Bio-Cognito: A Generalist AI Agent Framework for Medical and Computational Biology Deep Research

While the cross-over between AI and biomedical research has been tremendously successful from multiple standpoints, there exist fundamental drawbacks in the current approaches towards tackling these challenges.

2.1 AI agents in medical domains

Early medical AI agents addressed narrow clinical tasks, achieving notable advances in diagnostic imaging and decision support. Singhal et al. [5] demonstrated strong performance using MedPaLM for medical licensing examinations, attaining a USMLE score of 67.6% through transformer-based large language model (LLM) architectures applied to clinical reasoning. At the same time, Li et al. [6] applied this approach for clinical conversation purposes as well, since domain adaptation for question answering using the LLaMA-based ChatDoctor proved itself to be rather efficient in practice. However, the problem with all the aforementioned approaches is their monolithic structure, as the entire process is performed by one model without any specialization of sub-agents.

The monolithic approach exhibits three principal limitations. First, multi-task specialization results in competing objectives within a shared parameter space, limiting performance across any single domain. Second, these models do not natively support structured biological formats such as FASTA sequences, VCF variants, or PDB protein files. Third, clinical safety verification—including drug interaction checking, dosage validation, and audit trail maintenance—cannot be performed within a single-model architecture. Recent prompt-engineering approaches have partially addressed some of these issues; however, the fundamental challenge of inter-agent coordination remains unresolved.

2.2 Computational biology frameworks

In contrast to the above, there have been enormous developments in the field of computational biology in connection with protein structure predictions and genomics analysis. The study by Abramson et al. [7] demonstrated unprecedented 92.4% GDT-TS (Global Distance Test-Template Score) results achieved in protein structure prediction through AlphaFold and AlphaFold3 via deep learning. Additionally, Gao et al. [8] demonstrated the effectiveness of transformer-based architectures for medical image segmentation, further illustrating the expanding role of transformers in biological data processing. Deep learning segmentation methods have shown strong performance across a range of medical imaging modalities; a comprehensive review of classical, machine learning, and deep learning segmentation approaches for breast cancer imaging highlights that U-Net architectures can achieve accuracy levels up to 99.7% in specialized contexts [9]. For genomic analysis tools such as Nucleotide Transformer and DNABERT-2, transformer-based genomics sequence processing is implemented. Explainable multi-task learning approaches have further demonstrated the utility of integrating multi-modality biological data for improved genomic and clinical analysis [10].

However, despite all successes shown by some computational biology tools in their fields of application, there are numerous critical limitations that these tools possess concerning their incorporation into medical environments. As seen from tools such as scGPT (for single-cell analysis), and GeneGPT (genomic question answering), the problem concerns the lack of support for multi-modality in personalized medicine. Recent attempts have been made to create biomedical foundation models, e.g., the research by Zhang et al. [11] with BiomedGPT. Even though the most basic integration of vision and language modalities is provided by those models, the issue remains unsolved because of the inability to meet all requirements of the environment regarding deployment and safety validation. Importantly, no framework is capable of working with input data in the format of VCF and FASTA files directly.

2.3 Generalist AI architectures

In this context, the emerging field of general-purpose AI agents can provide an opportunity for better solutions to multitasking problems. Some examples of generalists include AutoGPT and LangChain platforms, used for creating tool-using AI agents. Improvements on this subject also include the use of multi-agent architectures that have been successful in implementation, such as MetaGPT and ChatDev systems. Moreover, recently, multi-agent architectures have found applications in the medical field, in particular in the study by Zuo et al. [12], where a knowledge-based hierarchical multi-agent LLM framework is applied for medical diagnosis.

A structured comparison of the key capabilities of these existing medical AI approaches relative to Bio-Cognito is provided in Table 2.

Table 2. Comparison of existing medical AI approaches

Figure 2 illustrates how a central orchestrator agent manages domain-specific sub-agents while maintaining coherent reasoning across tasks.

Figure 2. Illustration of how a central agent manages domain agents and maintains the consistency of reasoning capabilities

When applied to medical contexts, however, these approaches exhibit inherent limitations. Unlike software engineering or web-browsing tasks, medical applications require domain-specific expertise, rigorous safety verification, and computationally efficient execution on resource-constrained clinical hardware. Prior attempts at medical multi-agent systems have been insufficient in delivering the degree of domain specialization and safety validation required for clinical use. The prompt-engineering-based adaptations of these frameworks do not account for the aforementioned problems and neither offer solutions regarding the ingestion of structured biological data nor clinical safety at inference time.

2.4 Efficiency and scalability

Moreover, one can easily identify the lack of attention paid to the implementation efficiency and scalability of AI algorithms in medicine. While multimodal AI algorithms such as Med-Flamingo by Moor et al. [13], which incorporate the visual-language modeling along with the use of medical images in them, demonstrate high performance in the field, they still involve complex computations that cannot be performed within the clinical infrastructure framework. Hybrid approaches integrating transformer-based feature extraction with lightweight classifiers have demonstrated that clinically competitive performance can be achieved without prohibitive computational overhead, a principle that informs Bio-Cognito's quantization-aware deployment design.

Efforts towards solving this problem include the QLoRA work by Dettmers et al. [14] that uses quantization techniques, among other techniques, but does not consider domain-specific compression algorithms. Additionally, the deployment problem includes the need to integrate with clinical workflows and meet certain regulatory requirements. The mentioned solutions offer improvements in particular aspects of deployment but still fail to provide a comprehensive solution to be used in a clinical environment that includes EHR integration and other requirements [15].

2.5 Positioning Bio-Cognito

Unlike those suggested before, which exhibited advantages in certain domains, lacked in-depth research, but had a wide scope, or vice versa, Bio-Cognito addresses the problems found through its innovative aspects as follows:

• Using our hierarchical multi-agent architecture, we allow for the agents' specialization in certain domains without loss of coordination, unlike generalist architectures like AutoGPT-Medical or LangChain that use reasoning in only one loop rather than the cooperation of several sub-agents in it.
• Because of the native computational biology support in our framework, the analysis can be performed not only with clinical data, but also in different file formats (FASTA, VCF, PDB, FASTQ).
• Because of the nature of the approach, we can unify quantization and clinical safety validation in one pipeline and solve the problem of versatility and applicability of current medical AI solutions.

In other words, Bio-Cognito is an approach combining the efforts of specialized biomedical agents under a single hierarchical architecture, providing for both deep domain knowledge and deployment optimization with rigorous clinical safety validation.

Bio-Cognito was evaluated across medical knowledge assessment, computational biology tasks, and clinical decision support scenarios. The experimental evaluation yields consistent, statistically significant improvements over strong baselines while maintaining computational efficiency suitable for clinical deployment.

4.1 Experimental setup

4.1.1 Dataset configuration

Our evaluation employs four primary datasets representing diverse medical and biological domains: MedQA containing 12,723 multiple-choice questions from medical licensing examinations, Autonomous Medical Evaluation for Guideline Adherence (AMEGA) with 2,850 clinical guideline adherence scenarios [18], United States Medical Licensing Examination (USMLE)-Style simulations encompassing 5,200 clinical reasoning cases, and a curated Multi-Modal Medical dataset integrating 467 patient cases with genomic (VCF- Variant Call Format), imaging (DICOM- Digital Imaging and Communications in Medicine), and clinical (HL7 FHIR- Health Level Seven Fast Healthcare Interoperability Resources) data modalities.

Each dataset was selected to test a distinct capability of the framework. MedQA and USMLE-Style cases assess general clinical reasoning across multiple specialties, AMEGA evaluates adherence to structured clinical guidelines, and the Multi-Modal Medical set is specifically designed to stress-test the framework's capacity to integrate heterogeneous data sources. For MedQA and USMLE, questions were drawn from publicly available standardized examination repositories. For the Multi-Modal Medical dataset, cases were constructed by pairing synthetic patient records with matched genomic variants and imaging studies, following stratified sampling across cardiology, oncology, and infectious disease domains to reduce specialty bias. All datasets were partitioned into 70/15/15 train/validation/test splits with no case overlap between partitions. Readers should note that the multi-modal dataset is simulation-based and is used here for feasibility evaluation; prospective validation on de-identified real patient data constitutes a key direction for future work.

The characteristics of all evaluation datasets and their primary metrics are summarized in Table 5.

Table 5. Experimental dataset characteristics and evaluation metrics

The choice of primary evaluation metric for each task reflects the nature of the underlying problem: Pass@1 accuracy and diagnostic accuracy are appropriate for multiple-choice settings where a single correct answer exists, guideline adherence rate captures rule-following behaviour in structured clinical scenarios, and F1-score accounts for both precision and recall in the multi-label multi-modal integration task. Where tasks share a multiple-choice format, accuracy is reported uniformly to enable direct cross-task comparison.

4.1.2 Baseline methods and implementation

We compare Bio-Cognito against five state-of-the-art approaches: GPT-4 with medical prompting [19], MedPaLM-2 [20], AutoGPT-Medical [21], and domain-specific models including ClinicalBERT [22]. All experiments utilize T4v2 GPU infrastructure with 4-bit NF4 quantization and LoRA adaptation (rank r = 8, α =16) for computational efficiency.

To ensure a fair comparison, all baseline models were evaluated on the same T4v2 hardware under identical input conditions. For GPT-4 and MedPaLM-2, which do not natively support 4-bit quantization, inference was conducted at full precision via their respective APIs; reported latency figures for these models therefore reflect API round-trip time rather than on-device inference and should not be directly compared with on-device latency values for Bio-Cognito. For LangChain Medical and AutoGPT-Medical, we used publicly available configurations without additional prompt tuning beyond what is documented in the respective repositories, to avoid inadvertently favouring or disadvantaging these systems. Clinical BERT was evaluated using its standard fine-tuned checkpoint without further task-specific adaptation.

Statistical significance is assessed using paired t-tests with Bonferroni correction for multiple comparisons, reporting the results as µ ± 1.96σ/√n confidence intervals, where p < 0.05 indicates significance.

4.2 Performance analysis

4.2.1 Medical knowledge and reasoning

The experimental evaluation demonstrates consistent performance gains for Bio-Cognito across all medical benchmarks, with improvements reaching statistical significance in every primary comparison. On MedQA, Bio-Cognito achieved an accuracy of 79.5% ± 2.1%, representing a Δ = +13.6% improvement over the GPT-4 baseline (70.0% ± 2.8%; p < 0.001, paired t-test with Bonferroni correction). On AMEGA, the system achieved a guideline adherence rate of 87.6% ± 2.4% (Δ = +20.8% over GPT-4; p < 0.001). On USMLE-Style cases, diagnostic accuracy reached 82.4% ± 2.6%, corresponding to a Δ = +20.6% improvement (p < 0.001). The Pass@1 accuracy of the baseline model and all Med-Agent variants (FP16, INT8, NF4) across these benchmarks is visualized in Figure 3.

Figure 3. Pass@1 accuracy of baseline model and Med-Agent variants (fp16, int8, nf4) across medical benchmarks (AMEGA, MedQA, USMLE)

When evaluated at up to three attempts (Pass@3), all Med-Agent variants approach ceiling performance across benchmarks, as shown in Figure 4, demonstrating the framework's strong consistency under repeated inference.

Figure 4. When allowed up to three attempts (Pass@3), all Med-Agent models perform nearly flawlessly across medical benchmarks, showing only slight differences from the baseline

Table 6. Performance comparison on medical benchmarks

The largest difference is observed in the clinical safety metric, where Bio-Cognito achieves 82.7% ± 2.8%, a Δ = +28.8% improvement over the GPT-4 baseline of 64.2% ± 4.1% (p < 0.001). This result is attributable to the multi-agent consultation mechanism and the dedicated clinical safety validation layer, which independently checks each recommendation against drug-interaction rules and clinical guideline constraints before any output is returned to the user. Detailed performance comparisons across all baselines are reported in Table 6.

4.2.2 Multi-modal integration capabilities

The multi-modal processing pipeline of Bio-Cognito demonstrates substantial performance advantages in integrated clinical settings, aligning with advances in multimodal analysis methods for predictive biomedicine [23]. On the Multi-Modal Medical dataset, the architecture performs with the F1-score of 0.884 ± 0.032 compared to the best baseline (GPT-4 + concatenated inputs) with 0.712 ± 0.047, for a gain Δ = +24.2% (p < 0.001). The gain from multi-modal processing is especially evident in those examples where interpretation of genomics requires a combination with clinical history. Thus, modality-specific processing proves superior to naive input concatenation.

The multi-modal fusion is based on an attention-weighted combination of all available modalities as follows:

$R_{\text {unified}}=\sum_{i=1}^k \alpha_i \cdot \phi_i\left(m_i\right)$

where, the learned attention weights of the respective modalities are {α₁ = 0.34, α₂ = 0.41, α₃ = 0.25}. These weights were determined by training on a multi-modal dataset and proved stable across all cross-validation folds (variance < 0.03), suggesting constant modality priorities. Notably, the highest attention is paid to the clinical modality (α₂ = 0.41). The priority of the clinical modality can be considered clinically valid, since patient medical records usually have the greatest diagnostic value unless the case includes new, urgent imaging data.

4.3 Ablation studies

To evaluate the contribution of individual components, systematic ablation experiments were conducted. In each configuration, a single component was removed while all remaining components were held constant. Results are reported for the test split of MedQA dataset [24, 25].

Results of ablation experiments, listed in Table 7, show that each component contributes uniquely to one particular aspect of the architecture's performance. Multi-agent coordination has the greatest influence on diagnostic accuracy (+5.3%). Consistent with the function of consensus among agents, the multi-agent mechanism is responsible for solving clinically ambiguous problems. Validation on the clinical safety checklist has the strongest effect on safety score (+13.3%), but does not affect diagnostic accuracy, which shows that this component implements a separate safety-check function. The largest single-component accuracy drop is caused by omitting multi-modal fusion (−6.9%). As expected, multi-modal fusion is crucial for processing multi-modal tasks. Omitting integration with the knowledge base also leads to the significant drop in accuracy (−8.2%) and in the safety score, as expected.

Table 7. Ablation study results on MedQA dataset

Analysis of the effects of quantization on the system performance suggests that NF4 reduces computation time by 41.7% and memory usage by 73.2% compared to the full-precision model while preserving 98.4% of its performance:

$\eta_{\text {quant}}=\frac{\text { Accuracy}_{N F 4}}{\text { Accuracy}_{F P 32}} \times \frac{\text { Latency}_{F P 32}}{\text { Latecy}_{N F 4}}$

Such computational savings are highly relevant in clinical environments where GPU memory should be shared between multiple software applications.

The latency test results indicate that when multi-agent coordination is turned off, it causes a decrease in inference time by 19.3% (198.3 ms against 245.8 ms). This compromise might be worthwhile in instances of emergency triage where speed is needed, not accuracy. Hence, multi-agent coordination may be an option that may be turned off under time constraints. The deactivation of LoRA quantization causes an almost doubling in inference latency (421.7 ms), confirming that NF4 is critical for real-time performance on T4 GPUs.

4.4 Domain generalization and transfer learning

Bio-Cognito demonstrates robust cross-domain generalization capabilities. When evaluated on specialized computational biology tasks, the framework demonstrated strong performance across diverse biological data types.

Cross-domain performance results across specialized biological tasks are reported in Table 8.

Table 8. Cross-domain performance on specialized biological tasks

The framework's ability to natively process structured biological file formats (VCF, FASTA, PDB) while preserving clinical reasoning capabilities represents a key differentiating advantage. In contrast, general-purpose language models require all inputs to be serialized as plain text, a process that discards structural information critical for tasks such as variant pathogenicity classification or protein binding-site identification. A comparison of the average performance of all Med-Agent variants against the baseline model across all medical benchmarks is presented in Figure 5.

Figure 5. Average performance of all med-agent variants v/s baseline model across all medical benchmarks

4.5 Expert validation and clinical assessment

The clinical utility of Bio-Cognito was assessed through a structured pilot expert validation study. Two board-certified clinicians, one in general internal medicine and one with a background in clinical bioinformatics, independently reviewed a set of 10 synthetic patient cases processed by the framework. Each expert rated the system's outputs on a five-point Likert scale across four dimensions: clinical accuracy, safety of recommendations, interpretability of reasoning, and practical utility in a clinical workflow. The rating instrument was administered after each case, and experts were blinded to baseline comparison outputs during their initial assessments.

Expert consensus, measured as the proportion of cases on which both reviewers assigned ratings within one Likert point, reached 97.0%. The mean overall rating was 4.2 ± 0.3 out of 5.0. This level of inter-rater agreement is consistent with published benchmarks for AI-assisted clinical documentation review. It should be noted that this constitutes a small-scale pilot validation; broader evaluation involving a larger panel of clinicians across multiple specialties and institutional settings is necessary before deployment claims can be made.

A composite deployment readiness score was computed across four dimensions: regulatory compliance, clinical integration readiness, technical readiness, and clinical effectiveness, weighted equally:

Readiness$_{\text {deploy}}=\frac{1}{4}\left(C_{\text {reg}}+C_{\text {int}}+C_{\text {tech}}+C_{\text {eff}}\right)$

where, C_reg = 0.887, C_int = 1.000, C_tech = 1.000, and C_eff = 0.908, yielding an overall score of 0.949. The regulatory compliance dimension (0.887) reflects the absence of formal ISO 13485 certification and prospective clinical trial data, both of which are prerequisites for clinical-grade deployment in most jurisdictions.

4.6 Discussion of results

The experimental results indicate that Bio-Cognito achieves statistically significant improvements across all evaluated benchmarks, with the most pronounced gains observed in clinical safety scoring and multi-modal integration tasks. The key findings are as follows:

Medical Knowledge: Consistent improvements of 13–29% over strong baselines across standardised medical benchmarks support the validity of the hierarchical multi-agent reasoning approach.

Multi-Modal Integration: The 24.2% improvement in multi-modal F1-score demonstrates the benefit of modality-specific processing pipelines over text-serialisation-based approaches.

The ablation analysis indicates that Bio-Cognito's performance gains arise from the synergistic interaction of its components rather than any single architectural element. Multi-agent coordination, clinical safety validation, and domain knowledge integration each contribute measurably to distinct performance dimensions, a pattern consistent with findings in related multi-agent clinical systems.

Statistical significance testing (p < 0.001, Bonferroni-corrected paired t-tests) provides confidence in the reported improvements across all primary comparisons. The cross-domain results further suggest that the framework's learned representations generalise across specialised biological reasoning tasks, though the performance levels on simulation-based datasets should be interpreted with caution pending validation on real clinical data.

From an information systems perspective, the deployment readiness score of 0.949 and the pilot expert consensus rating of 4.2/5.0 are encouraging indicators of system integration potential. However, practical adoption will depend on factors beyond model performance, including clinician trust, alignment with existing hospital workflows, governance frameworks for model updates, and interoperability with EHR systems via HL7 FHIR interfaces — all of which represent priorities for future system-engineering work.

These results suggest that the pragmatic integration of hierarchical multi-agent coordination, safety-aware training, and domain-adaptive quantization produces measurable and reproducible improvements over both monolithic LLMs and modular pipeline-based systems on standardised medical benchmarks. Performance in uncontrolled clinical environments may differ, and prospective validation studies are required before operational deployment conclusions can be drawn.