Utility of language model and physics-based approaches in modifying MHC Class-I immune-visibility for the design of vaccines and therapeutics
FigureProtein therapeutics promise to revolutionize medicine, with an arsenal of applications tha... more FigureProtein therapeutics promise to revolutionize medicine, with an arsenal of applications that include disrupting protein interactions, acting as potent vaccines, and replacing genetically deficient proteins. Therapeutics must avoid triggering unwanted immune responses towards the therapeutic protein or viral vector proteins. In contrast, vaccines require a robust immune reaction targeting a broad range of pathogen variants. Therefore, computational methods modifying proteins’ immune visibility, while maintaining functionality, are needed. This paper focuses on visibility to cytotoxic T-lymphocytes, which use the MHC Class I pathway to detect destruction targets.To explore the limits of modifying MHC-I immune visibility within the distribution of naturally occurring sequences, we developed a novel machine learning technique,CAPE-XVAE, that combines a language model with reinforcement learning to modify a protein’s immune visibility. Our results show thatCAPE-XVAEeffectively modi...
Figure 4 from The Immunopeptidome from a Genomic Perspective: Establishing the Noncanonical Landscape of MHC Class I–Associated Peptides
Comparison of COD-dipp ncMAPs with other studies. Because the COD-dipp ncMAPs are restricted to t... more Comparison of COD-dipp ncMAPs with other studies. Because the COD-dipp ncMAPs are restricted to the 3-frame translation (3FT) of protein-coding genes, sequences from the literature were aligned to the same 3FT database for comparison purposes. The intersection is based on genomic coordinates to deal with sequences that partially match (i.e., longer, shorter, or partially overlapping). Because the Venn is generated by overlapping genomic coordinates of the ncMAPs, the original counts for each study are listed from left to right (i.e., on the right-hand side of panel C, the notation 29/41 refers to 29 instances for Chong and colleagues 2020 and 41 for COD-dipp). A, Comparison with peptide-PRISM published ncMAPs at a 10% FDR. COD-dipp ncMAPs were restricted to 3 studies in common with Erhard and colleagues 2020. B, Comparison with peptide-PRISM published ncMAPs at a 1% FDR. COD-dipp ncMAPs were restricted to 3 studies in common with Erhard and colleagues 2020. C, Comparison of the atla...
Building Trust in Deep Learning-based Immune Response Predictors with Interpretable Explanations
The ability to predict whether a peptide will get presented on Major Histocompatibility Complex (... more The ability to predict whether a peptide will get presented on Major Histocompatibility Complex (MHC) class I molecules has profound implications in designing vaccines. Numerous deep learning-based predictors for peptide presentation on MHC class I molecules exist with high levels of accuracy. However, these MHC class I predictors are treated as black-box functions, providing little insight into their decision making. To build turst in these predictors, it is crucial to understand the rationale behind their decisions with human-interpretable explanations. We present MHCXAI, eXplainable AI (XAI) techniques to help interpret the outputs from MHC class I predictors in terms of input peptide features. In our experiments, we explain the outputs of four state-of-the-art MHC class I predictors over a large dataset of peptides and MHC alleles. Additionally, we evaluate the reliability of the explanations by comparing against ground truth and checking their robustness. MHCXAI seeks to increa...
Image recognition tasks typically use deep learning and require enormous processing power, thus r... more Image recognition tasks typically use deep learning and require enormous processing power, thus relying on hardware accelerators like GPUs and FPGAs for fast, timely processing. Failure in real-time image recognition tasks can occur due to incorrect mapping on hardware accelerators, which may lead to timing uncertainty and incorrect behavior. Owing to the increased use of image recognition tasks in safetycritical applications like autonomous driving and medical imaging, it is imperative to assess their robustness to changes in the computational environment as parameters like deep learning frameworks, compiler optimizations for code generation, and hardware devices are not regulated with varying impact on model performance and correctness. In this paper we conduct robustness analysis of four popular image recognition models (MobileNetV2, ResNet101V2, DenseNet121 and InceptionV3) with the ImageNet dataset, assessing the impact of the following parameters in the model's computational environment: (1) deep learning frameworks; (2) compiler optimizations; and (3) hardware devices. We report sensitivity of model performance in terms of output label and inference time for changes in each of these environment parameters. We find that output label predictions for all four models are sensitive to choice of deep learning framework (by up to 57%) and insensitive to other parameters. On the other hand, model inference time was affected by all environment parameters with changes in hardware device having the most effect. The extent of effect was not uniform across models. PyTorch PyTorch GPU 1 Cyclist 80ms PyTorch TensorFlow GPU 1 Cat 75ms PyTorch PyTorch GPU 2 Cyclist 150ms True label: Cyclist Ref. inf. time: 80ms TensorFlow TensorFlow GPU 1 Cyclist 85ms TensorFlow PyTorch GPU 1 Cyclist 90ms TensorFlow TensorFlow GPU 2 Cyclist 130ms Framework Model Source Device Classification Inf. time
Figure 5 from The Immunopeptidome from a Genomic Perspective: Establishing the Noncanonical Landscape of MHC Class I–Associated Peptides
Origins of ncMAPs. A, Peptide length distribution of canonical (dark gray) and noncanonical (ligh... more Origins of ncMAPs. A, Peptide length distribution of canonical (dark gray) and noncanonical (light gray) MAPs. B, Annotation of ncMAPs across gene features. C, Analysis of ncMAPs that could originate from nORF. Upstream start codons of noncanonical MAPs are analyzed for their potential to initiate translation and produce ORFs (left-hand side) as a source of ncMAPs. The frequencies of different start codons for positively predicted TIS are shown on the right-hand side. D, Analysis of ncMAPs from intronic regions that may originate from IR events. Translation of MAPs from IR sources should be in-frame with the corresponding upstream exons. E, Analysis of ncMAPs that could originate from frameshift mutations in cancer. ncMAPs are aligned to an in-silico translated protein database of COSMIC somatic frameshift mutations. F, Summary indicating whether the ncMAPs can be accounted for by any of the analyses conducted in panels C, D, or E.
Proceedings of the International Conference on Parallel Architectures and Compilation Techniques
Finding the right heuristics to optimize code has always been a difficult and mostly manual task ... more Finding the right heuristics to optimize code has always been a difficult and mostly manual task for compiler engineers. Today this task is near-impossible as hardware-software complexity has scaled up exponentially. Predictive models for compilers have recently emerged which require little human effort but are far better than humans in finding near optimal heuristics. As any machine learning technique, they are only as good as the data they are trained on but there is a severe shortage of code for training compilers. Researchers have tried to remedy this with code generation but their synthetic benchmarks, although thousands, are small, repetitive and poor in features, therefore ineffective. This indicates the shortage is of feature quality more than corpus size. It is more important than ever to develop a directed program generation approach that will produce benchmarks with valuable features for training compiler heuristics. We develop BenchPress, the first ML benchmark generator for compilers that is steerable within feature space representations of source code. BenchPress synthesizes compiling functions by adding new code in any part of an empty or existing sequence by jointly observing its left and right context, achieving excellent compilation rate. BenchPress steers benchmark generation towards desired target features that has been impossible for state of the art synthesizers (or indeed humans) to reach. It performs better in targeting the features of Rodinia benchmarks in 3 different feature spaces compared with (a) CLgen-a state of the art ML
Automatic speech recognition (ASR) models are prevalent, particularly in applications for voice n... more Automatic speech recognition (ASR) models are prevalent, particularly in applications for voice navigation and voice control of domestic appliances. The computational core of ASRs are deep neural networks (DNNs) that have been shown to be susceptible to adversarial perturbations; easily misused by attackers to generate malicious outputs. To help test the security and robustnesss of ASRS, we propose techniques that generate blackbox (agnostic to the DNN), untargeted adversarial attacks that are portable across ASRs. This is in contrast to existing work that focuses on whitebox targeted attacks that are time consuming and lack portability. Our techniques generate adversarial attacks that have no human audible difference by manipulating the audio signal using a psychoacoustic model that maintains the audio perturbations below the thresholds of human perception. We evaluate portability and effectiveness of our techniques using three popular ASRs and two input audio datasets using the metrics-Word Error Rate (WER) of output transcription, Similarity to original audio, attack Success Rate on different ASRs and Detection score by a defense system. We found our adversarial attacks were portable across ASRs, not easily detected by a state-of-the-art defense system, and had significant difference in output transcriptions while sounding similar to original audio.
Poster: Reordering Tests for Faster Test Suite Execution
2018 IEEE/ACM 40th International Conference on Software Engineering: Companion (ICSE-Companion), 2018
As software takes on more responsibility, it gets increasingly complex, requiring an extremely la... more As software takes on more responsibility, it gets increasingly complex, requiring an extremely large number of tests for effective validation. Executing these large test suites is expensive, both in terms of time and energy. Cache misses are a significant contributing factor to execution time of software. In this paper, we propose an approach that helps order test executions in a test suite in such a way that instruction cache misses are reduced, and thereby execution time. We conduct an empirical evaluation with 20 subject programs and test suites from the SIR repository, EEMBC suite and LLVM Symbolizer, comparing execution times and cache misses with test orderings maximising instruction locality versus a traditional ordering maximising coverage, as well as random permutations. Performance gains were considerable for programs and test suites where the average number of different instructions executed between tests was high. We achieved an average execution speedup of 6.83% and a m...
When creating test cases for software, a common approach is to create tests that exercise require... more When creating test cases for software, a common approach is to create tests that exercise requirements. Determining the adequacy of test cases, however, is generally done through inspection or indirectly by measuring structural coverage of an executable artifact (such as source code or a software model). We present ReqsCov, a tool to directly measure requirements coverage provided by test cases. ReqsCov allows users to measure Linear Temporal Logic requirements coverage using three increasingly rigorous requirements coverage metrics: naïve coverage, antecedent coverage, and Unique First Cause coverage. By measuring requirements coverage, users are given insight into the quality of test suites beyond what is available when solely using structural coverage metrics over an implementation.
Using large pre-trained models for image recognition tasks is becoming increasingly common owing ... more Using large pre-trained models for image recognition tasks is becoming increasingly common owing to the well acknowledged success of recent models like vision transformers and other CNN-based models like VGG and Resnet. The high accuracy of these models on benchmark tasks has translated into their practical use across many domains including safety-critical applications like autonomous driving and medical diagnostics. Despite their widespread use, image models have been shown to be fragile to changes in the operating environment, bringing their robustness into question. There is an urgent need for methods that systematically characterise and quantify the capabilities of these models to help designers understand and provide guarantees about their safety and robustness. In this paper, we propose Vision Checklist, a framework aimed at interrogating the capabilities of a model in order to produce a report that can be used by a system designer for robustness evaluations. This framework pr...
The major histocompatibility complex (MHC) class-I pathway supports the detection of cancer and v... more The major histocompatibility complex (MHC) class-I pathway supports the detection of cancer and viruses by the immune system. It presents parts of proteins (peptides) from inside a cell on its membrane surface enabling visiting immune cells that detect non-self peptides to terminate the cell. The ability to predict whether a peptide will get presented on MHC Class I molecules helps in designing vaccines so they can activate the immune system to destroy the invading disease protein. We designed a prediction model using a BERT-based architecture (ImmunoBERT) that takes as input a peptide and its surrounding regions (N and C-terminals) along with a set of MHC class I (MHC-I) molecules. We present a novel application of well known interpretability techniques, SHAP and LIME, to this domain and we use these results along with 3D structure visualizations and amino acid frequencies to understand and identify the most influential parts of the input amino acid sequences contributing to the ou...
Automatic speech recognition (ASR) systems are prevalent, particularly in applications for voice ... more Automatic speech recognition (ASR) systems are prevalent, particularly in applications for voice navigation and voice control of domestic appliances. The computational core of ASRs are deep neural networks (DNNs) that have been shown to be susceptible to adversarial perturbations; easily misused by attackers to generate malicious outputs. To help test the correctness of ASRS, we propose techniques that automatically generate blackbox (agnostic to the DNN), untargeted adversarial attacks that are portable across ASRs. Much of the existing work on adversarial ASR testing focuses on targeted attacks, i.e generating audio samples given an output text. Targeted techniques are not portable, customised to the structure of DNNs (whitebox) within a specific ASR. In contrast, our method attacks the signal processing stage of the ASR pipeline that is shared across most ASRs. Additionally, we ensure the generated adversarial audio samples have no human audible difference by manipulating the aco...
The immunopeptidome from a genomic perspective: Establishing immune-relevant regions for cancer vaccine design
A longstanding disconnect between the growing number of MHC Class I immunopeptidomic studies and ... more A longstanding disconnect between the growing number of MHC Class I immunopeptidomic studies and genomic medicine hinders cancer vaccine design. We develop COD-dipp to genomically map the full spectrum of detected canonical and non-canonical (non-exonic) MHC Class I antigens from 26 cancer studies. We demonstrate that patient mutations in regions overlapping physically identified antigens better predict immunotherapy response when compared to neoantigen predictions. We suggest a vaccine design approach using 140,966 highly immune-visible regions of the genome annotated by their expression and haplotype frequency in the human population. These regions tend to be highly conserved, mutated in cancer and harbor 7.8 times more immunogenicity. Intersecting pan-cancer mutations with these immune surveilled regions revealed a potential to create off-the-shelf multi-epitope vaccines against public neoantigens. Here we release COD-dipp, a cancer vaccine toolkit as a web-application (https://w...
Solidity is an object-oriented and high-level language for writing smart contracts that are used ... more Solidity is an object-oriented and high-level language for writing smart contracts that are used to execute, verify and enforce credible transactions on permissionless blockchains. In the last few years, analysis of smart contracts has raised considerable interest and numerous techniques have been proposed to check the presence of vulnerabilities in them. Current techniques lack traceability in source code and have widely differing work flows. There is no single unifying framework for analysis, instrumentation, optimisation and code generation of Solidity contracts. In this paper, we present SIF, a comprehensive framework for Solidity contract analysis, query, instrumentation, and code generation. SIF provides support for Solidity contract developers and testers to build source level techniques for analysis, understanding, diagnostics, optimisations and code generation. We show feasibility and applicability of the framework by building practical tools on top of it and running them on 1838 real smart contracts deployed on the Ethereum network. Index Terms-high level languages, software testing, code instrumentation, program analysis 4 Node [2]-> Node [3];
In black-box testing, the tester creates a set of tests to exercise a system under test without r... more In black-box testing, the tester creates a set of tests to exercise a system under test without regard to the internal structure of the system. Generally, no objective metric is used to measure the adequacy of black-box tests. In recent work, we have proposed three requirements coverage metrics, allowing testers to objectively measure the adequacy of a black-box test suite with respect to a set of requirements formalized as Linear Temporal Logic (LTL) properties. In this report, we evaluate the effectiveness of these coverage metrics with respect to fault finding. Specifically, we conduct an empirical study to investigate two questions: (1) do test suites satisfying a requirements coverage metric provide better fault finding than randomly generated test suites of approximately the same size?, and (2) do test suites satisfying a more rigorous requirements coverage metric provide better fault finding than test suites satisfying a less rigorous requirements coverage metric? Our results...
Proceedings of the 10th ACM/IEEE International Symposium on Empirical Software Engineering and Measurement, 2016
Software energy consumption has emerged as a growing concern in recent years. Managing the energy... more Software energy consumption has emerged as a growing concern in recent years. Managing the energy consumed by a software is, however, a difficult challenge due to the large number of factors affecting it -namely, features of the processor, memory, cache, and other hardware components, characteristics of the program and the workload running, OS routines, compiler optimisations, among others. In this paper we study the relevance of numerous architectural and program features (static and dynamic) to the energy consumed by software. The motivation behind the study is to gain an understanding of the features affecting software energy and to provide recommendations on features to optimise for energy efficiency. In our study we used 58 subject desktop programs, each with their own workload, and from different application domains. We collected over 100 hardware and software metrics, statically and dynamically, using existing tools for program analysis, instrumentation and run time monitoring. We then performed statistical feature selection to extract the features relevant to energy consumption. We discuss potential optimisations for the selected features. We also examine whether the energy-relevant features are different from those known to affect software performance. The features commonly selected in our experiments were execution time, cache accesses, memory instructions, context switches, CPU migrations, and program length (Halstead metric). All of these features are known to affect software performance, in terms of running time, power consumed and latency.
Uploads
Papers by Ajitha Rajan