range from a molecular association process in explicit solvent to more complex processes such as host-guest association, peptide conformational sampling, and protein folding.

Large enhancement of response times of a protein conformational switch by computational design.

The design of protein conformational switches-or proteins that change conformations in response to a signal such as ligand binding-has great potential for developing novel biosensors, diagnostic tools, and therapeutic agents. Among the defining properties of such switches, the response time has been the most challenging to optimize. Here we apply a computational design strategy in synergistic combination with biophysical experiments to rationally improve the response time of an engineered protein-based Ca2+-sensor in which the switching process occurs via mutually exclusive folding of two alternate frames. Notably, our strategy identifies mutations that increase switching rates by as much as 32-fold, achieving response times on the order of fast physiological Ca2+ fluctuations. Our computational design strategy is general and may aid in optimizing the kinetics of other protein conformational switches.